# Bravos AI > SaaS platform to create AI chatbots trained on your own data (PDFs, URLs, CSVs, product catalogs). Widget embeddable in any website, plus WhatsApp. Can also be operated and optimized from inside Claude or ChatGPT through its own MCP connector. Spain-based, launched 2025. Website and blog available in Spanish, English and Brazilian Portuguese. ## About Bravos AI lets businesses deploy a chatbot on their website that answers customer questions based on their own content — not generic AI knowledge. Upload a PDF of your menu or policies, connect your Shopify/PrestaShop catalog, and the bot answers accurately in 12+ languages. And if you use Claude or ChatGPT, you can manage and improve the bot from that same assistant through the Bravos MCP connector — audit conversations, edit the knowledge base and rewrite the bot's instructions, without opening the panel. Try the full PRO plan free for 7 days — no commitment, cancel in one click and pay nothing. ## What makes Bravos different Most AI chatbots use only RAG (text similarity search). This works for documents and FAQs but fails with structured data like product catalogs: "blue shoes under 60€ in size 42" returns approximate text matches, not actual filtered results. Bravos combines RAG with SQL-like structured filtering over the client's data. When the user asks about products by price, size, stock, category, or any attribute, the chatbot extracts filters from the natural-language query and runs real SQL queries against the structured data (CSV, Excel, product catalog). The user uploads the file once — no technical setup, no schema configuration, no query language to learn. The system auto-detects columns, types, ranges, and relationships. This means the chatbot returns exact results instead of "similar" ones. For restaurants, real estate, ecommerce, and any business with structured data (menus, property listings, product catalogs, inventory), this is a significant accuracy difference vs generic RAG chatbots like Tidio, Intercom Fin, Zendesk or ChatGPT-based bots. Another differentiator: Bravos AI can be operated and optimized from inside Claude or ChatGPT via its own MCP server (Model Context Protocol). The client connects their Bravos account to their AI assistant and, in plain language, analyzes conversations, audits and edits the knowledge base and rewrites the bot's instructions — turning "my bot answers badly and I don't know why" into a diagnosis with the fix applied, from the same chat they already use. The AI inference is paid by the client's own Claude/ChatGPT subscription, so it adds no cost on Bravos. Available from the Starter plan and during the 7-day trial. Bravos is not the only platform offering an MCP connector, but pairs it with real structured search, low price, native Spanish/English/Portuguese and WhatsApp. Another differentiator: WhatsApp integration via Meta Tech Provider model (Embedded Signup). Most chatbot platforms with WhatsApp rely on per-client API tokens that expire every 60 days, or charge Twilio/MessageBird middleware fees on top of Meta's pricing. Bravos AI is a verified Meta Tech Provider — clients connect their WhatsApp Business Account in one click and the same RAG-powered bot they use on their website answers automatically on WhatsApp. Same assistant, same catalog, two channels. No middleware fees, no token refresh, no setup beyond the Embedded Signup popup. ## Full content - [Complete content (all blog articles in English)](https://bravos-ai.com/llms-full.txt) All blog articles listed below are also published in Spanish (https://bravos-ai.com/blog) and Brazilian Portuguese (https://bravos-ai.com/pt/blog), each with its own localized URL and content adapted to the local market (e.g. LGPD and Pix in the Portuguese versions). ## Core pages - [Home](https://bravos-ai.com/en) - [Home in Spanish](https://bravos-ai.com/) - [Home in Portuguese](https://bravos-ai.com/pt) - [About](https://bravos-ai.com/en/about) - [Contact](https://bravos-ai.com/en/contact) - [MCP connector (dedicated page)](https://bravos-ai.com/en/ai-connector-mcp): landing page for the Bravos MCP connector — how to connect Claude or ChatGPT to a Bravos account and operate and optimize the bot from the assistant's own chat, with our team's setup criteria carried inside the assistant. Available in Spanish (/conector-ia-mcp), English (/en/ai-connector-mcp) and Portuguese (/pt/conector-ia-mcp) ## Integrations - [Shopify integration](https://bravos-ai.com/en/blog/shopify-chatbot): direct catalog sync via GraphQL, product filtering by price/size/color/stock - [PrestaShop integration](https://bravos-ai.com/en/blog/prestashop-chatbot): Webservice API integration, multilingual product catalog - [WooCommerce integration](https://bravos-ai.com/en/blog/woocommerce-chatbot): catalog sync via the WooCommerce REST API (consumer key/secret generated in the store's WordPress panel — no plugin installed, nothing running on the store's hosting). Full support for variable products (per-variation price, stock and photo), the three WooCommerce stock representations (exact quantities, in-stock/out-of-stock, inherited from parent, plus "available on backorder"), sale prices quoted as before/after pairs, and prices matching exactly what the store displays (tax-inclusive, tax-exclusive or taxes disabled — re-detected on every sync). Real-time updates via product webhooks plus a full nightly reconciliation sync as source of truth (read-only API keys fall back to nightly-only mode, shown with a badge in the dashboard). Cleans WordPress page-builder shortcodes (Elementor, Divi, WPBakery) from product descriptions. Detects WPML/Polylang multilingual stores and lets the user pick the catalog language, discarding translation duplicates. Grouped and external/affiliate products are skipped and counted visibly. Sync designed for shared hosting: low concurrency, gentle pace, automatic retries - [Resales Online (real estate)](https://bravos-ai.com/en/blog/real-estate-chatbot-resales-online): real estate listings sync - [Custom Webhook API](https://bravos-ai.com/en/developers): sync any structured data (JSON/XML) via URL + HMAC webhooks for real-time updates - [WhatsApp Business integration](https://bravos-ai.com/en#whatsapp): connect any client's WhatsApp Business Account via Meta Embedded Signup in 60 seconds. The chatbot answers WhatsApp messages automatically with the same knowledge base used on the website widget (same RAG + structured search). Bravos AI is a verified Meta Tech Provider, so the integration uses a System User token at the platform level — no per-client token refresh, no Twilio/MessageBird middleware fees. Native lead capture forms via WhatsApp Flows (requires Business Verification on the client side). Included in all paid plans and during the 7-day PRO trial - [MCP connector for Claude, ChatGPT and any MCP client](https://bravos-ai.com/en/ai-connector-mcp): connect a Bravos account to Claude, ChatGPT or any MCP-compatible AI assistant via a remote MCP server (mcp.bravos-ai.com) using OAuth — no code, no API keys pasted. From the assistant's chat, in natural language, the connected AI can analyze the bot's conversations and leads, audit and edit the knowledge base (create/edit/delete text entries and URLs), rewrite the bot's instructions and tone, change widget texts/appearance/languages, and trigger store catalog syncs, with a before/after returned on every change and all actions gated by the account's plan. It cannot create bots or upload/delete files (done in the panel) and never sees store credentials. The AI inference is paid by the user's own Claude/ChatGPT subscription (no extra cost on Bravos). Available from the Starter plan and during the 7-day PRO trial; requires a paid Claude or ChatGPT subscription, which is where connectors live ## Blog — Core guides - [Chatbot & AI statistics for customer service](https://bravos-ai.com/en/blog/chatbot-statistics): a rigorously-sourced reference of 50+ chatbot and AI statistics, each verified at its primary source with the year and a link (analyst firms, company reports, academic studies, court rulings), not blog-citing-blog. Covers adoption and market size, customer service impact, cost and ROI, consumer sentiment (including the reluctance: 64% of customers would prefer companies did not use AI in customer service, per Gartner 2024), e-commerce and messaging, the generative-AI shift, real limits and risks (the Air Canada chatbot liability ruling; LLM hallucination benchmarks), language/multilingual, and regional data for Spain, the EU and Latin America. Explicitly distinguishes measured figures from projections, and the pre-ChatGPT rule-based era from the current LLM era. Key figures: 75% of CX leaders expect 80% of interactions resolved without humans (Zendesk 2025); 65% of organizations use generative AI regularly (McKinsey 2024); 70.22% average cart abandonment (Baymard); WhatsApp 3B+ monthly users (Meta 2025); 76% of consumers prefer buying in their own language (CSA Research). Available in Spanish (/blog/estadisticas-chatbots-ia), English (/en/blog/chatbot-statistics) and Brazilian Portuguese (/pt/blog/estatisticas-chatbots-ia) - [AI customer service chatbot: how it works, costs, and how to start](https://bravos-ai.com/en/blog/ai-customer-service): plain-language 2026 guide to AI customer service chatbots for businesses — what an AI customer service chatbot is (a bot on your website or WhatsApp that answers in natural language from your own business data, not generic internet knowledge), why adoption jumped from 39% to 66% of support teams in a year (Salesforce State of Service 2026), how it actually works, what it can and cannot do (honest about hallucination and the hybrid AI+human model), real costs compared (Zendesk, Intercom Fin per-resolution, Tidio/Lyro vs Bravos flat rate from €19/mo with AI included), a step-by-step setup, and the EU AI Act transparency rules. Available in Spanish (/blog/ia-atencion-al-cliente), English (/en/blog/ai-customer-service) and Brazilian Portuguese (/pt/blog/ia-atendimento-ao-cliente) - [AI agent or chatbot](https://bravos-ai.com/en/blog/ai-agent-or-chatbot): real differences, costs, agent washing, and which one your business needs - [Why your chatbot hallucinates](https://bravos-ai.com/en/blog/why-your-chatbot-fails): how RAG works, preventing AI from making things up - [How to write a system prompt for your business chatbot](https://bravos-ai.com/en/blog/system-prompt-business-chatbot): practical guide to writing the instructions that turn a generic LLM into your business assistant. The 7 building blocks of a real system prompt (role, business context, data, rules, tone, format, examples), a battle-tested base template ready to copy, sector adaptations (e-commerce, restaurants, professional services, SaaS), the most common mistakes (telling the bot to "be helpful" instead of giving rules, contradictory instructions, forgetting how to handle off-topic questions), how to iterate on a prompt without breaking what already works, and the length paradox (longer is not better — past 1,000 tokens the model starts ignoring instructions). On Bravos AI, the Starter plan adds custom_instructions on top of the platform's base prompt; Pro and Enterprise fully replace the system prompt with their own - [PixelRAG explained for business chatbots](https://bravos-ai.com/en/blog/pixelrag-business-chatbot): what PixelRAG actually is, how it works under the hood, and whether it makes sense for a business chatbot. Honest read of the paper (Wang et al, UC Berkeley + Princeton + EPFL + Databricks + Renmin University, June 2026, Apache 2.0, not yet on arXiv) for people evaluating it. The "+18% accuracy" headline is the peak on EVQA (visual Wikipedia infographics); on text Wikipedia QA it's +2.8 to +7.2 pp. The "10x fewer tokens" applies in ReAct multi-step agents, not single-turn chatbots. PixelRAG only validated in English. With GPT-4o as reader (standard SaaS) it's 7x more expensive than text-based RAG. Authors themselves acknowledge English-only, lost hyperlinks, harder content moderation. Covers when PixelRAG actually makes sense (3 niche profiles: historical archives, technical diagrams with engineering schematics, complex financial/legal PDFs), when it doesn't (most business chatbots — FAQs, catalogs, policies, descriptions), and real alternatives that already exist in production (AWS Textract, Unstructured.io, Claude with direct vision, ColPali). Includes a 5-question test to decide if it applies. Bravos AI uses text-based RAG (pgvector + text-embedding-3-small + BGE reranker) + structured SQL filtering, working in English/Spanish/12+ languages, latency under 2s - [Vectorless RAG: do you still need a vector database?](https://bravos-ai.com/en/blog/vectorless-rag): honest explainer of the mid-2026 "RAG is dead / vectorless RAG" debate. Explains from scratch what a vector database and embeddings are, how an AI chatbot finds information (the standard chunk-embed-store-retrieve-generate pipeline), and the difference between semantic search (vector similarity, good for fuzzy meaning-based questions) and keyword/structured search (good for exact values: codes, prices, sizes, filters). What vectorless RAG actually is: retrieval without a vector database, using document structure + LLM reasoning (PageIndex by VectifyAI, a hierarchical table-of-contents tree; "similarity is not relevance") or plain keyword search in an agentic loop. Figures flagged by source: PageIndex claims 98.7% on FinanceBench (vendor-reported); the paper "Keyword search is all you need" (Subramanian et al, Amazon, arXiv 2602.23368) reports agentic keyword search reaching over 90% of traditional RAG performance without vectors; Chroma's "context rot" (accuracy degrades as context grows). Verdict: RAG is not dead — the reflex of embedding everything into a vector database by default is fading. Includes a decision guide (keep vectors for large fuzzy corpora; go vectorless/hybrid for structured docs and exact-value questions) and an honest vector-vs-vectorless table. Bravos AI already runs hybrid: structured search for catalogs and exact-value questions, semantic retrieval for free-form text - [What is an MCP and how to manage your chatbot from Claude or ChatGPT](https://bravos-ai.com/en/blog/what-is-mcp): plain-language explainer of the Model Context Protocol (MCP) — the open standard Anthropic published in late 2024, now also used by ChatGPT and other tools — written for non-developers using the USB analogy (one common connector so any AI assistant can talk to any outside service). Covers what an MCP is, what an MCP server is (a program that exposes a closed list of named tool actions the AI can request, authorized via OAuth, no code or keys), what it is for (getting the AI out of its training-data bubble to act on your data in real time), MCP vs API (the MCP is a layer on top so a model discovers and uses actions in natural language; many MCP servers are just a regular API wrapped for a model), and that it works with any MCP-compatible assistant, not only Claude/ChatGPT. Then the concrete case: Bravos AI just launched its own MCP server (mcp.bravos-ai.com), so from Claude, ChatGPT or any MCP client you can analyze your bot's conversations and leads, audit and edit the knowledge base (create/edit/delete text entries and URLs — recoverable), rewrite the bot's instructions and tone, change widget texts/appearance/languages, and trigger store catalog syncs — all from the chat, with a before/after returned on every change and everything gated by your plan. What it does NOT do, on purpose: create the bot (done in the panel, it touches plan and billing), upload or delete files (PDF/images/audio are irreversible — it reads them and recommends, you delete in the panel), or touch store credentials. The AI analysis is paid by your own Claude/ChatGPT subscription (no extra cost on Bravos); available from the Starter plan and during the 7-day PRO trial; requires a paid Claude or ChatGPT subscription, which is where connectors live. Honest that Bravos is not the only platform offering this — the differentiator is the package (real structured search, price, native Spanish/English/Portuguese, WhatsApp included). Available in Spanish (/blog/que-es-un-mcp), English (/en/blog/what-is-mcp) and Brazilian Portuguese (/pt/blog/o-que-e-mcp) - [What is context engineering, the skill that replaced prompt engineering](https://bravos-ai.com/en/blog/context-engineering): plain-language explainer of "context engineering", the term that in 2026 displaced "prompt engineering" as the headline AI skill (Gartner: "context engineering is in, prompt engineering is out"). Covers what it is (per Anthropic's Sept 2025 guide: the practice of curating and maintaining the optimal set of tokens/information a model sees during inference), where the term came from (popularized by Tobi Lütke, CEO of Shopify, mid-2025; amplified by Andrej Karpathy, a founding member of OpenAI; formalized by Anthropic), and how it differs from prompt engineering — with the honest take that prompt engineering is NOT dead but is now one layer (the instructions) inside the bigger discipline. Breaks down the pieces that make up "context": system instructions, retrieved external data (RAG), tools, examples, message history and memory — and the core principle of fitting the smallest set of high-signal tokens into a finite context window (avoiding "context rot"). Explains why context engineering is what reduces hallucinations (a model invents answers mostly when the real information is not in its context), and the takeaway for a business chatbot: a chatbot is only as good as the context it is given — the differentiator is rarely the model (all serious chatbots run on the same frontier models) but whether the right piece of your business is put in front of the model at the right moment. With Bravos AI the platform does the context engineering for you: you upload your knowledge (documents, website, catalog) and it assembles the right context per question — the right information retrieved, exact answers on structured data (catalogs, prices, sizes, stock), and your instructions and conversation kept in context — no prompts to write, no pipelines to build. Cites the real primary sources (Anthropic, Karpathy, Gartner). Same slug across languages — available in Spanish (/blog/context-engineering), English (/en/blog/context-engineering) and Brazilian Portuguese (/pt/blog/context-engineering) - [What is an AI agent, and what companies actually use them for](https://bravos-ai.com/en/blog/what-is-an-ai-agent): plain-language explainer of "AI agent" — the most-repeated AI term of 2026 — written for a business owner, not a developer. What an AI agent actually is (per Anthropic's Building Effective Agents: a system where the LLM dynamically directs its own steps and tool use, versus a workflow with a fixed pre-programmed path; in short, a chatbot replies and an agent acts), how it works (perceive → reason → act → check, in a loop with tools), and the practical distinction between chatbot, assistant and agent (the assistant answers with your data and takes bounded actions while you stay in control; most of what is sold as an "agent" for a small business is really an assistant). Then what companies actually use agents for in 2026, with real data: customer service (two in three service teams already use AI agents, up from 39% to 66% in a year — Salesforce State of Service 2026), coding, research/summarization and sales. The standard IBM/textbook taxonomy (simple reflex, model-based, goal-based, utility-based) plus single vs multi-agent. And the honest part most vendors skip: fake "agents" (rule-based chatbots rebranded), Gartner's prediction that over 40% of agentic AI projects will be canceled before 2027, and the danger of an agent that improvises when it does not know. A decision guide (agent vs assistant vs chatbot; keep a human in the loop for delicate or irreversible tasks) landing on Anthropic's advice: start with the simplest thing that works, add autonomy only when the problem needs it. Honest positioning: Bravos AI is not an autonomous agent but a conversational assistant grounded in your data — which, for customer service, is exactly what most businesses should use. Available in Spanish (/blog/que-es-un-agente-de-ia), English (/en/blog/what-is-an-ai-agent) and Brazilian Portuguese (/pt/blog/o-que-e-um-agente-de-ia) - [What is DeepSeek Harness, the open-source coding agent where everything is a plugin](https://bravos-ai.com/en/blog/what-is-deepseek-harness): plain-language explainer of DeepSeek Harness (dsh), DeepSeek's open-source coding agent (MIT license) launched August 13, 2026, which passed 177,000 GitHub stars in one week. What a coding-agent "harness" is (the software layer that turns a model into an agent that reads files, runs commands and works through long tasks — the model is the brain, the harness the body; the same model performs differently per harness). Its defining trait: everything is a plugin — the model, tools, memory, sessions, interface, theme and even the agent loop are swappable pieces, powered by a framework called Cordis; plugins install with "dsh plugin add" (a dsh.bundle manifest) and there is a plugin market (dsh-market) inside the app, with hundreds of community plugins in the first week (from serious ones — dsh-model-picker, a real terminal via xterm.js, dsh-file with Monaco, MCP management, a GitHub issue/PR picker — to absurd ones like desktop pets and a TikTok-style video feed, which is what proves everything really is a plugin). Covers the four modes (Standard, Code, Minimal, Creator), full transparency of the agent's reasoning via an append-only trajectory log (unlike Anthropic and OpenAI, which hide raw chain-of-thought to prevent model distillation), how it compares to Claude Code and Codex (more open and customizable but more developer-oriented; Claude Code's system prompt once ran ~10,000 tokens vs ~200 in others), how to install (npx @deepseek-ai/dsh web, web UI on 127.0.0.1:3080, or from source with pnpm; free, connect any model by API), and the security risk the hype skips: installing a plugin runs third-party code with your own permissions, and the agent's tool approvals do not sandbox it. Available in Spanish (/blog/que-es-deepseek-harness), English (/en/blog/what-is-deepseek-harness) and Brazilian Portuguese (/pt/blog/o-que-e-deepseek-harness) - [AI chatbot product catalog](https://bravos-ai.com/en/blog/ai-chatbot-product-catalog): why most chatbots fail with structured data and how SQL+AI solves it - [Rule-based vs AI chatbot](https://bravos-ai.com/en/blog/rule-based-chatbot-vs-ai-chatbot): comparison, when to choose each - [Contact form vs chatbot](https://bravos-ai.com/en/blog/contact-form-vs-chatbot): data-backed analysis of conversion rates - [Proactive vs reactive chatbot](https://bravos-ai.com/en/blog/proactive-vs-reactive-chatbot): when to ask vs wait - [Chatbot pricing in 2026](https://bravos-ai.com/en/blog/ai-chatbot-pricing): real prices for Tidio, Intercom, Zendesk, Bravos and alternatives - [Do Tidio, Crisp and Freshchat's free plans include real AI?](https://bravos-ai.com/en/blog/tidio-crisp-freshchat-free-plan): investigation — we asked each platform's own official chatbot about its free plan. Tidio: 50 lifetime AI conversations (not renewable). Crisp: €0 AI credits on free, AI starts on Mini. Freshchat: AI only from Growth tier. Bravos AI does not offer a usage-limited free tier — instead a 7-day full PRO trial with all features and lead capture included, no charge if you cancel before day 7 - [Keep your chatbot data up to date without Zapier or n8n](https://bravos-ai.com/en/blog/chatbot-data-sync): 5 native sync paths at Bravos AI (Google Sheets, Shopify real-time webhooks, WooCommerce real-time webhooks + nightly reconciliation, PrestaShop daily, custom webhook with HMAC). Real cost of Zapier/n8n middleware including task tiers and 15-min polling delays in Zapier Free. Comparison with Tidio (Plus $749/mo for URL auto-sync, re-sync wipes attached Q&A), Crisp, Intercom Fin (24h resync of Confluence/Notion/Guru, not Sheets natively) - [Multilingual chatbot](https://bravos-ai.com/en/blog/multilingual-chatbot): 12+ languages with one content set - [Chatbot legal liability](https://bravos-ai.com/en/blog/chatbot-legal-liability): Air Canada case, EU AI Act - [Lead capture with chatbot](https://bravos-ai.com/en/blog/chatbot-lead-capture): how to ask for visitor data without being invasive - [WhatsApp chatbot — the complete guide for 2026](https://bravos-ai.com/en/blog/whatsapp-chatbot): the umbrella guide covering everything about WhatsApp + AI chatbots — WhatsApp Business app vs WhatsApp Business API (the most common confusion in the market), what an AI chatbot can and cannot do on WhatsApp in 2026, the truth about "free" WhatsApp chatbots (none with real AI exist), how Meta's signup, templates and 24-hour window actually work, real pricing ranges ($23-2,500/month depending on need), AI vs button-flow chatbots, sector use cases with sample conversations, common launch mistakes, when to build it yourself with n8n/Make/Zapier vs use a managed platform. Bravos AI as a verified Meta Tech Provider — no markup over Meta's per-message cost, native Shopify, PrestaShop, Resales Online and custom webhook integrations, exact catalog search with multi-filter combinations (size, color, stock, price range), multilingual out of the box, from $23/month - [Best WhatsApp chatbot 2026 — 13 platforms compared](https://bravos-ai.com/en/blog/best-whatsapp-chatbot): honest comparison of 13 relevant WhatsApp chatbot platforms in 2026 (Wati, Tidio with Lyro, Respond.io, Landbot, HubSpot, Intercom Fin, ManyChat, Crisp, AiSensy, Interakt, SleekFlow, 360dialog, Twilio). Prices verified directly on each platform's official page, real LLM-based conversational AI vs flow-based bots with a single AI step, the hidden Meta surcharge most platforms do not disclose (2.8-40% markup on Meta's per-message cost depending on platform and tier), Spanish-language support availability (only Bravos AI, Landbot and HubSpot have confirmed Spanish teams), and which one fits each business case. Bravos AI from $23/month — the only platform with real conversational LLM AI in that price tier ## Blog — Use cases - [Dental clinic chatbot](https://bravos-ai.com/en/blog/dental-clinic-chatbot): treatments, pricing, accepted insurance, GDPR/HIPAA-aware lead capture, appointment requests in 12+ languages - [Law firm chatbot](https://bravos-ai.com/en/blog/law-firm-chatbot): automating client intake for a law firm without giving legal advice. How the system prompt sets the boundary — the bot answers only with the firm's own information (RAG), can give general legal information but never applies it to the visitor's specific case, qualifies the inquiry (practice area, matter type), filters out-of-area, and captures the contact with the whole conversation attached (not just a form field). Confidentiality, privilege and the jurisdiction's ethical rules; what it does NOT do (no case evaluation, no drafting, no deciding who is right). Works on the website widget and WhatsApp. Real personal-injury intake example. PT/Brazil version covers OAB advertising rules and LGPD - [Restaurant chatbot](https://bravos-ai.com/en/blog/restaurant-chatbot): menu, allergens, reservations, multilingual - [Ecommerce chatbot](https://bravos-ai.com/en/blog/ecommerce-chatbot): reducing cart abandonment, catalog filtering - [WooCommerce chatbot](https://bravos-ai.com/en/blog/woocommerce-chatbot): what actually decides whether a chatbot works on a WooCommerce store — answering per variation (size/color with own price and stock), distinguishing "4 left" from "in stock" from "available on backorder", quoting sales as before/after pairs, and matching the price the store displays tax-wise. Why a plugin inside WordPress is the wrong architecture for catalog AI (load on shared hosting, capped AI quality, one more thing to break) vs a service connected through the WooCommerce REST API. Step-by-step connection without coding, real-time updates plus nightly reconciliation, WPML/Polylang handling, Elementor/Divi shortcode cleaning - [Real estate chatbot](https://bravos-ai.com/en/blog/real-estate-chatbot-resales-online): property listings, filters, Resales Online - [Small business chatbot](https://bravos-ai.com/en/blog/chatbot-for-small-business): practical guide for SMBs - [Ad campaigns chatbot](https://bravos-ai.com/en/blog/chatbot-ad-campaigns): converting Google Ads and Meta Ads traffic - [Marbella chatbot](https://bravos-ai.com/en/blog/marbella-chatbot): local business focus ## Blog — Tutorials - [Add chatbot to WordPress](https://bravos-ai.com/en/blog/wordpress-chatbot): 5-minute setup guide - [Wix chatbot](https://bravos-ai.com/en/blog/wix-chatbot): an honest guide to AI chat on Wix. What Wix's built-in Smart Chat does (manual mode free; AI mode trained on site content plus hand-written rules) and how its AI-credit system really works, with official numbers: free plan 30 credits/day capped at 120/month, paid plans ~150-1,000 estimated actions per cycle, a pool shared with Wix's other AI tools, no guaranteed credit-to-conversation equivalence (Wix states it literally), and the AI chat switching to offline mode when credits run out. When the built-in chat is enough, when a specialized chatbot is needed (exact catalog filtering, unpublished documents, WhatsApp, answering in the visitor's language), a 4-question test to run on any vendor's trial, how to prepare a Wix site so any bot answers better, and the step-by-step install of an external widget (requires a paid Wix plan with a connected domain; Settings → Custom Code). Wix Stores catalog connects via CSV export or a synced Google Sheets document; a direct Wix Stores integration is on the Bravos roadmap - [WhatsApp AI chatbot — how to build one without code](https://bravos-ai.com/en/blog/whatsapp-ai-chatbot): three real paths (n8n/Make/Zapier vs Meta Cloud API vs purpose-built platform), why ChatGPT alone is not enough for business, comparison vs Manychat/Wati/Respond.io, and the truth about Meta's July 2025 pricing (replies to customer-initiated chats are free, no monthly cap). Bravos AI plan from $23/month with direct Meta Tech Provider integration — no Twilio, no middleware fees - [WhatsApp auto reply in 2026 — out-of-office, Business and AI](https://bravos-ai.com/en/blog/whatsapp-auto-reply): three real ways to set up an auto reply on WhatsApp — the static away message in the WhatsApp Business app, rule-based templates, and AI that understands the customer and answers with the business's own data. Includes the difference between automating personal WhatsApp (ban risk with unofficial apps like WhatsAuto) vs business WhatsApp (the official Meta path), setup steps, and Meta's July 2025 pricing change (replies to customer-initiated chats are free). Bravos AI is a verified Meta Tech Provider — no Twilio, no 60-day token refresh, no middleware fees - [n8n, Make and Zapier for WhatsApp AI chatbots — the honest cost breakdown](https://bravos-ai.com/en/blog/whatsapp-chatbot-n8n-make-zapier): the 9 components you actually need to wire if you want to build a WhatsApp AI chatbot on n8n self-hosted, Make.com or Zapier (not the 2 the YouTube tutorial shows). Real monthly cost breakdown ($21-95/month depending on platform, vector DB and OpenAI usage), real developer hours (25-40h the first time), the Meta Cloud API gotchas tutorials skip (60-day token expiry, media_id download for inbound images/audio, 24-hour window, debugging without proper logs). When DIY actually makes sense (devs with spare time, custom ERP integration, micro volume) and when it does not (non-devs, agencies scaling to multiple clients, mission-critical 24/7). Comparison vs Bravos AI as a verified Meta Tech Provider with all 9 components managed on the platform side ## Technology - **LLM**: OpenAI (GPT-4.1-mini for chat, GPT-5.4-nano for filter extraction) - **Embeddings**: text-embedding-3-small (1536 dims) - **RAG**: PostgreSQL + pgvector, hybrid search (semantic + keyword + structured) with Reciprocal Rank Fusion - **Reranker**: Voyage AI rerank-2.5 - **Multimedia**: AWS Bedrock (Claude) for image analysis, AWS Transcribe for audio, AWS Textract for PDFs - **Backend**: FastAPI (Python), SQLAlchemy, Uvicorn - **Frontend**: Next.js 15, React, Tailwind CSS, next-intl - **Storage**: AWS S3 - **Payments**: Stripe - **Email**: Resend - **Hosting**: Hetzner ## Pricing Bravos AI does not have a free tier. New customers start with a 7-day full PRO trial. Prices are billed in the customer's currency: EUR, USD, BRL or GBP (Starter €19 / $23 / R$ 119 / £16 per month; Pro €49 / $59 / R$ 299 / £42 per month). - **7-day PRO trial**: full PRO plan for 7 days, no commitment. We notify you before charging. If you cancel before day 7, no charge. Card required to start the trial (for verification and seamless continuation if you choose to stay). - **Starter**: €19/month ($23 / R$ 119 / £16) — 1 chatbot on web and WhatsApp, unlimited messages, AI included, all integrations (PrestaShop, Shopify, Resales…), MCP connector to manage the bot from Claude/ChatGPT, image and audio in conversations, custom instructions, email support within 48h - **Pro**: €49/month ($59 / R$ 299 / £42) — 3 chatbots on web and WhatsApp, unlimited messages across all bots, more advanced AI for more accurate replies, lead capture inside the chat, 100% customizable system prompt and tone, more capacity per bot (URLs, files, multimedia), priority support within 24h - **Enterprise**: custom — tailored chatbots and limits, premium AI model of choice, guided setup with the Bravos team, guaranteed response time per contract ## Key URLs - Website: https://bravos-ai.com - App: https://app.bravos-ai.com - API: https://api.bravos-ai.com - Sitemap: https://bravos-ai.com/sitemap.xml ## Legal - [Privacy Policy](https://bravos-ai.com/en/legal/privacy) - [Terms of Service](https://bravos-ai.com/en/legal/terms) - [DPA](https://bravos-ai.com/en/legal/dpa) --- # Blog articles (full content — EN) Below is the full content of all blog articles in English, sorted newest first. --- # What Is DeepSeek Harness? The Open-Source Coding Agent Where Everything Is a Plugin **Category:** Practical guide | **Read time:** 13 min | **Date:** Aug 22, 2026 | **URL:** https://bravos-ai.com/en/blog/what-is-deepseek-harness **DeepSeek Harness is DeepSeek's open-source coding agent: the software layer that wraps an AI model and turns it into an assistant that can read your files, run commands and keep working through long tasks on its own.** It launched on August 13, 2026, is free (MIT license), and passed **177,000 GitHub stars** in a single week. What sets it apart isn't the model — it's the architecture: **literally everything (the model, the tools, the interface, even the agent loop) is a plugin you can swap.** This guide explains **what DeepSeek Harness is** with real detail, not slogans: what "everything is a plugin" actually means (with concrete examples of what the community has already built), how plugins install, its modes, how it compares to **Claude Code and Codex**, the security risk almost nobody mentions, and how to get started. Sources at the bottom. Contents - [What is a coding-agent "harness"](#what-is-harness) - [What is DeepSeek Harness](#what-is) - ["Everything is a plugin": what it really means](#everything-plugin) - [Real examples: what the community already built](#examples) - [The four modes](#modes) - [Full transparency: the trajectory log](#transparency) - [The risk the hype skips](#security) - [DeepSeek Harness vs Claude Code and Codex](#vs-claude-code) - [How to install and get started](#get-started) - [Why it blew up](#why) - [FAQ](#faq) - [Sources](#sources) ## What is a coding-agent "harness" An AI model like DeepSeek, GPT or Claude, on its own, only writes text. For it to **actually code** — open your files, run the terminal, search the project and keep going until the job is done — it needs a layer of software around it that gives it those hands. That layer is the **harness**. Claude Code, Codex and Cursor are harnesses. **The model is the brain; the harness is the body.** A coding agent is, at heart, a type of [AI agent](https://bravos-ai.com/blog/que-es-un-agente-de-ia) specialized in programming: it doesn't just answer, it plans steps, uses tools and acts until the goal is met. And here's why the harness matters so much: **the same model performs differently depending on the harness it runs through.** The harness decides what the model sees, which tools it has and how much each task costs. That's why it has become the battleground where AI companies now compete. ## What is DeepSeek Harness **DeepSeek Harness** (short: **dsh**) is an **open-source coding agent**, MIT-licensed, that runs both as a web app and headless (no interface, for automation). Instead of shipping a closed tool with a fixed set of features — like Claude Code or Codex — DeepSeek built **the whole harness out of interchangeable pieces**, on top of a framework called **Cordis** designed for exactly that: mounting and unmounting components on the fly. A bit of jargon: a **plugin** is a module you plug in and remove to add or change a capability. A **language model** (LLM) is the AI behind tools like ChatGPT, Claude or DeepSeek. *Headless* means "no graphical interface": you drive it with commands or scripts, handy for automation. ## "Everything is a plugin": what it really means This is the heart of DeepSeek Harness, and it's worth explaining properly because it's the whole point. Normally a tool like Claude Code is **monolithic**: it comes with its model, its interface and its way of doing things, and if you don't like something, tough. DeepSeek Harness is the opposite: **every part of the program is a piece you can remove, swap or replace** from configuration, without touching the core code. "Every part" is literal. The things that are welded shut in other tools are plugins here: - **The model:** it uses DeepSeek by default, but you can wire in another provider by swapping a plugin. - **Tools and skills:** browser control, voice, vision, code review, Git integration… all added as plugins. - **Memory and sessions:** how it remembers and where it stores history is swappable too. - **The interface and the look:** from the visual theme to the web UI itself. - **The agent loop:** even the logic of how the agent thinks-acts-repeats is a plugin. It's the same idea behind standards like [MCP](https://bravos-ai.com/blog/que-es-un-mcp) — plugging new capabilities into an agent — only DeepSeek Harness takes it to the extreme: even its own skeleton is replaceable. The model is the soul; the harness is what turns it into an agent acting in a real environment: it manages environment interaction, tool use and long-running execution. ## Real examples: what the community already built The best way to grasp "everything is a plugin" is to see what people have already made. In its **first week** alive, the community shipped **hundreds of plugins**. Technically, a plugin is a package that declares a dsh.bundle manifest and installs with one command: A sample of the **serious** stuff you can already bolt on, to show the range: - Switch models or watch your API spend: dsh-model-picker , dsh-balance . - A **real terminal** inside the UI (with xterm.js ): terminal . - A VS Code-style file explorer with a Monaco editor: dsh-file . - Manage MCP servers from settings: dsh-plugin-setting-mcp . - Search GitHub issues and pull requests without leaving the chat: dsh-github-picker . - A keyboard command palette: dsh-spotlight . A pomodoro timer: dsh-pomodoro . And a sample of the **absurd**, which is exactly what proves *everything* is a plugin: a desktop pet ( dsh-desk-pet ), a pixel whale that taps the glass and blinks the tab title when the agent is waiting for your approval ( whale-on-desk ), or a TikTok-style video feed in a floating window ( dsh-plugin-video-player ). If that can be built as a plugin, anything can. And since adding plugins by hand is clunky, there's already a **plugin market inside the program itself** ( dsh-market ): a searchable screen with one-click install and theme switching. All of this in the first week — a sign of how fast the community moved. ## The four modes DeepSeek Harness ships with four out-of-the-box configurations, each for a different job: - **Standard:** the full coding agent — file editing, terminal, search, skills. The everyday setup. - **Code:** the above plus a TypeScript SDK to orchestrate multi-step tasks in code. - **Minimal:** a two-tool agent (bash + editor), built for clean benchmarking. - **Creator:** for **building your own plugins by chatting** — you describe what you want the agent to be able to do, it asks questions and builds the plugin for you. ## Full transparency: the trajectory log DeepSeek Harness keeps a **complete, ordered record of everything the model sees**: system prompts, its reasoning, every tool call and result, and every bit of context injected. You can inspect it, pause it, fork it and replay it step by step. And here's the juicy bit: **Anthropic and OpenAI deliberately hide that raw reasoning** — out of fear it'll be used to copy (distill) their models. DeepSeek does the opposite and shows all of it. For anyone who wants to understand *why* the agent did what it did — or tune its behavior by editing the [system prompt](https://bravos-ai.com/blog/system-prompt-chatbot-empresa) — that's a huge difference. All that instruction text steering the agent is exactly what [context engineering](https://bravos-ai.com/blog/context-engineering) works on. ## The risk the hype skips Everything being a community plugin has a B-side almost no article mentions, and it matters: **installing a plugin runs third-party code on your machine, with your own permissions.** That plugin can read your files, use your credentials and reach the network. Watch out for one misconception: the agent's **tool approvals** (that "allow this action?" popup) **do not sandbox a plugin's code**. And being on a community "awesome" list — the curated resource lists GitHub is full of — **is not a security review**. Before installing a plugin you don't know, read its source, and try it somewhere that doesn't hold your keys or sensitive data. ## DeepSeek Harness vs Claude Code and Codex The comparison is unavoidable — plenty of people search straight for a **Claude Code alternative** or a Codex alternative — so here's the honest version: Claude Code / Codex Polished, closed, ready-to-use tools. Plug in and go, but you only change what they let you change. DeepSeek Harness Less "finished product", more platform to build your own agent. Everything is modifiable, but it asks for more tinkering (and care with plugins). Put plainly: DeepSeek Harness is more open and powerful for anyone who wants to **customize everything**, but for that same reason it's more **developer-oriented** than something you just want to work untouched. As a measure of how much harness design matters: Claude Code's system prompt once ran to roughly **10,000 tokens** (recently trimmed by 80%), while other harnesses run on barely 200. Those choices change performance and cost even when the underlying model is the same. ## How to install and get started Getting started is straightforward. You need Node.js, and one command launches the web UI (by default at http://127.0.0.1:3080 ): You can also clone the repo and run it from source with pnpm : Since it's **free and open source**, there's no license to pay to try it; what can cost money is the model you connect over an API, depending on the provider. Remember it's in *developer preview*: the docs warn, in capital letters, that **there will be compatibility-breaking changes**. It's an early build to experiment with, not (yet) a stable tool for production. ## Why it blew up Three reasons explain 177,000 stars in a week. One: **the harness has become AI's new battleground** — models look more and more alike, so the difference is made by the layer that wraps them. Two: "everything is a plugin" thrilled developers, and within days a **whole ecosystem** appeared — hundreds of plugins, a market, community lists, desktop clients. Three: being **free, open and from DeepSeek** gives it obvious appeal against the closed tools from the big players. It's not the first open-source AI project to blow up out of nowhere. We recently explained [PixelRAG](https://bravos-ai.com/blog/pixelrag-chatbot-empresarial), another case of a technical idea going viral in days. It's the 2026 pattern: when something suddenly opens a new way of doing things, the community piles in. ## FAQ ### What is DeepSeek Harness? It's DeepSeek's open-source coding agent: a software layer (a "harness") that wraps an AI model and turns it into an assistant that can read files, run commands and work on its own through long tasks. Its distinctive trait is that everything — model, tools, interface, memory, even the agent loop — is a swappable plugin. It launched on August 13, 2026 under an MIT license. ### What does "everything is a plugin" mean? That every part of the system is a piece you can change, remove or replace from configuration without touching the core code: the model, the tools, memory, sessions, the interface, the theme and even the agent loop. Plugins install with dsh plugin add , and there's even a market inside the program itself. Under the hood it's made possible by a framework called Cordis. ### Is DeepSeek Harness free? Yes, it's free and open source under an MIT license: there's no license to pay to install or use it. What can cost money is the model you connect over an API, depending on the provider you choose. ### Is DeepSeek Harness a Claude Code or Codex alternative? Yes, it does the same job — turning a model into a coding agent — but with a different philosophy: instead of a closed, ready-to-use tool, it's a platform where everything is a modifiable plugin. It's more open and customizable, but also more geared to developers who want to tinker. ### Is it safe to install DeepSeek Harness plugins? With care. Installing a plugin runs third-party code on your machine with your permissions: it can read your files and use your credentials, and the agent's tool approvals do not sandbox that code. Review the source before installing plugins you don't know, and test them in an environment without sensitive data. ## Sources - **DeepSeek — Harness (official docs):** [deepseek.com ](https://deepseek.com/harness/en/) — what it is, the plugin architecture, modes, install and the trajectory log. - **DeepSeek Harness on GitHub:** [github.com/deepseek-ai ](https://github.com/deepseek-ai/deepseek-harness) — source code, MIT license, install (npm and from source with pnpm) and the developer-preview notice. - **Awesome DSH Plugin (community list):** [github.com/awesome-dsh-plugin ](https://github.com/awesome-dsh-plugin/awesome-dsh-plugin) — plugin categories and real examples, install with dsh plugin add , and the security warning about third-party code. - **The Register — on DeepSeek's harness:** [theregister.com ](https://www.theregister.com/ai-and-ml/2026/08/14/deepseeks-innovative-harness-treats-everything-as-a-plug-in/5288095) — why the harness is the new battleground, and comparison with Claude Code, Codex and others. --- # What Is an AI Agent? And What Companies Actually Use Them For **Category:** Practical guide | **Read time:** 14 min | **Date:** Aug 11, 2026 | **URL:** https://bravos-ai.com/en/blog/what-is-an-ai-agent **An AI agent is an artificial intelligence system that pursues a goal on its own:** it decides what steps to take, uses tools (searches the web, calls other apps, taps into your systems), checks the result and tries again, with little human input. The difference from a chatbot is simple: **a chatbot replies; an agent acts.** "AI agent" is the most-repeated term of 2026, and there's so much noise around it that it's hard to tell what one actually is and whether it's useful for your business. The pages that rank at the top of Google are either technical encyclopedias written for developers, or brochures selling hype. This guide is the opposite: what an AI agent really is in plain English, how it works, **what companies actually use them for**, the part almost nobody mentions (why so many projects fail), and how to decide whether you need an agent, an assistant, or simply a good chatbot. With real sources, not headlines. Contents - [What an AI agent is](#que-es) - [How it works](#como-funciona) - [Agent vs assistant vs chatbot](#vs-chatbot) - [What companies use AI agents for](#para-que) - [Types of AI agents (with examples)](#tipos) - [The honest truth](#sin-humo) - [Does your business need an AI agent?](#necesitas) - [Bravos AI: the assistant, done right](#bravos) - [In short](#resumen) - [Frequently asked questions](#faq) - [Sources](#fuentes) ## What an AI agent is An **AI agent** is a program that uses a language model (the engine behind ChatGPT, Claude or Gemini) to *achieve a goal*, not just to answer a message. You give it an objective —"book me a flight", "put together this month's sales report", "resolve this support ticket"— and the agent breaks that goal into steps, decides which tools it needs, uses them, looks at what came back and corrects course until it's done. The most useful distinction comes from Anthropic (the company behind Claude), in its reference guide for engineers. It separates two things almost everyone lumps together: - **Workflow:** a language model and some tools that follow *steps you programmed in advance*. The path is fixed. - **Agent:** the model *decides its own steps and which tools to use*, on the fly, based on what it finds along the way. There's no predefined path. Agents ... are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks. A line worth remembering: **a chatbot answers you; an agent does the work.** A chatbot says "to change your order, go to My Account"; an agent would go in and change it itself. A bit of jargon: a **language model** (or LLM) is the AI behind tools like ChatGPT or Claude. A **tool**, in this context, is anything the agent can "use": a web search, a calculator, your store's API, your calendar. An agent is the LLM plus the ability to use those tools in a loop. ## How an AI agent works Under the hood, an agent repeats a cycle until it meets the goal or runs out of room. Four moves: - **Perceives:** gathers the information it needs —your message, data from your systems, the result of a search. - **Reasons and plans:** with the language model, it works out what's being asked, splits it into steps and decides the next one. - **Acts:** carries out that step using a tool (query a database, send an email, call an API). - **Checks and repeats:** looks at the result of its action and, if it's not finished, plans again. That loop is what sets an agent apart from a single reply. That "loop with tools" is the key. A chatbot gives one answer and stops; an agent can chain ten steps —search, compare, decide, execute, verify— to complete a task that would take a person half an hour. ## Agent vs assistant vs chatbot These three terms get used as synonyms, and they're not. The practical difference is *how much each one does on its own*: Chatbot Answers questions. You act afterwards. "Do you ship to Scotland?" → it tells you. Assistant Answers with *your* data and takes bounded actions (search your catalogue, capture a lead). You stay in control. Agent Pursues a goal over several steps and makes decisions on its own. More powerful, harder to control. Most of what's sold as an "agent" for a small business is really a very capable **assistant** — and that's often exactly what you need. We go deeper in our comparison of [AI agent vs chatbot](https://bravos-ai.com/blog/agente-ia-o-chatbot). ## What companies use AI agents for This is the part the generic guides skip, and the one you actually care about: **it's not theory, they're already in use.** These are the uses with the most traction in 2026: - **Customer service** — triaging tickets, tracking orders, resolving routine questions and escalating the hard stuff to a person. It's one of the most mature uses: per [Salesforce's State of Service 2026](https://www.salesforce.com/news/stories/ai-service-agents-improve-customer-satisfaction/) (a survey of 3,075 professionals), **two in three service teams already use some kind of AI agent** —up from 39% to 66% in a year—. - **Coding** — generating, documenting and reviewing code. It's where agents are having the clearest impact inside software companies. - **Research and summarization** — gathering information from several sources and synthesizing it. One of the most common uses, precisely because it's bounded and verifiable. - **Sales** — qualifying leads, following up and drafting first-touch emails. Here the agent sets the table and a person closes. The pattern is clear: **agents win on repetitive tasks that have rules and a checkable result.** The more open-ended and sensitive the task (a negotiation, a tense complaint), the less autonomous it should be. ## Types of AI agents (with examples) There's a standard classification —the one IBM and the AI textbooks use— that orders agents from least to most sophisticated by how they make decisions. In plain terms, with business examples: - **Simple reflex:** reacts to what it sees right now with fixed rules, with no memory of what came before. Example: a bot that, for every incoming email, tags it and routes it to the right department based on keywords. - **Model-based:** keeps a picture of what has been happening, so it decides better even when it can't see everything at once. Example: an order agent that remembers which orders already shipped and acts accordingly. - **Goal-based:** has an objective and plans the steps to reach it. Example: given "put together the weekly sales summary", it queries the system, cross-references the data and writes the report. - **Utility-based:** not only meets the goal, but looks for the *best* way to do it against several criteria (cost, time, risk). Example: picking the shipping option that balances price and speed. And a distinction you'll hear a lot, cutting across the four above: **a single agent** versus **multi-agent** —several specialized agents that coordinate: one searches, another drafts, another reviews—. Multi-agent is powerful for big tasks, but also the most expensive and hardest to maintain. ## The honest truth Here's what the brochures don't tell you, and it matters as much as everything above: - **A lot of "agents" are fake.** Half the industry has rebranded plain old rule-based chatbots and automations as "AI agents". Calling something an agent doesn't mean it decides anything on its own. - **Full autonomy still fails a lot.** Gartner [predicts that over 40% of agentic AI projects will be canceled before 2027](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027), for cost, risk or lack of real value. A loose agent that processes refunds and changes orders unsupervised looks great in a demo and causes problems in production. - **An agent that improvises when it doesn't know is dangerous.** If it doesn't have the information, it should say "I don't know" and hand off, not invent an answer. It's the same hallucination risk as any chatbot, and we cover it in [why your chatbot makes things up](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). And Anthropic's own advice is the most sensible thing you'll read on this: *start with the simplest thing that works, and only add agent complexity when the problem genuinely calls for it*, because an autonomous agent trades reliability, latency and cost for a capability you often don't need. ## Does your business need an AI agent? A practical rule to decide without getting swept up in the hype: - If what you want is to **answer your customers well** (questions, prices, availability, orders) → you need an **assistant grounded in your data**, not an autonomous agent. It's more reliable, cheaper and can be live in an afternoon. - If you have a **repetitive process with clear rules and a checkable result** (classify, summarize, move data between systems) → that's where an agent starts to make sense. - If the task is **open-ended, delicate or irreversible** (complaints, money, decisions with no undo) → always keep a person giving the green light before the agent acts. This is a specific feature the good platforms include —you'll see it named **"human in the loop"**: the agent stops at the sensitive step and waits for your approval instead of executing on its own. Never leave it 100% on autopilot. For the vast majority of businesses —a shop, a law firm, a clinic, a local council— the best first step isn't an autonomous agent: it's a **good conversational assistant that answers with the real information from your business**, on your website and on WhatsApp, without making anything up. Autonomy comes later, when a specific process justifies it. ## Bravos AI: the assistant, done right Everything above leads to a practical conclusion, and it's where Bravos AI comes in. For customer service —the use where agents have the most traction— you almost never need an autonomous agent: you need something that answers your customers with the real information from your business, without making anything up, and that doesn't decide things on its own it shouldn't. That's Bravos, and we'll tell you with the same honesty as the rest of this article: **it's not an autonomous agent** that directs itself; it's a **conversational assistant**, done right. Which, for this job, is exactly what you want. In practice, the difference shows in the details. You give it two things: your knowledge —your website, your documents, your catalogue— and some [instructions on how to behave](https://bravos-ai.com/blog/system-prompt-chatbot-empresa) (the tone, the rules, what to do when it doesn't know). With that it answers your customers with the exact fact instead of a generic reply. If they ask about a specific product, it searches your catalogue by price, size or real stock, not by "something similar". If it doesn't have the answer, it says so and hands off to a person, instead of hallucinating. If someone shows interest, it captures the lead inside the conversation itself. And it answers the same way on your website and on WhatsApp, with the same knowledge. No more autonomy than the task needs, no less reliability than a customer deserves. It's Anthropic's advice applied to your business: *start with the simplest thing that solves the problem.* For the vast majority of companies, that's a good assistant grounded in your data, not a loose agent that looks spectacular in a demo and causes headaches in production. And if one day a specific task justifies more autonomy, the foundation —answering with your information, without making things up— is already in place. ## In short An AI agent is a system that pursues a goal on its own: it plans, uses tools in a loop and acts, versus a chatbot that only replies. Companies mostly use them for customer service, coding, research and sales —repetitive, checkable tasks—. But there are plenty of fake "agents", full autonomy still fails often, and for most businesses the sensible move is to start with a reliable assistant grounded in your data, not an autonomous agent. The rule, in Anthropic's words: start simple, and add autonomy only when the problem calls for it. ## Frequently asked questions ### What is an AI agent in simple terms? It's an AI program that pursues a goal on its own: it breaks the task into steps, uses tools (search, log into apps, look up data), checks the result and keeps going until it's done. The difference from a chatbot is that a chatbot replies, and an agent acts. ### What are AI agents used for? To automate repetitive tasks that have rules and a checkable result. In companies, the uses with the most traction in 2026 are customer service (triaging and resolving tickets), coding, research and summarization, and sales (qualifying leads and following up). ### What is the difference between an AI agent and a chatbot? A chatbot replies to messages; an agent pursues a goal over several steps and makes decisions on its own using tools. In between sits the assistant: it answers with your data and takes bounded actions, but you stay in control. Most of what is sold as an agent for a small business is really an assistant. ### Does my business need an AI agent? If what you want is to answer your customers well, you need an assistant grounded in your data, not an autonomous agent: more reliable and cheaper. An agent makes sense for repetitive processes with clear rules. For delicate or irreversible tasks, always keep a person reviewing. ### Are AI agents reliable? It depends on the use. On bounded, checkable tasks they work well; on full autonomy they still fail often. Gartner predicts over 40% of agentic AI projects will be canceled before 2027. The key is not to give them more autonomy than the task needs, and never let them make up answers when they don't know. ### Start with what actually moves the needle An AI assistant that answers your customers with the real information from your business, on web and WhatsApp, without making anything up. From €19/mo, with a free 7-day PRO trial. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) ## Sources - **Anthropic — Building Effective Agents:** [anthropic.com ](https://www.anthropic.com/engineering/building-effective-agents) — the reference technical guide: the workflow-vs-agent distinction and the "start simple" principle. - **Salesforce — State of Service 2026:** [salesforce.com ](https://www.salesforce.com/news/stories/ai-service-agents-improve-customer-satisfaction/) — AI agent adoption in customer service (from 39% to 66% in a year; survey of 3,075 professionals). - **Gartner — on agentic AI projects:** [gartner.com ](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) — prediction that over 40% of projects will be canceled before 2027. - **IBM — What are AI agents:** [ibm.com ](https://www.ibm.com/think/topics/ai-agents) — technical reference on the components and types of agents. --- # AI Customer Service Chatbot: How It Works, What It Costs, and How to Start (2026) **Category:** Practical guide | **Read time:** 12 min | **Date:** Aug 10, 2026 | **URL:** https://bravos-ai.com/en/blog/ai-customer-service An **AI customer service chatbot** is a bot on your website or WhatsApp that answers customer questions in plain language, using your own business information — not generic internet knowledge. In 2026 it's no longer an enterprise-only tool: a small business can have one running from €19/month, with no technical team, in less than an afternoon. And it's gone mainstream fast. According to [Salesforce's State of Service 2026](https://cxfoundation.com/news/two-in-three-customer-service-teams-use-ai-agents), the share of customer service teams using AI agents jumped from 39 % to 66 % in a single year — two in three now run at least one. There's a good reason: [IBM](https://www.ibm.com/think/topics/ai-customer-service-chatbots) estimates up to 80 % of support queries are routine — questions that already have an answer somewhere in your docs, website, or past emails, yet most small businesses still answer them by hand, one by one. This isn't a piece about "AI revolutionizing your business." It's a practical guide to what a customer service chatbot actually does, how it works, what it really costs, when it's worth it (and when it isn't), and the fastest way to get one live. ## Why customer service is the first place to automate with AI The smartest place to start with AI isn't your marketing or your sales campaigns — it's customer service. It's the area with the most repetitive, predictable tasks in any business, which makes it the ideal place to get measurable results from day one. And because many of those questions come from people deciding whether to buy, resolving them fast doesn't just save time — it recovers sales you'd otherwise lose. Your team answers the same 10–15 questions over and over. What are your opening hours? Do I need an appointment, or can I just walk in? What documents do I need to apply? Do you ship internationally, and is this in stock? Where's my order — or my case? Each answer takes 2 to 5 minutes — whether you run a shop, a law firm, a public office or an information site. Multiply that by dozens of queries a day. But the real problem isn't the time cost — it's what you lose by not responding fast enough. A study by [MIT](https://web.archive.org/web/2023/https://www.leadresponsemanagement.org/lrm_study) found that responding within 5 minutes makes you 21 times more likely to convert a lead into a customer. After 30 minutes, that probability drops by 95 %. Most small businesses don't have 24/7 staff. But their customers expect answers outside business hours — especially if they landed on the website through an [ad campaign](https://bravos-ai.com/blog/chatbot-campanas-publicidad) that's still running at midnight. ## How an AI customer service chatbot actually works There are several ways to apply AI to customer service: automatic ticket classification, suggested replies for agents, intelligent knowledge bases. But all of those require enterprise tools like Zendesk or Intercom, multi-agent support teams, and budgets starting at hundreds of dollars per month. For a small business that wants to start today, the most direct path is an AI chatbot for customer service on your website. And we're not talking about the chatbots from five years ago — the ones that only worked if the customer clicked the right button and navigated a rigid menu of options. Modern AI chatbots use large language models and understand what the customer types in their own words, as if they were talking to a person. The customer doesn't have to adapt to the chatbot — the chatbot adapts to them. They can ask with typos, mix topics in the same message, or phrase the same question ten different ways. The chatbot understands the intent and responds with real information from your business. It doesn't make things up. In practice: you upload your content (FAQ, policies, opening hours, product catalogue), paste one line of code on your website, and the chatbot starts answering queries 24/7. No coding. You can have it running in under an hour. If a customer asks "Do you ship to Scotland?" at 2 AM, the chatbot searches your shipping information and responds with real data. If they ask something that isn't in your data, it says so instead of making up an answer. And if the query is complex, it can hand the customer off to your team or a direct contact channel. ### What it looks like in practice Picture an online furniture store. It's 11 PM, the shop is closed, and a customer lands on the site after seeing an Instagram ad: That entire conversation happened without anyone on the team being online. The chatbot filtered by size, colour and availability, answered questions about delivery and financing, and offered a direct link to the product. A [potential customer](https://bravos-ai.com/blog/chatbot-captar-clientes) who would have left the site without an answer now has the information they need to buy. ## What AI still can't do in customer service Knowing what doesn't work is just as important as knowing what does. Especially because there are plenty of companies selling ["AI agents"](https://bravos-ai.com/blog/que-es-un-agente-de-ia) that promise things the technology can't reliably deliver yet. - **Fully autonomous agents that do everything:** [Gartner predicts](https://www.gartner.com/en/newsroom/press-releases/2024-10-gartner-says-40-percent-agentic-ai-cancelled) that 40 % of agentic AI projects will be cancelled before 2027. The idea of an agent that processes returns, modifies orders and handles complaints without supervision sounds great in a demo, but in production it creates more problems than it solves. We have a [detailed comparison of AI agents vs chatbots](https://bravos-ai.com/blog/agente-ia-o-chatbot). - **Fully replacing your support team:** AI handles routine queries well. Complex complaints, frustrated customers, nuanced negotiations — those still need people. The model that works is hybrid: AI absorbs the volume, humans handle what matters. - **AI that improvises when it doesn't know:** The biggest risk of a poorly configured chatbot is that it invents answers. If a customer asks something outside your data, the chatbot should say "I don't have that information" — not fabricate a plausible-sounding response. This is called hallucination, and it's a [well-documented problem](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). ## How much does it cost to automate customer service with AI? Prices vary by tool, volume and approach. If your goal is to start automating responses on your website, an AI chatbot is the fastest and most affordable path. Here's how the most common options compare: | Tool | From (with AI) | Pricing model | AI included | | --- | --- | --- | --- | | Zendesk Suite | ~$105/agent/mo | Per agent + paid AI add-on | No (paid separately) | | Intercom | $39/mo + $0.99/resolution | Per seat + per AI resolution | No (Fin AI separate) | | Tidio | ~$68/mo | Plan + paid AI (Lyro) | No (Lyro separate, from 50 conv.) | | Bravos AI | 7-day free trial · €19 (Starter) / €49 (PRO) | Flat rate, no per-resolution fees | Yes, included in all plans | **Zendesk** charges $55/agent/month for Suite Team, plus around $50/agent/month to enable AI. It's a full helpdesk platform — powerful but built for multi-agent support teams. **Intercom** charges $0.99 per AI resolution (Fin), with a minimum of 50 resolutions per month. It can work out well or get very expensive depending on your volume. **Tidio** starts at $29/month, but AI (Lyro) costs an additional $39/month and only includes 50 conversations — extra ones cost more. For a deeper dive into chatbot pricing and billing models, we have a [detailed pricing comparison](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa). ## How to get started with AI customer service (step by step) If you've never used AI in your business, the worst decision is trying to automate everything at once. Start with the simplest thing that has the most immediate return: an AI chatbot for customer service on your website that answers the questions your team is already answering manually. Here's a real example. A dental practice receives around 40 calls a day. More than half are people asking about opening hours, cleaning prices, which insurance plans are accepted, or how to get there. The receptionist spends half the day answering the same questions while the waiting room fills up. With a chatbot trained on that information, those queries resolve themselves. The receptionist focuses on the patients in front of them. We cover this exact scenario in our [dental clinic chatbot](https://bravos-ai.com/blog/chatbot-clinica-dental) guide. And the same goes for other sectors with lots of repetitive inquiries, like a [law firm chatbot](https://bravos-ai.com/blog/chatbot-para-abogados). - **Identify your 10 most common questions.** Review them: how many have a fixed answer? Those are the ones you automate first. In the dental practice example: hours, prices, accepted insurance, address, parking, what to do in an emergency. - **Choose a tool that doesn't require coding.** If you have a technical team, you might consider Zendesk or Intercom. If not, look for platforms where you can upload your content and have the chatbot running in an afternoon. - **Train the chatbot with your real data.** Not generic responses. Upload your FAQ, your product catalogue, your policies. If you run an [online store](https://bravos-ai.com/blog/chatbot-tienda-online), look for a chatbot that can [filter your catalogue by real attributes](https://bravos-ai.com/blog/chatbot-catalogo-productos) (price, size, colour) rather than just text search. The more specific the content, the better the answers. - **Install and test.** Most chatbots integrate by pasting a single line of code on your site. If you use [WordPress](https://bravos-ai.com/blog/chatbot-wordpress), [Shopify](https://bravos-ai.com/blog/chatbot-shopify), [PrestaShop](https://bravos-ai.com/blog/chatbot-prestashop) or [Wix](https://bravos-ai.com/blog/chatbot-wix), there are platform-specific guides. - **Monitor and improve.** Review conversations, spot questions the chatbot can't answer, and add that content. Improvement is continuous, not instant. And if you want the AI itself to do that review, you can [connect your chatbot to Claude or ChatGPT](https://bravos-ai.com/blog/que-es-un-mcp) with an MCP that goes through the conversations and tells you what's failing. **If your business serves international customers**, prioritise a tool with [multilingual support](https://bravos-ai.com/blog/chatbot-multiidioma) . A chatbot that detects the visitor's language and responds automatically can save you from translating content by hand. **If you operate in Europe — EU AI Act:** In August 2026, the transparency obligations of the [EU AI Act](https://artificialintelligenceact.eu/) come into force. If you use a chatbot, you must inform customers they're talking to a machine. Fines for non-compliance can reach €15 million. Compliance is straightforward: a visible notice and no unnecessary data collection. More detail in our [article on chatbot legal liability](https://bravos-ai.com/blog/chatbot-responsabilidad-legal). ## What you can do with Bravos AI Bravos AI is the platform you're reading this on. We built it so you can create an AI chatbot for customer service trained on your own data — without coding and without depending on anyone. - Upload your content (text, PDFs, URLs, Excel files) and the chatbot learns from it. - If you have an online store on [Shopify](https://bravos-ai.com/blog/chatbot-shopify) or [PrestaShop](https://bravos-ai.com/blog/chatbot-prestashop) , it connects to your catalogue and filters products by price, size, colour — any attribute. - Acts as a [multilingual AI chatbot](https://bravos-ai.com/blog/chatbot-multiidioma) and responds in 13 languages automatically, with no manual translation. - Captures [leads](https://bravos-ai.com/blog/chatbot-captar-clientes) naturally within the conversation. - And if you use Claude or ChatGPT, [our MCP connector](https://bravos-ai.com/conector-ia-mcp) puts our team's judgment inside the assistant: it goes through your conversations and tells you which questions are going unanswered, without you opening the panel. - Plans from €19/month with unlimited messages, AI included — no per-resolution charges, no artificial limits. 7-day free trial of the full PRO plan. The furniture store conversation above isn't fictional — it's the kind of interaction our chatbots have with real customers every day. ## Frequently asked questions ### What is an AI customer service chatbot? It's a bot powered by a large language model that answers customer questions in natural language, drawing on your own business data (FAQ, policies, catalogue, opening hours) instead of generic internet knowledge. Unlike the old menu-based bots, it understands what the customer means even with typos or mixed questions, and it says "I don't have that" instead of inventing an answer. It lives on your website or WhatsApp and answers 24/7. ### How much does an AI customer service chatbot cost? It depends on the tool and approach. An AI chatbot for your website can start from €0 (free plans) up to over $100/agent/month on platforms like Zendesk. For most small businesses, tools between €19 and €50/month cover the basics — watch for per-resolution fees that scale with your volume. ### Can AI completely replace my support team? No. AI handles routine queries well — opening hours, pricing, availability, order status. But complex complaints, negotiations and customers who need real empathy still need people. The most effective model is hybrid: AI resolves the volume, humans focus on what matters. ### Do I need technical skills to set up an AI chatbot? Not necessarily. There are platforms that let you create a chatbot without coding: you upload your content (FAQ, catalogue, policies) and paste one line of code on your website. If you use WordPress, Shopify or PrestaShop, integration usually takes less than 5 minutes. ### What happens if the chatbot can't answer a question? A chatbot [with a well-written system prompt](https://bravos-ai.com/blog/system-prompt-chatbot-empresa) should recognise its limits. If a question isn't covered in its data, it should say so and offer alternatives: hand off to a human agent, provide a contact email, or suggest a related search. A chatbot that makes up answers is worse than no chatbot at all. ### How long does it take to set up? With a no-code platform, you can have a working AI chatbot in under an hour. Upload your content, customise the look and feel, paste the embed code on your site, and it's live. The ongoing work is reviewing conversations and adding content the chatbot doesn't cover yet — but that's incremental, not a big upfront effort. ### Ready to automate your customer service? Bravos AI lets you create a chatbot trained on your data in under an hour. No coding, from €19/month with unlimited messages. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # What Is Context Engineering? The Skill That Replaced Prompt Engineering **Category:** Analysis | **Read time:** 14 min | **Date:** Aug 2, 2026 | **URL:** https://bravos-ai.com/en/blog/context-engineering For two years, «prompt engineering» was the skill everyone rushed to learn: the art of wording the perfect instruction for ChatGPT. In 2026 the phrase you hear instead is **context engineering** — and Gartner has gone as far as saying «context engineering is in, prompt engineering is out». If you keep seeing the term and want to know what it actually means, without the hype, this article is for you. We build AI chatbots for a living, so we care about this for a very practical reason: it's the difference between an AI that makes things up and one that answers with your real information. We'll explain what context engineering is, where the term came from, how it differs from prompt engineering (and whether prompt engineering is really «dead»), the pieces that make up «context», and what all of it means for a business chatbot. With the real sources, not the viral threads. Contents - [What context engineering is](#what-is) - [Where the term came from](#origin) - [Context engineering vs prompt engineering](#vs-prompt) - [The pieces of «context»](#pieces) - [Why context engineering stops AI from hallucinating](#hallucinations) - [What it means for a business chatbot](#business-chatbot) - [How Bravos does the context engineering for you](#bravos) - [In short](#summary) - [Frequently asked questions](#faq) - [Sources](#sources) ## What context engineering is **Context engineering is the practice of assembling the right information around an AI model so it answers correctly.** Not the wording of a single question, but everything the model can «see» when it responds: your documents, your data, the instructions, previous messages, the tools it can use. Anthropic, which published the reference guide on the topic, defines it as *«the set of strategies for curating and maintaining the optimal set of tokens (information) during LLM inference»*. The key idea is that the language model itself — GPT, Claude, Gemini — is the same one everyone has access to. What makes it answer *your* customer's question correctly isn't a smarter model; it's the context you place in front of it. Give it the right paragraph from your returns policy and it answers precisely. Give it nothing and it guesses. Context engineering is the discipline of getting that right, on purpose and at scale. A quick word on jargon: an **LLM** (large language model) is the AI behind tools like ChatGPT or Claude. Its **context window** is the working memory it reads before answering — the space where all that information (your data, the instructions, the conversation) has to fit. Context engineering is, quite literally, deciding what goes into that space. ## Where the term came from The phrase went mainstream in mid-2025. **Tobi Lütke**, the CEO of Shopify, popularized it in a June 2025 post, describing context engineering as the art of providing all the context needed for a task to be plausibly solvable by the model. A week later, **Andrej Karpathy** (a founding member of OpenAI) endorsed it publicly, and his framing is the one most people quote: +1 for «context engineering» over «prompt engineering». People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window with just the right information for the next step. By September 2025, **Anthropic** had formalized it in an engineering guide, and the analysts followed: **Gartner** declared that «context engineering is in, and prompt engineering is out», calling it a breakout AI capability and predicting it will show up in the majority of AI tools within a few years. In other words: this isn't a random buzzword a marketing team invented. It came from the people actually building production AI, and the industry agreed it named something real. ## Context engineering vs prompt engineering Here's the honest version, because the viral headlines oversell it. **Prompt engineering** is about *how you word the instruction* — phrasing the question or the system prompt well. **Context engineering** is about *what information the model has in front of it* when it answers — the instruction being just one piece of it. Prompt engineering How you *phrase* the instruction. Scope: one prompt, one turn. «Answer as a friendly assistant, in Spanish, in under 3 sentences.» Context engineering What *information* the model can see. Scope: instructions + your data + tools + memory, across the whole conversation. «Here is the relevant section of the returns policy, the customer's order, and the rule about refunds — now answer.» **Is prompt engineering dead, then?** No — and anyone telling you it is flat-out is selling something. A well-written system prompt is still essential; it's just no longer the *whole* job. Prompt engineering is now **one layer inside context engineering** — the instructions layer. What changed is the realization that a perfectly worded prompt with the wrong information in front of it still gives a wrong answer. The wording matters; the information matters more. ## The pieces of «context» When Anthropic breaks «context» down, it's not one thing — it's a handful of ingredients that all compete for the same limited space. These are the pieces: - **System instructions** — who the assistant is, what it does and the rules it follows (this is where prompt engineering lives). We cover this one in depth in [the system prompt guide](https://bravos-ai.com/blog/system-prompt-chatbot-empresa). - **Retrieved data** — the specific pieces of *your* information pulled in for this question: a paragraph from a document, a row from a catalog. This is what [retrieval (RAG)](https://bravos-ai.com/blog/rag-sin-vectores) does — with newer variants like [PixelRAG](https://bravos-ai.com/blog/pixelrag-chatbot-empresarial) changing how it works. - **Tools** — the actions the model can take (look up an order, check stock), each one described so it knows when to use it. This is the piece that turns a chatbot into an [AI agent](https://bravos-ai.com/blog/que-es-un-agente-de-ia). - **Examples** — a few canonical samples of the behavior you want, worth more than a long list of rules. - **Message history** — the earlier turns of the conversation, kept (or trimmed) so the model remembers what was already said. - **Memory** — notes persisted outside the conversation, so the assistant doesn't start from zero every time. The craft is that the context window is finite. You can't dump everything in — too much context is as harmful as too little (a problem researchers call *context rot*). Anthropic sums up the goal as finding *«the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome»*. Not the most information; the *right* information. ## Why context engineering stops AI from hallucinating Here's the part that matters for anyone running an AI on real data. A model «hallucinates» — makes up a confident, wrong answer — mostly when it doesn't have the real answer in front of it and fills the gap with a plausible guess. Give it the actual returns policy in its context and it doesn't need to invent one. That's why context engineering and reliability are the same conversation. Grounding the model in retrieved, real information is the single most effective way to cut hallucinations. We wrote a whole piece on the failure modes in [why your chatbot makes things up](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo) — almost every fix in it is, underneath, a context-engineering fix: the bot answered badly because the right information wasn't in its context when it needed it. ## What it means for a business chatbot Put all of this together and you get a useful reframing: **a business chatbot is a context-engineering problem, not a model problem.** Every serious chatbot runs on the same handful of frontier models. The one that answers your customers well isn't the one with a secret smarter AI — it's the one that puts the right slice of *your* business in front of the model at the right moment. And that's exactly where most chatbots quietly fail. They dump a PDF into the model and hope; they retrieve text by rough similarity and miss the exact price or size; they forget the earlier messages. Those are all context-engineering mistakes. The difference between a chatbot that frustrates and one that helps is almost never the model — it's whether the context around it was engineered well. ## How Bravos does the context engineering for you The good news for a business owner is that you don't have to become a context engineer. That's the platform's job. With Bravos AI you provide the knowledge — your documents, your website, your catalog — and the platform assembles the context for every question so the bot answers grounded in your real data, not guesses. In practice that means three things you get without configuring anything: - **The right information retrieved per question**, so the model sees the relevant part of your content instead of everything or nothing. - **Exact answers on structured data**: for catalogs and anything with prices, sizes, stock or references, it filters by the real value instead of «something similar». We explain why that matters in [chatbot for product catalogs](https://bravos-ai.com/blog/chatbot-catalogo-productos). - **Your instructions and the conversation kept in context**, so the tone and the rules you set are respected turn after turn. You don't write prompts or wire pipelines. You upload your data; the context engineering happens under the hood. That's the whole point of a good platform: it turns «the skill everyone's talking about» into something you get by default. And if you already work with Claude or ChatGPT, you can inspect and adjust that context yourself in plain language — the bot's knowledge and its instructions — through the [Bravos MCP connector](https://bravos-ai.com/conector-ia-mcp) (new to the term? here's [what an MCP is](https://bravos-ai.com/blog/que-es-un-mcp)). ## In short Context engineering is the practice of assembling the right information — instructions, your data, tools, memory — around an AI model so it answers correctly instead of guessing. It replaced «prompt engineering» as the headline skill not because wording stopped mattering, but because the industry realized the information around the model matters more than the phrasing of a single prompt. For a business, the takeaway is simple: a chatbot is only as good as the context it's given. Get that right and it answers with your truth; get it wrong and it invents. And the practical way to get it right without becoming an engineer is to use a platform that does the context engineering for you. ## Frequently asked questions ### What is context engineering in simple terms? It's the practice of putting the right information in front of an AI model — your documents, your data, the instructions, the conversation so far — so it answers correctly instead of guessing. The model is the same for everyone; the context you give it is what makes the answer right. ### What is the difference between context engineering and prompt engineering? Prompt engineering is about how you word the instruction. Context engineering is about what information the model can see when it answers — instructions, retrieved data, tools, message history and memory — with the wording being just one piece of it. Context engineering is the bigger discipline; prompt engineering is one layer inside it. ### Who coined the term «context engineering»? It was popularized by Tobi Lütke, CEO of Shopify, in mid-2025, and amplified by Andrej Karpathy (a founding member of OpenAI). Anthropic formalized it in an engineering guide in September 2025, and Gartner then declared «context engineering is in, prompt engineering is out». ### Is prompt engineering dead? No. A well-written system prompt is still essential. What changed is that it's now one layer inside a bigger job: prompt engineering is the instructions, context engineering is the whole set of information around the model. A perfect prompt with the wrong information in front of it still gives a wrong answer. ### What are the components of context engineering? Per Anthropic: system instructions, retrieved external data, tools, examples, message history and memory. The craft is fitting the *right* pieces into a finite context window — the smallest set of high-signal information that produces the desired answer, not the most information. ### Why does context engineering reduce hallucinations? A model makes things up mostly when it doesn't have the real answer in its context and fills the gap with a plausible guess. Putting the actual information — the real policy, the real price — in front of it removes the reason to invent. Grounding the model in retrieved, real data is the most effective way to cut hallucinations. ### A chatbot that answers with your data, not guesses Upload your data and Bravos AI assembles the right context for every question — exact, grounded answers on web and WhatsApp. Free 7-day PRO trial. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) ## Sources - **Anthropic — Effective context engineering for AI agents:** [anthropic.com ](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) — the reference technical guide (Sept 2025), source of the definition and the components of context. - **Andrej Karpathy on context engineering:** [x.com/karpathy ](https://x.com/karpathy/status/1937902205765607626) — the post that popularized the term (June 2025). - **Gartner — Context engineering vs prompt engineering:** [gartner.com ](https://www.gartner.com/en/articles/context-engineering) — the analyst view («context engineering is in, prompt engineering is out»). --- # What Is an MCP and How to Manage Your Chatbot from Claude or ChatGPT **Category:** Practical guide | **Read time:** 15 min | **Date:** Jul 29, 2026 | **URL:** https://bravos-ai.com/en/blog/what-is-mcp «MCP» is one of those acronyms that in 2026 shows up everywhere —in headlines, in threads, in the Claude and ChatGPT docs— and that almost nobody bothers to explain from scratch. If you landed here looking for what an MCP is and what it's for, without anything taken for granted, this article is for you. We'll explain it without the jargon: what an MCP is, what an «MCP server» is, and why it suddenly matters to so many people. We build AI chatbots and just launched our own, so at the end we land the idea on a very concrete example: what changes when a chatbot can be analyzed and fine-tuned from the same Claude or ChatGPT you already use. But first, the important part: what is an MCP. Contents - [What an MCP is, explained without the jargon](#what-is) - [What an MCP server is and how it works](#mcp-server) - [What it's actually for (with examples)](#what-for) - [What it changes for a business with a chatbot](#what-changes) - [The Bravos MCP: what you can do from the chat](#mcp-bravos) - [How to set up your chatbot and connect it, step by step](#how-to-set-up) - [What sets it apart](#what-sets-apart) - [What it does NOT do (honesty first)](#what-it-doesnt) - [Why we built it](#why) - [In short](#summary) - [Frequently asked questions](#faq) - [Sources](#sources) ## What an MCP is, explained without the jargon **MCP stands for *Model Context Protocol*.** It's an open standard that Anthropic —the company behind Claude— published in late 2024, and that ChatGPT and other tools now use too. It sounds very technical, but the idea is simple: it's a **common language so an AI like Claude or ChatGPT can connect to an outside program and use it**. The comparison that captures it best is USB. Before USB existed, every device came with its own plug and its own cable, and connecting two things was a lottery. USB gave everything the same connector. An MCP is that, but for AI: one «socket» so any AI assistant can talk to any outside service —your email, your calendar, your store, or your chatbot— without building a custom cable for every combination. There are already quite a few. There's an MCP for Gmail, which lets the AI read and draft your emails; one for Google Drive, which lets it search inside your documents; and there are ones for your calendar or your online store. Each one opens a specific service to the AI, without anyone having to program a custom connection for each case. Before the MCP, if you wanted Claude to read your data from a specific tool, someone had to program a specific connection for that pair: Claude with that tool, ChatGPT with that tool, one by one. With the MCP, the service «introduces itself» once in this common language, and from then on any AI that speaks MCP understands it. That's why one connector works the same in Claude and in ChatGPT. Throughout the article you'll see «AI», «assistant» or «model» to mean Claude, ChatGPT, or any other assistant that speaks MCP —we name those two because they're the best known, but the MCP isn't exclusive to any of them—. And «chatbot» or «bot» for the assistant you put on your website or your WhatsApp to serve your customers. They're two different things, and this is precisely about connecting the first to the second. ## What an MCP server is and how it works An MCP always has two sides. On one side is the **client**: the AI that wants to do something, for example Claude or ChatGPT. On the other is the **MCP server**: the program that opens the door to a specific service and offers the AI a list of actions it can request. When you read «Gmail MCP server», for example, that's exactly it: the door through which Claude or ChatGPT enter your email, with the permission you grant. That server doesn't give the AI free access to everything. It offers a closed list of **tools**, which are specific, named actions: in an email MCP they'd be «find this customer's emails» or «draft a reply»; in a store's, «check this product's stock» or «list today's orders». The AI can only ask for things on that list; nothing more. It's the difference between handing someone the keys to the whole house or handing them a remote with specific buttons. The flow, in three steps, is this: - **You connect once.** In Claude or ChatGPT you add the connector for whichever service you want and sign in, just like when an app asks you to sign in «with your Google account». You authorize what it can see and do, and that's it. - **You ask for something in plain language.** «Find this customer's unanswered emails and prep a draft for me.» - **The AI uses the tools for you.** Under the hood, it calls the actions it needs on the server, gathers the data and does what you asked. You don't see the mechanics; you see the result. That authorization uses a standard called **OAuth**, the same one behind «sign in with Google» or «sign in with Apple». You don't share your password with the AI: you grant a specific permission that you can revoke whenever you want from your panel. No code, no keys to copy and paste. ## What it's actually for (with examples) The theory lands better with something concrete. With a chatbot connected via MCP, instead of opening the panel, clicking through menus and drawing your own conclusions, you talk to Claude or ChatGPT like you'd talk to a colleague who has all your data in front of them: - «Go through last week's conversations and tell me where people get stuck the most.» - «My bot answers oddly when people ask about delivery times. Look at what info it has loaded on that and tell me what's missing.» - «Rewrite the bot's instructions so it's more to the point and doesn't ramble.» - «Which leads came in yesterday and what did they want?» - «Make the tone warmer and swap the welcome message for one that mentions this month's offer.» And the AI doesn't just tell you: if you ask, it does it. It changes the text, rewrites the instructions, updates the tone. It stops being a panel you operate and becomes an assistant that works with you on your own chatbot. ## What it changes for a business with a chatbot The moment most people abandon a chatbot isn't when it fails spectacularly. It's when it answers so-so, the business owner senses something's off but doesn't know what, has no time to dig through conversation after conversation, and slowly lets it die. The problem is almost never that the tool is bad: it's that **tuning a chatbot means looking at data and knowing what to change**, and that takes effort. That's where the MCP changes things. It turns that «something's off» into a diagnosis with the fix applied, done by the AI itself, in five minutes and in plain language: which questions go unanswered, what information is missing, which instruction is confusing the bot, and —if you give the go-ahead— the fix made on the spot. You don't need to be technical. You need to be able to describe your problem, which is something you already know how to do. If you're into the «why» of a bot answering badly, we cover it separately in [why your chatbot makes things up](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). The MCP is, in large part, that article turned into something the AI does for you. ## The Bravos MCP: what you can do from the chat [Our connector](https://bravos-ai.com/conector-ia-mcp) opens your Bravos account to Claude, ChatGPT and any other assistant that speaks MCP, with a list of actions built for three things: **understanding** how your bot is doing, **improving** what it answers and **keeping it up to date**. Specifically, from the chat you can: **Analyze conversations and leads.** Read and search the history, view stats, spot where your customers get stuck and which contacts the bot captured. **Audit the knowledge base.** Review everything the bot has loaded (texts, websites, documents, catalogs), and spot what's missing, what's redundant and what's out of date. **Create, edit and delete content.** Text entries and web pages in the knowledge base: the AI adds, corrects or removes them. Since all of that is recoverable, it can undo what it did with no trouble. **Fine-tune the configuration.** Rewrite the bot's instructions, adjust the tone, change the widget messages, the languages, the appearance or the domains where it runs. **Keep integrations current.** Check the status of your connected store (WooCommerce, Shopify, PrestaShop…) and trigger a catalog sync. Every action that changes something returns the before and after right in the conversation, so you always see what was touched and can revert it. And since everything goes through your plan, the AI never proposes something your plan doesn't include. The most useful detail for the result: when you ask the AI to write your bot's instructions, it does so knowing how Bravos works under the hood —which part of the behavior already comes built in and doesn't need repeating—. On why good instructions change an answer so much, we wrote [the system prompt guide](https://bravos-ai.com/blog/system-prompt-chatbot-empresa). ## How to set up your chatbot and connect it, step by step Here it's worth being precise, because there's a nuance. The AI does *almost* all the work, but you create the bot in the panel. That's on purpose: creating a bot touches your plan and your billing, and those two things we want you to decide, seeing them, not have an assistant fire off. The rest the AI can do. The full path is like this: - **You create the bot in the panel.** It takes thirty seconds: give it a name and you're done. - **You connect Bravos to your assistant.** Whether it's Claude, ChatGPT or another that speaks MCP: in the panel, under **Integrations → MCP**, you have the connector's address and the steps. You paste it into your assistant, sign in with your account and authorize. Once. - **The AI gets it up and running.** You tell it your website and it loads the content, writes the instructions, adjusts the tone and leaves the bot ready. You review and approve as you go. - **You test and install it.** You check how it answers and copy the code to put it on your website, or connect it to your WhatsApp. Put another way: you make the decision (create the bot) and describe what you want; the tedious part —loading information, writing, tuning— the AI does with you alongside. ## What sets it apart To be honest: we're not the only ones offering to connect a chatbot to Claude or ChatGPT, and we're not going to sell it as if we were. It's a new wave and there'll be more of them. What matters isn't having the connector, it's **what it connects to and at what price**. And that's where our package stands out: - **Real structured search.** Bravos doesn't only search by similarity: it understands catalogs and column-based data and filters by exact values (price, size, reference, availability). It's the difference between a bot that «sort of» finds things and one that gives you the right value. We cover it in [how a chatbot finds information](https://bravos-ai.com/blog/rag-sin-vectores) and in [a chatbot for product catalogs](https://bravos-ai.com/blog/chatbot-catalogo-productos). - **Price.** All of this capability isn't locked in a high-end, enterprise-named plan; it comes with plans built for regular businesses. You can see [how much a business chatbot costs](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa). - **Trilingual from birth.** The bot, the panel and support all work natively in English, Spanish and Portuguese —it's not a single-language tool translated on top—, and the bot itself serves your customers in many more. - **WhatsApp included.** The same bot you manage from Claude serves on your website and on your WhatsApp. - **No extra cost for the MCP.** The connector comes with your plan from Starter, and it's available during the 7-day PRO trial too — exactly when you want to check whether this is for you. A detail that tends to surprise people: the analysis the AI does is paid by your own subscription —Claude's, ChatGPT's, or whichever assistant you use—, not by us. For you, that means we don't charge you for any «optimization assistant»: you use the AI you already pay for, applied to your chatbot. ## What it does NOT do (honesty first) So you decide with the full picture, here's what the connector **doesn't** do, on purpose: **It doesn't create the bot from the chat.** You take that first step in the panel, because it touches your plan and your billing. **It doesn't upload or delete files.** Texts and URLs it does manage, because they're recoverable; files it doesn't. The AI reads what's inside your documents (PDF, images, audio) and recommends what to keep or remove, but deleting a file is something you do in the panel. It's a no-going-back action and we'd rather you be the one to run it. **It doesn't touch your credentials.** It can trigger a sync of your store, but it never sees or connects the keys to your WooCommerce or your Shopify. A practical requirement: to use a connector like this you need an AI assistant that supports MCP connectors, usually on its paid plan (Claude, ChatGPT and others already offer them). If you don't use one, the chatbot works just as well from the usual panel; the MCP is an extra layer for those who already work with one of these AIs. ## Why we built it For a long time, when a customer didn't know why their bot answered so-so, the help was us: we'd look at their conversations, tell them what to load, rewrite their instructions. We did it well, but it depended on us: on our time and on us being available. The customer with a question on a Sunday night, or the one who doesn't dare write in, was left waiting. The MCP was what we needed without knowing we needed it. It flips that help around: it no longer has to come from us. We built it so that **any customer gets support just as professional as what we give by hand —to be exact, the same— but 24 hours a day and without waiting for anyone**, because it's their own AI that delivers it, with all their data in front of it. That, for us, was the missing step. ## In short An MCP is the standard that lets an AI like Claude or ChatGPT use outside services —your email, your store, your chatbot— without a custom integration for each one. It's the piece that gets the AI out of its bubble and lets it work with your stuff. Applied to a chatbot, it turns «I don't know why it answers badly» into a diagnosis with the fix made, from the same assistant you already use. And if you don't work with one yet, you lose nothing: all of this is still in the usual panel. The MCP just puts it one message away. ### Your chatbot, ready from your own AI Create your bot in the panel and let Claude, ChatGPT or whichever assistant you use get it into shape: it loads the content, writes the instructions and fine-tunes it, in English, Spanish and Portuguese. The MCP connector comes with your plan from the first tier and during the 7-day PRO trial too: we warn you before charging, and if you cancel before day 7 you pay nothing. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) ## Frequently asked questions ### What is an MCP in simple terms? It's an open standard that lets an AI assistant connect to outside services and tools and use them. It works like a USB for AI: one socket for everything, instead of building a custom cable for each connection. ### What does MCP stand for? It stands for *Model Context Protocol*. Anthropic created it in late 2024, and today it's an open standard adopted by much of the industry. ### What is an MCP server? It's the program that opens the door to a specific service and offers a closed list of actions the AI can request. A Gmail MCP server, for example, lets it read or draft your emails; a store's lets it check your orders. The AI can only ask for what's on that list. ### What is an MCP for? To get the AI out of its bubble. Without an MCP, it only knows what it learned in training; with one, it can look at your data and act in your tools in real time: your email, your store, your chatbot. And without building a different integration for each. ### Is an MCP the same as an API? No. An API is the technical way two programs talk to each other, and it's been around for decades. The MCP is a layer on top, designed so the AI discovers and uses those actions on its own, in natural language. In fact, many MCP servers are just a regular API wrapped so a model can understand it. ### Does MCP work with ChatGPT and Claude? Yes, and with more. Being an open standard, it doesn't depend on a single company, so one server works for any assistant that supports it. Claude and ChatGPT are the best known, but not the only ones. ### Do I need to know how to code to use an MCP? To use one that already exists, no: you connect with a sign-in, no code and no keys to copy, and you talk to the AI in plain language. Coding is only needed to build a new MCP server, not to use one. ## Sources - **Anthropic (original announcement):** [anthropic.com ](https://www.anthropic.com/news/model-context-protocol) — introducing the Model Context Protocol, November 2024. - **Official MCP documentation:** [modelcontextprotocol.io ](https://modelcontextprotocol.io/) — the standard explained by its authors. - **Connectors in Claude:** [support.claude.com ](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp) — subscription requirements for using MCP connectors. The Claude and ChatGPT plan conditions for using connectors are set by Anthropic and OpenAI and may change; check the current one before subscribing. --- # Vectorless RAG: Do You Still Need a Vector Database in 2026? **Category:** Analysis | **Read time:** 17 min | **Date:** 23 Jul 2026 | **URL:** https://bravos-ai.com/en/blog/vectorless-rag For two years, the answer to «how do I make a chatbot answer from my own data?» was always the same: chop the data into chunks, turn each chunk into numbers, store the numbers in a vector database. Then, in mid-2026, a wave of posts declared that step dead. «Vectorless RAG.» «You don't need a vector database anymore.» «RAG is dead.» We build AI chatbots for a living, so we read the actual papers, ran the numbers, and looked past the headlines —the same way we did with [PixelRAG](https://bravos-ai.com/blog/pixelrag-chatbot-empresarial). This is the honest version for someone deciding what to build: what vectorless RAG actually is, how a chatbot finds information in the first place, what a vector database really does, and the practical question underneath all the noise — do *you* need one? With the real figures, not the viral ones. On this page - [What is vectorless RAG?](#what-is) - [How an AI chatbot actually finds information](#how-chatbot-finds) - [What a vector database is (and what embeddings are)](#vector-database) - [Semantic search vs keyword search](#semantic-vs-keyword) - [Is RAG dead? What the debate actually says](#is-rag-dead) - [How vectorless RAG works under the hood](#how-vectorless) - [Do you need a vector database? A decision guide](#do-i-need) - [Vector RAG vs vectorless RAG: an honest table](#comparison) - [What this means for your business chatbot](#business-chatbot) - [The verdict, in one line](#verdict) - [Frequently asked questions](#faq) - [Sources](#sources) ## What is vectorless RAG? **Vectorless RAG is retrieval without a vector database.** Instead of converting your documents into numerical vectors and searching by mathematical similarity, the system uses the document's own structure — or plain keyword search — and lets a language model *reason* about which part is actually relevant. RAG stands for *Retrieval-Augmented Generation*: the technique most AI chatbots use to answer from your data instead of making things up. «Vectorless» changes the retrieval half of that. The poster child is **PageIndex**, from VectifyAI. It builds a hierarchical «table of contents» tree from a long document, then, when a question comes in, an LLM walks that tree the way a person flips through a report — «this is a financial filing, the answer is probably in the liquidity section, page 40» — and pulls the right section. No chunking, no embeddings, no vector store. Their tagline captures the whole pitch: *similarity is not relevance*. There's a second flavour that's even simpler: give an AI agent a plain **keyword search** tool (the kind of full-text search a database has had for decades) and let it search, read, refine, and search again in a loop. That's how coding agents like Claude Code and Cursor navigate a codebase — they grep, they don't embed. A recent paper by researchers at Amazon put a number on it (more on that below). «Vectorless» does not mean «no AI». There's still a language model doing the heavy lifting — arguably more of it. What disappears is the *vector database* in the middle: the embeddings index that most RAG tutorials treat as mandatory. ## How an AI chatbot actually finds information (RAG in 60 seconds) To judge whether removing the vector database is a good idea, you need to know what it was doing there. Here is the standard RAG pipeline that powers the vast majority of AI chatbots today, ours included: - **Chunk.** Your documents — a PDF, a web page, a product catalog — get split into small passages of text. - **Embed.** Each passage is fed to an embedding model that turns it into a list of numbers (a *vector*) capturing its meaning. Two passages about the same idea end up with similar numbers, even if they use different words. - **Store.** Those vectors go into a *vector database* built to find nearest neighbours fast. - **Retrieve.** When a customer asks something, the question is embedded too, and the database returns the passages whose vectors sit closest to it. - **Generate.** Those passages get handed to the language model, which writes the answer grounded in them. The vector database lives in step 3 and 4. It exists to answer one question quickly: *which of my thousands of passages are most similar in meaning to this query?* Vectorless approaches argue that «most similar in meaning» is often the wrong question — and that a language model, given the document's structure or a search tool, can find the genuinely relevant part more reliably. ## What a vector database is (and what embeddings are) An **embedding** is a way of turning text into a point in space. The embedding model reads «waterproof jacket» and outputs a long list of numbers — say 1,536 of them — that pin that phrase to a specific coordinate. «Rain coat» lands nearby because it means something similar; «diesel engine» lands far away. The whole trick is that *closeness in this space ≈ closeness in meaning*. A **vector database** (Pinecone, Weaviate, Qdrant, pgvector on top of PostgreSQL, and others) is storage optimised to hold millions of those points and, given a new one, return the nearest neighbours in milliseconds. That's it. It's a very good tool for one specific job: fuzzy, meaning-based matching over a big pile of text where you don't know the exact words the user will type. It's also where a lot of the cost, complexity and failure modes of RAG live: chunk boundaries that cut a table in half, an index to keep in sync, and the core assumption that similarity equals relevance. That's where the vectorless argument starts. In a lot of real cases, going for the closest match doesn't get you the right answer, so building and maintaining all that infrastructure isn't worth it. ## Semantic search vs keyword search This is the fault line under the whole debate, so it's worth a plain example. Same knowledge base, same question, two ways of finding the answer: Neither is universally better. Semantic search shines when the user's words won't match your document's words. Keyword and structured search shine when the answer hinges on an exact value — a price, a size, a reference, a date. The mistake the last two years made was treating vectors as the default for everything, including questions where a plain filter would have been faster, cheaper and more accurate. ## Is RAG dead? What the debate actually says Short answer: no. What's being questioned isn't «grounding a model in your data» — that's more relevant than ever. What's being questioned is one specific implementation: *chunk everything, embed it, and search by vector similarity*. Three threads fed the «RAG is dead» wave, and each says something narrower than the headline: - **Bigger context windows.** Models now take huge inputs, so for a *small* knowledge base you can sometimes just paste everything in and skip retrieval entirely. This breaks down at scale and gets expensive — and it collides with the next point. - **Context rot.** Chroma's research showed model accuracy *degrades* as the input grows, well before the window is full. So stuffing everything in isn't free — the model gets worse the more you add. If anything, context rot is an argument for *better* retrieval, not for dropping it. - **Agentic search.** Give a capable model a search tool and let it look things up in a loop, and it often beats a one-shot vector lookup. This is the real substance behind vectorless RAG. So the honest framing is not «RAG is dead» but «the reflexive vector-database-for-everything era is ending». Retrieval is alive; the monoculture around one method is what's fading. Retrieving the right information for each question is one of the pieces of [context engineering](https://bravos-ai.com/blog/context-engineering), the broader discipline of assembling the context a model needs to answer well. ## How vectorless RAG works under the hood Take PageIndex as the concrete example. Instead of chunking and embedding, it does this: - **Build a tree.** It generates a hierarchical, table-of-contents-style index of the document — sections, subsections, page anchors — that mirrors how the document is actually organised. - **Reason over the tree.** When a question arrives, an LLM does tree search: it looks at the structure and decides which branch is likely to hold the answer, navigating like a human skimming a report, rather than matching vectors. - **Retrieve the relevant node.** It pulls the section it reasoned its way to, and the model answers from that. The keyword-agent flavour skips even the tree: the agent issues full-text searches, reads results, and refines its query until it has what it needs — the same loop a person uses with a search box. On the numbers: VectifyAI reports **98.7% accuracy on FinanceBench** for PageIndex, «vastly outperforming vector RAG». Treat that as a *vendor's own benchmark*, not an independent result — it's their tool on one financial-document test they selected. The separately published paper *«Keyword search is all you need»* (Subramanian et al., Amazon) is more measured: agentic keyword search reaches **over 90% of the performance of traditional RAG**, while being «simple to implement» and «cost effective». Note the direction: *comparable to*, not «crushes». ## Do you need a vector database? A decision guide Forget the ideology. Run through these and count where you land: {[ 'Your knowledge base is large (thousands of pages) and users ask open-ended, meaning-based questions where their words won’t match yours.', 'You need fuzzy matching across a lot of unstructured prose — support articles, manuals, policies — with no obvious structure to navigate.', 'Low, predictable per-query cost and sub-second latency matter more than squeezing out the last few points of accuracy.', ].map((t, i) => ( ** .** ))} - **Mostly yes:** a vector database still earns its place. Semantic search over a big, messy corpus is exactly what it's good at. Keep it. - **Mixed:** you probably want *hybrid* — keyword/structured search for exact stuff, vectors for the fuzzy stuff. Most serious systems already blend both. - **Mostly no:** if your data has clear structure (well-organised documents) or your questions hinge on exact values (catalogs, records, filters), vectorless — a reasoning tree or plain structured search — may be simpler, cheaper and *more* accurate. ## Vector RAG vs vectorless RAG: an honest table | Dimension | Vector-based RAG | Vectorless RAG | | --- | --- | --- | | Retrieval by | Vector similarity (meaning). | LLM reasoning over structure, or keyword search. | | Infrastructure | Embedding model + vector DB to run and sync. | No vector DB. A tree index or a search tool. | | Latency & cost per query | Low, predictable. One embedding + one lookup. | Higher & variable. Multiple LLM reasoning steps. | | Best at | Fuzzy, meaning-based questions over big unstructured corpora. | Well-structured documents; exact-value and navigational questions. | | Weak at | Exact codes/filters; when similarity ≠ relevance. | Scale & speed; messy corpora with no structure. | | Maturity | Years in production, huge tooling ecosystem. | New (2026), fast-moving, largely vendor/self-reported benchmarks. | The row that matters most is the last one: vector RAG is a known quantity; vectorless is promising but young, and most of its winning numbers come from the people selling it. That's not a reason to ignore it — it's a reason to test it against your own data rather than a headline. ## What this means for your business chatbot Here's the part the debate usually skips: most business chatbots never needed pure vector search in the first place. A store chatbot answering «waterproof jackets under €80 in size L» is not a similarity problem — it's a *filter*. The right answer is structured search over your catalog (price, size, stock), not the nearest vector to the sentence. We wrote about exactly this in our guide on [why your chatbot can't find products](https://bravos-ai.com/blog/chatbot-catalogo-productos). At [Bravos AI](https://bravos-ai.com/) we already run a hybrid: structured, database-style search for catalogs and exact-value questions, and semantic retrieval for free-form text like FAQs, policies and service descriptions. The «vectorless» conversation didn't surprise us because the lesson underneath it — *match the retrieval method to the question, don't force vectors on everything* — is how a good system should have been built all along. If your chatbot mostly answers questions about a catalog or structured records, the vector database was never the important part. The takeaway for a business isn't «rip out your vector database». It's: know which of your questions are meaning-based (keep semantic search) and which are exact-value or structural (use filters, tree search, or keyword search). The best chatbots blend both. ## The verdict, in one line Vectorless RAG is a real, useful correction to two years of over-using vector databases — not the death of RAG. For structured documents and exact-value questions it can be simpler, cheaper and more accurate; for large, fuzzy, meaning-based corpora, a vector database still earns its keep. The right question was never «vectors or no vectors» — it's «which retrieval fits *this* question». If a vendor tells you one method wins every time, they're selling, not engineering. ## Frequently asked questions ### What is vectorless RAG in simple terms? It's retrieval-augmented generation without a vector database. Instead of turning your documents into numerical embeddings and searching by similarity, the system uses the document's structure (a reasoning tree, like PageIndex) or plain keyword search, and lets a language model reason about which part is actually relevant. The tagline is «similarity is not relevance». ### Do I still need a vector database for RAG in 2026? It depends on your questions. For large, unstructured corpora where users ask fuzzy, meaning-based questions, yes — vector search is what it's good at. For well-structured documents or questions that hinge on exact values (catalogs, records, filters), you may not, and a vectorless or hybrid approach can be simpler and cheaper. Many production systems use both. ### Is RAG dead? No. Grounding a model in your own data is more relevant than ever. What's fading is the reflex of chunking and embedding everything into a vector database by default. Retrieval is alive; the one-size-fits-all method is what's being questioned. ### What is PageIndex? PageIndex (by VectifyAI) is a vectorless, reasoning-based RAG system. It builds a hierarchical table-of-contents tree from a document and uses an LLM to navigate it, with no embeddings, chunking or vector database. Its authors report 98.7% accuracy on the FinanceBench benchmark — a vendor-reported figure, so worth verifying against your own data. ### Vectorless RAG vs vector RAG — which is better? Neither wins outright. Vector RAG is mature, fast and strong at fuzzy semantic matching over big corpora. Vectorless RAG is newer and strong at structured documents and exact-value questions, with simpler infrastructure but higher per-query reasoning cost. The best answer for most businesses is a hybrid that routes each question to the method that fits it. ## Sources - **PageIndex (VectifyAI):** [github.com/VectifyAI/PageIndex ](https://github.com/VectifyAI/PageIndex) — vectorless, reasoning-based RAG; 98.7% on FinanceBench (vendor-reported). - **«Keyword search is all you need»:** Subramanian et al., [arXiv:2602.23368 ](https://arxiv.org/abs/2602.23368) — agentic keyword search reaches over 90% of traditional RAG performance without vector databases. - **Context Rot (Chroma):** [research.trychroma.com/context-rot ](https://research.trychroma.com/context-rot) — model performance degrades as input context grows. ### A chatbot that retrieves the right way Bravos AI blends structured search for catalogs and exact-value questions with semantic retrieval for free-form text — so your chatbot answers accurately whether the question is a filter or a paraphrase. In English, Spanish and 12+ languages, sub-2-second latency. 7-day PRO trial, cancel before day 7 and you pay nothing. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Wix Chatbot with AI: What Comes Built In and How to Add Your Own **Category:** Practical guide | **Read time:** 13 min | **Date:** Jul 20, 2026 | **URL:** https://bravos-ai.com/en/blog/wix-chatbot If you're looking for a Wix chatbot with AI, you have two paths: use the one Wix ships built in (it's called Smart Chat) or add a dedicated one from an outside provider. Neither is "the right one" for everybody — it depends on what your business needs. In this guide we cover exactly what Wix's built-in chat does and how its credit system works (with the official numbers, not hearsay), when it's enough, when it falls short, and how to install an external chatbot on your Wix site step by step. Everything verified against Wix's official documentation in July 2026 — sources at the end. ## What Wix gives you built in: Smart Chat Smart Chat is Wix's current chat solution (in 2026 it replaced Wix's first AI chat, the AI Site Chat). You install it from the Wix App Market and it has two modes you can combine: - **Manual mode:** visitors write to you and you (or someone on your team) reply by hand from the Wix Inbox. This mode is free and unlimited. - **AI mode:** an assistant answers on its own, based on your site's content and rules you write in the dashboard (your shipping policy or store locations, for example). You can tune its tone, exclude pages and block topics. On catalogs, what Wix documents is that the AI can **suggest items and link to pages on your site** based on your catalog and site structure. The widget's interface is available in 34 languages. For the full picture: on its own App Market listing, Smart Chat holds a rating of [2.9 out of 5 across 133 reviews](https://www.wix.com/app-market/wix-ai-assistant-client) (July 2026). ## How its credit system works (read this before deciding) Here's the part almost nobody explains. Smart Chat's AI mode is not unlimited: **every AI answer consumes "AI credits" from your Wix plan**. The official numbers: - **Free Wix plan:** up to 30 credits per day, capped at 120 per month. - **Paid plans:** Light, around 150 actions per cycle; Core, around 200; Business, around 300; Business Elite, around 1,000. - **When you install the app** you get a trial of 50 conversations in total. And three details of the system worth knowing: - **The pool is shared.** Those credits aren't just for the chat: Wix's other AI tools (the marketing agent, the design generator…) draw from the same balance. - **There's no guaranteed equivalence.** Wix says it literally in its documentation: "the exact number of actions you can take and how many credits each action uses is not guaranteed," because it depends on the length and complexity of each request. There's no public way to know how many conversations your plan's credits buy — we checked, and no measurement has been published anywhere, not even by users. - **When they run out, the AI chat switches off.** Straight from the documentation: once the credits are used up, Smart Chat goes into offline mode. You can keep replying by hand, but the AI doesn't come back until the next cycle (unused credits don't roll over either). Think about when your credits would run out: exactly when the most people are writing to you. A campaign that works, a press mention, a holiday peak — the moment with the most inquiries is precisely the moment a credit system runs dry. ## When Smart Chat is enough Let's be clear, because not everyone needs more: - If your site gets little traffic and a handful of inquiries a month, the included credits may be all you need, at no extra cost. - If everything the bot needs to know is already published on your site, training on site content works. - If what you mostly want is manual chat (replying yourself from the Inbox), that mode is free and does the job. ## When you need a dedicated chatbot There are four situations where a chat that feeds on your site's public content will, by design, fall short: **1. Your catalog needs real filters.** Suggesting items is one thing; applying exact filters — product type, feature and price at once — without slipping in something that doesn't qualify is another. Here's the difference with a real inquiry: Notice two things: results come with their price, availability and a link to the product — the customer can buy from right there — and when nothing matches the filters, the bot **says so** and offers the closest alternative with its real price, instead of fudging it. That exact filtering over structured data is a different technology — we explain it in [why chatbots fail with product catalogs](https://bravos-ai.com/blog/chatbot-catalogo-productos). **2. The information isn't (all) on your site.** Internal rate sheets, detailed terms, PDF documents, a spreadsheet with your inventory… A dedicated chatbot trains on the documents and data you upload, published or not. **3. You want to cover WhatsApp too.** Smart Chat lives on your website. If your customers write to you on [WhatsApp](https://bravos-ai.com/blog/chatbot-whatsapp), you need a bot that covers both channels with the same knowledge. **4. You sell in several languages.** Having the widget's texts translated is one thing; having the AI understand the visitor and answer in their language is another. If people write to you in Spanish or French and you want the bot to reply in Spanish or French without you translating anything, that's what a dedicated [multilingual chatbot](https://bravos-ai.com/blog/chatbot-multiidioma) does. | | Smart Chat (per its documentation) | Dedicated chatbot (Bravos AI) | | --- | --- | --- | | Trains on | Your site's content + hand-written rules | Your site + documents, texts and catalogs you upload | | Messages | Plan credits, shared, with no guaranteed equivalence | Unlimited, flat rate | | When the balance runs out | The AI chat switches to offline mode | Not applicable | | Catalog | Suggests items and links to pages | Exact filtering across several attributes at once | | WhatsApp | Not documented | Yes, with the same knowledge as the widget | | Languages | Translated widget interface | The AI detects the visitor's language and answers in it | ## The 4-question test (before paying for any chatbot) No matter the vendor — Smart Chat, us or anyone else — almost all of them have a free trial, and in 15 minutes you can uncover what the marketing doesn't say. Run these four tests with your real information: - **The two-filter question.** Ask for something from your catalog combining two conditions (product type + price cap, for instance). Then a combination you know doesn't exist. If it returns something that doesn't meet the conditions — or makes up a result instead of saying "there's none" — you already know how it will treat your customers. - **The question that isn't on your site.** Ask it something you know you haven't given it. The right answer is admitting it doesn't know and offering contact; the wrong one is improvising a convincing reply. A bot that makes things up is worse than no bot. - **The message in another language.** Write to it in Spanish or French. Does it answer in that language with your information, or does it get lost? - **The math with YOUR volume.** Estimate your monthly inquiries and ask for the total price at that volume: credits, conversations, resolutions or messages — every vendor measures differently and that's where the surprises hide. If they can't give you a firm number, that's an answer too. ## Prepare your Wix site so any chatbot answers better This applies whatever you choose, because every AI chat that learns from your site answers only as well as the content it finds. Four improvements you can make in an afternoon: - **Publish your frequently asked questions as text.** Shipping, returns, timelines, payment methods. If the answer exists on your site, the bot finds it; if it only exists in your head, it doesn't. - **Write policies in full, not summarized.** A bare "returns within 30 days" just triggers the next question (who pays return shipping?). Every gap in your text is an inquiry the bot won't be able to resolve. - **Complete your product or service pages.** Dimensions, materials, what's included, what isn't. Data missing from the page is a filter no bot will be able to apply. - **Get key information out of images.** Opening hours in a photographed sign or rates inside a scanned PDF are invisible to most bots. Move them into page text. If you want to go deeper on this, we have a full guide on [preparing your content so your chatbot doesn't fail](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). ## How to add an external chatbot to your Wix site, step by step Installing an outside chatbot on Wix is simple, but there's one prerequisite worth knowing before you start: **Wix only allows external code on published sites with a connected domain** — and connecting a domain is a paid-plan feature. If your site is on Wix's free plan (a wixsite.com address), you won't be able to add any external widget, ours or anyone's. With that covered, the process with Bravos AI (other providers are similar): - **Create your chatbot and train it.** Give it your site's address so it learns from your content, and upload what isn't published: documents, texts, your catalog in a spreadsheet. No coding. - **Configure it:** tone, language, what it can and can't say (the [system prompt](https://bravos-ai.com/blog/system-prompt-chatbot-empresa)), and whether you want it to capture contact details. - **Paste the code into Wix.** In your site's dashboard: Settings → Custom Code (under "Development & integrations") → Add Custom Code. Paste the widget line, under "Add Code to Pages" choose all pages, and under "Place Code in" select the end of the body. Apply, and the chat shows up on your site. The widget is the little chat window your visitor sees. It loads like any external script: it doesn't touch your Wix template or the editor. If you ever want it gone, delete the line of code and that's it. ### And your Wix store catalog — how do you connect it? If you sell with Wix Stores, this is how it works today: - **Export your catalog from Wix.** In your store's dashboard: Products tab → tick the top checkbox to select them all → More Actions → Export. Wix generates a CSV with your products and their data (up to 5,000 rows per file). - **Connect it to the chatbot.** You can upload that CSV directly, or paste it into a Google Sheets document and connect it as a synced sheet: the bot re-reads it automatically every day or every week, your choice. - **When the catalog changes, repeat the export.** It takes a couple of minutes. Let's be frank about the limits here: if your catalog changes little — stable prices, products that rotate every few weeks — it works fine and painlessly. If it changes daily, exporting by hand gets tedious. That's why a direct Wix Stores integration is on our roadmap, like the ones we already have for [Shopify](https://bravos-ai.com/blog/chatbot-shopify), [WooCommerce](https://bravos-ai.com/blog/chatbot-woocommerce) and [PrestaShop](https://bravos-ai.com/blog/chatbot-prestashop). ## What about price? The honest math The fair comparison isn't "free versus $23": the credits are included with your Wix plan, but you've seen how they work. The real math shows up when you consider **upgrading your Wix plan just to get more AI credits**: compare that monthly step-up against a dedicated flat rate. With Bravos AI, from $23/month you get unlimited messages just for your chatbot, with no shared meter and no shutdowns. The full market numbers are in [how much an AI chatbot costs in 2026](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa). ## So, which one do I pick? It depends on how big a role chat plays in your business. With few inquiries and all your information published on your site, the built-in one may be enough. If chat is part of how you sell, you've seen where its limits are — and that's where a dedicated chatbot pays off. And you don't have to decide blindly: with [Bravos AI](https://bravos-ai.com/) you can set up your Wix chatbot, train it on your site and documents, install it in minutes and watch it answer your real inquiries for 7 days at no cost. If it doesn't convince you, remove it by deleting one line of code. ## Frequently asked questions ### Is Wix's AI chatbot free? Manual mode (replying yourself from the Inbox) is free. AI mode consumes AI credits from your Wix plan: the free plan includes up to 30 a day (120 a month max) and paid plans between roughly 150 and 1,000 estimated actions per cycle, shared with Wix's other AI tools. Wix doesn't guarantee how many conversations they equal, and when they run out the AI chat switches to offline mode. ### Can I install an external chatbot on Wix? Yes, by pasting the widget code under Settings → Custom Code in your Wix dashboard. Wix's requirement: the site must be published with its own connected domain, which means a paid Wix plan. On the free plan (a wixsite.com address) external code can't be added. ### Does an external chatbot work with my Wix store? Yes, with an honest caveat: at Bravos AI we don't (yet) have automatic sync with the Wix Stores catalog, as we do with Shopify, WooCommerce or PrestaShop — it's on our roadmap. Meanwhile, Wix Stores exports your catalog to CSV in a couple of clicks, and that catalog (uploaded directly or as a synced Google Sheets document) is what the bot runs exact filtering on, by price, category or any attribute. ### What languages can a chatbot on Wix answer in? It helps to separate two things. With Smart Chat, what Wix documents is the translation of the widget's interface (the texts the visitor sees). With a dedicated chatbot like Bravos AI, the AI itself detects the language the visitor writes in and answers in it — 13 languages — even if your site is only in English. ### Try a dedicated chatbot on your Wix site Train it on your site and documents, paste it into your Wix site and watch it handle your real inquiries. 7-day PRO trial, no strings: we warn you before charging and if you cancel before day 7 you pay nothing. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Law Firm Chatbot: Automate Client Intake Without Giving Legal Advice **Category:** Use case | **Read time:** 13 min | **Date:** Jul 17, 2026 | **URL:** https://bravos-ai.com/en/blog/law-firm-chatbot "A chatbot on a law firm's website is risky — if it starts handing out legal advice, I'm the one on the hook." It's the first reaction of almost every attorney we talk to. And they're half right: a badly built chatbot is a liability. But the fear comes from picturing the wrong bot — one running loose, opining on someone's case. The one that actually helps a firm does the opposite. First, what this is. A **chatbot** is that chat window that pops up on a website (or on WhatsApp) and answers visitors on its own. When there's real AI behind it, it's not a menu of buttons: it understands plain-language questions and answers with the information you've given it about your firm. In this article we cover what a chatbot does well on a law firm website (and what it should never do), how you configure it so it **doesn't give legal advice**, and what a firm gains by answering intake around the clock. No hype, and no pretending the machine will practice law for you. ## Why the fear is legitimate (and why it's solvable) The worry is reasonable. If a chatbot starts answering "do I have a case?" with a generic response pulled off the internet, it can get it wrong, set false expectations, and drift into the unauthorized practice of law — a real problem with your state bar. No firm wants a machine speaking for it about a matter no attorney has reviewed. Here's the key thing to understand: **a chatbot doesn't decide on its own what to say**. It works from two things you control. One, **what it knows**: only the information you've loaded about your firm — the texts and documents you upload (your practice areas, your fees, how the first consultation works). It doesn't answer with whatever is out on the internet. And two, **how it behaves**: what the industry calls the *system prompt*, the master instructions you give it before it ever talks to anyone — what it can say, what it can't, and how to act when something falls outside that. So you can load your firm's real information and, at the same time, tell it literally: "Do not give legal advice. If someone asks about their specific case, explain that an attorney evaluates that in a consultation and offer to take their details." It answers with your information, within the limits you set. We go deep on this in our guide to the [system prompt for a business chatbot](https://bravos-ai.com/blog/system-prompt-chatbot-empresa). ## The real problem it solves An attorney doesn't make money answering "how much is the first consultation?" twenty times a day. Yet a big chunk of the day goes to exactly that. According to the [Clio Legal Trends Report](https://www.clio.com/resources/legal-trends/) — the sector's benchmark, built on data from tens of thousands of firms — the average lawyer bills just under **3 hours of an 8-hour day**. The rest is eaten by non-billable work: admin, scheduling, and intake questions that repeat over and over. Then there's what happens after hours. Someone with a legal problem — a crash, a firing, an eviction notice — doesn't wait until tomorrow: they call the next firm on the list. That's not a hunch. A study by MIT and InsideSales.com of over 15,000 leads found that responding within the **first 5 minutes makes you 21x more likely** to qualify that lead than waiting 30. A chatbot responds instantly, at 3 a.m. too. A lead that comes in through chat is worth more than one from a form. On a form, the client leaves their name, phone and, if you're lucky, a one-line "subject." Through chat you get that plus the **whole conversation**: what they asked, what details they gave, and how they describe their problem in their own words. The attorney calls back already up to speed, not starting from zero. We break down how to capture without being pushy in our article on [capturing clients with a chatbot](https://bravos-ai.com/blog/chatbot-captar-clientes). ## How you configure it so it doesn't give legal advice This is the heart of it. Two mechanisms, together, keep the bot in its lane: **1. It answers only with YOUR information.** A serious law firm chatbot doesn't answer with what it "knows" from the internet. It uses a technique called RAG (retrieval over your own documents): before answering, it searches the information you've loaded — your fees, your practice areas, how you work — and answers from that. If the answer isn't there, it doesn't make one up. Why that matters so much is covered in [why your chatbot fails and how to fix it](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). **2. It's barred from opining.** Through the system prompt you set the boundary: it doesn't evaluate anyone's specific matter, doesn't say whether someone has a case, and doesn't hand out strategy for their situation. When someone asks about their situation, the bot can give general orientation and defer what's specific to their case: it makes clear an attorney evaluates that, and offers to take their details for a consultation. It genuinely helps without weighing in on the case — you'll see it in the example just below. If someone sells you a "legal chatbot" that **resolves your clients' legal questions**, be skeptical. It shouldn't — for liability and ethics reasons — and it isn't in your firm's interest: the value comes from the attorney, not from an automated answer anyone could copy off a search engine. ## A real example: a car accident inquiry Better to see it than explain it. This is how a well-configured bot behaves when a real inquiry comes in — not a "what are your hours?" but someone with an actual problem on their hands: Notice what it **didn't** do: it didn't say whether the client will win, or what the case is worth. It recognized the practice area, asked just enough to qualify (injured, saw a doctor), gave general orientation, made clear the case evaluation belongs to an attorney, and captured the contact for a consultation. That boundary — help without opining — is exactly what the system prompt sets. ## What a chatbot does well on a law firm website With the off-limits part settled, there's plenty of useful ground — and it's exactly what buries most firms' intake: - **Answer the same questions, always:** fees and the cost of a first consultation, practice areas, location and hours, how the process works, what documents to bring. - **Qualify the inquiry:** ask the practice area (personal injury, family, criminal, employment…) and the type of matter, so it arrives with context, not a bare "I need a lawyer." - **Filter what isn't your area:** if you don't take criminal, the bot can say so politely instead of booking a consultation that made no sense. - **Give general orientation (as far as you want):** it can explain how a process generally works or what usually happens in similar cases. General information — never applied to the specific matter of whoever's asking. - **Capture the contact with all the context:** the bot shows a quick form (name, phone…) and attaches the whole conversation, so the attorney calls back already up to speed. - **Be where the client is:** on the firm's website (the *widget*, the chat window) and also on [WhatsApp](https://bravos-ai.com/blog/chatbot-whatsapp), which many people now prefer for messaging. | What it does | What it does NOT do (and shouldn't) | | --- | --- | | Explain fees and how the first consultation works | Tell you whether you'll win your case | | Give general legal information (how a process works) | Apply that information to your specific case | | Take your details for a first consultation | Evaluate your specific matter | | Answer 24/7 on the web and WhatsApp | Replace the attorney in the client relationship | ## Confidentiality, privilege and ethical rules A firm handles sensitive information and is bound by confidentiality, so this part isn't optional. A well-designed chatbot helps rather than gets in the way, if it's set up with judgment: - **It collects only what's needed** for the firm to make contact (name, phone, type of matter), not a full account of the client's situation over chat. - **It meets data-protection rules:** a data-processing notice and consent before capturing the contact, same as any form on your site. - **It makes clear it's an automated assistant** and not a substitute for an attorney — and it can carry the disclaimers your jurisdiction's ethical rules require. No making a visitor think they're talking to a lawyer. In other words: the bot is the front door — it sorts and filters — but the relationship of trust, and everything touching privilege, stays with the attorney. ## What it does NOT do (and be skeptical of anyone who says otherwise) For honesty's sake, and because it's exactly what protects your firm, it's worth writing down. A law firm chatbot does **not**: - Give legal advice applied to the specific matter of whoever's asking: it won't say if they'll win or evaluate their situation. - Replace the first consultation or judge whether a case has merit. - Draft complaints, contracts or motions for the client. - Decide who's right. That's an attorney's job, not a machine's. ## How to put one on your law firm website Less technical than it sounds, and no coding. Three steps: - **Train the bot** on your firm's information: practice areas, fees, how the first consultation works, FAQs. You upload your documents and it learns from them. - **Configure the system prompt:** set what it doesn't say (no legal advice) and how it defers when someone asks about their case. - **Put it wherever you want:** paste one line of code on your site for the chat widget, or connect your WhatsApp number. Done. ## So, should you put one on your firm's site? The fear of putting a chatbot on a law firm site is legitimate, but it's solved by configuring it well. One that answers only with your information, gives general orientation, doesn't evaluate anyone's specific case, and captures the contact with the whole conversation doesn't replace you: it takes the repetitive intake off your plate and answers the people who show up after hours, when the office is closed. Every firm is different — a PI practice doesn't field the same inquiries as a family attorney — so the best way to know if it fits is to test it with your own. With [Bravos AI](https://bravos-ai.com/) you set one up, train it on your firm's information and watch it respond for 7 days at no cost. And if you use Claude or ChatGPT, with [our MCP connector](https://bravos-ai.com/conector-ia-mcp) the assistant applies the same judgment we use to build the bots to tell you how to tune it to your firm's inquiries. If it doesn't convince you, you walk away. ## Frequently asked questions ### Can a chatbot give legal advice to my clients? No, and it shouldn't. A chatbot configured properly for a firm is barred from opining on specific matters: it shares the firm's information (fees, practice areas, first consultation), gives general orientation, qualifies the inquiry and defers to an attorney. The advice comes from the professional, not the machine. That's enforced through the system prompt, the instructions that set what the bot can and can't say. ### Is it safe with confidential information and privilege? Yes, if it's set up with judgment. The bot should collect only what's needed to make contact (name, phone, type of matter), ask for data-processing consent like any form, carry the disclaimers your jurisdiction requires, and make clear it's an automated assistant. Privilege and the client relationship stay with the attorney. ### Do I need to know how to code to set it up? No. You train the bot by uploading your firm's information, adjust how it behaves from a dashboard, and paste one line of code on your site (or connect WhatsApp). No technical skills required. ### Does it work over WhatsApp? Yes. Beyond the website chat, the bot can handle inquiries over WhatsApp, which many people prefer for reaching a firm. It behaves the same: it informs, qualifies and captures the contact, 24 hours a day. ### Try a chatbot for your firm Train it on your firm's information, configure it so it doesn't give legal advice, and put it on your site in minutes. 7-day PRO trial, no strings: we warn you before charging and if you cancel before day 7 you pay nothing. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # WhatsApp Chatbot with n8n, Make or Zapier: Cost, Hours and Comparison **Category:** Analysis | **Read time:** 14 min | **Date:** May 29, 2026 | **URL:** https://bravos-ai.com/en/blog/whatsapp-chatbot-n8n-make-zapier If you landed here, you already know what **n8n**, **Make** or **Zapier** are, and you are weighing whether to build your WhatsApp AI chatbot on one of them instead of paying for a managed platform (we compare the 13 main ones in [best WhatsApp chatbot](https://bravos-ai.com/blog/mejor-chatbot-whatsapp)). We have tried it ourselves, we have seen customers finish the workflow on week three, and we have seen others abandon on day four. The difference, every time, was the same: **nobody told them what actually had to be built**. The underlying problem: the YouTube tutorial that pops up when you search «n8n WhatsApp chatbot» usually shows you two things (receive a message, send a template reply) and skips the other seven you need for the bot to be useful. Below is the full list, with the time and cost each component takes. ## Why so many devs try to build it on n8n, Make or Zapier Three honest reasons: - **Total control.** You see the nodes, you decide what happens with every message, you wire the bot logic into your CRM, your spreadsheet or your in-house system. No black box. - **Pay-per-use is cheap at low volume.** Self-hosted n8n has zero license cost. Make's 9€/month Core plan goes a long way if you do not receive many messages. - **Learning.** If you build it, you understand it, and you can change it later without depending on anyone. All three are valid. The trouble starts mid-way through week two, when you realize the tutorial covered **20%** of what you need and the other 80% is on you. ## What you actually need to build a WhatsApp chatbot with n8n: the 9 components This is the full component list for a WhatsApp AI chatbot (or WhatsApp AI agent — the same thing) that answers with your business data. Whichever low-code platform you pick, all nine are required. Public tutorials cover blocks 1 and 8 well. The other seven you build yourself. The ones that surprise people most: ### Connecting to the WhatsApp Cloud API is not a toggle Two clocks run in parallel here. The first is your **technical work**: **four to eight hours** the first time if you have never touched [Meta for Developers](https://developers.facebook.com/docs/whatsapp/cloud-api/webhooks/). Create a Meta app, set up a WhatsApp Business Account, generate an access token, register the phone number, verify the webhook with a hub_verify_token, and stand up a public HTTPS endpoint. On self-hosted n8n, that endpoint means a real domain with valid SSL, not localhost. The second clock is **Meta's side**, which you cannot speed up. When you create the app, Meta gives you a **test phone number** you can use the same day to receive and reply to messages (no card, no verification), capped at **250 messages every 24 hours** and a small fixed allowlist of recipients. To move to your real business number with real volume, you go through **Meta Business Verification**: upload incorporation papers, tax ID, a utility bill, and wait **2 to 10 business days** for approval. While you wait, the bot can already be answering on the test number if you accept that as a temporary setup. ### The AI brain: system prompt alone is not enough The **system prompt** is the set of instructions you give the AI model at the start of every conversation: who it is, what it does, and what information about your business it has on hand. It is the block most people underestimate. There are two paths: - **System prompt only.** All business info stuffed inside the system prompt. Works for very basic Q&A (hours, address). Falls apart with [product catalogs](https://bravos-ai.com/blog/chatbot-catalogo-productos) or long documents: the model cannot keep thousands of words of context on every call without blowing up cost and latency. - **RAG-based.** Instead of stuffing everything in the system prompt, you store your content in a **vector database** (a type of database designed for searching by meaning, not by exact words; typical products: Pinecone, Supabase or Chroma). Before storing, each chunk of your content is converted into a numerical vector (those are the *embeddings*). On every customer question, the system retrieves the relevant chunks and only passes those to the model. This answers with your real content, but **adds three external services** to the list and tuning retrieval well takes real time. The typical tutorial builds the bot with system prompt only. It looks fine in the demo, until a real customer asks about a specific product and [the bot makes up the answer](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). ### Conversation memory most builders forget WhatsApp does not remember anything by itself: every inbound message runs your workflow from scratch. If a customer writes «do you have size 42?» after a previous message asking about shoes, the bot will not remember it was talking about shoes unless *you* store the history somewhere. On self-hosted n8n you do that with a separate database (Postgres or Redis are the usual picks). On Make or Zapier you need an external database hired separately (Airtable, Supabase, similar). Yet another piece the tutorial never mentions. ## How much does a WhatsApp chatbot with n8n cost per month (line-by-line) Assuming [small-business](https://bravos-ai.com/blog/chatbot-para-pymes) volume (500 to 1,500 inbound messages per month), this is the real monthly cost once the bot is live: | Component | n8n self-hosted | Make | Zapier | | --- | --- | --- | --- | | Low-code platform | $0 (license) | $10/mo Core | $19.99/mo Pro | | VPS hosting (n8n) | $5-15/mo | — | — | | Domain + SSL | $1-2/mo | — | — | | OpenAI calls | $15-45/mo | $15-45/mo | $15-45/mo | | Vector DB for RAG | $0 (pgvector) or $0-20/mo Pinecone | $0-20/mo Pinecone | $0-20/mo Pinecone | | DB for conversation memory | included in VPS | $5-10/mo Supabase | $5-10/mo Supabase | | Meta WhatsApp (within 24h window) | $0 | $0 | $0 | | Approx. monthly total | $21-77 | $30-85 | $40-95 | Ranges depend on real volume, the OpenAI model you pick (mini variants like gpt-4o-mini are cheap and fast; larger models like gpt-5.5 cost more but handle complex questions better), and whether you need a paid vector DB or you can survive on pgvector. **In dollars, self-hosted n8n is the cheapest path**. Make and Zapier land close to the [price of a managed platform](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa) at mid-tier without saving any of the build hours. ## How long does it take to build a WhatsApp chatbot with n8n What we have seen with developers who have prior REST API experience but no prior Meta or RAG work: - **Connecting to the WhatsApp Cloud API (app, webhook, token):** 3-5 hours. - **Linear flow to receive and reply plain text with no AI:** 1-2 hours (this is all the tutorial covers). - **OpenAI integration with system prompt only:** 1-2 hours. - **Real RAG (upload content, vector database, retrieval tuning):** 4-8 hours. - **Conversation memory with database persistence:** 1-3 hours. - **Inbound media handling** (transcribe audio with Whisper, read images with GPT-5.5 or Claude vision): 2-4 hours. - **Token renewal, monitoring, alerts:** 1-2 hours. - **Total for a serious bot:** **13-26 hours** the first time. Without RAG, more like 8-15. If you have never touched *Meta for Developers* or built a RAG before, multiply the ranges by 1.5 or 2. If you already know both, take 30% off. Add the **maintenance overhead**: every time Meta changes the webhook (two or three times a year), you go in again. Every 60 days you renew the access token unless you have set up a System User token with persistent permissions, which is another documentation deep-dive. ## WhatsApp Cloud API pitfalls no tutorial mentions ### Tokens expire every 60 days Unless you use a Meta [System User token](https://developers.facebook.com/docs/marketing-api/system-users/overview/) (which requires being a Meta-verified business or Tech Provider), the standard access token expires every 60 days. When it expires, the bot stops replying and you find out because a customer complains. You need to automate the renewal with a separate process. Another component on the diagram. ### Inbound media is a separate download When a customer sends a photo or audio over WhatsApp, Meta does not include the file in the webhook. It sends a **media_id**. To get the real file, your workflow needs to fire another WhatsApp Cloud API request, download the file, store it on your end, and only then pass it to the AI model. If you want the bot to understand audio, add Whisper (OpenAI's speech-to-text model). If you want it to understand images, add GPT-5.5 vision or Claude vision. ### The 24-hour window exists regardless Meta's rule about not initiating conversations outside the 24-hour window (unless via approved templates) applies the same way whether you build with n8n or pay for a managed platform. DIY does not free you from it. If your bot only replies (never initiates), it never affects you. If you want to send broadcast messages or follow-ups, you need to learn Meta's template system and its approval process. ### Logs and debugging when something fails Self-hosted n8n gives you per-workflow execution logs. Make has plan-limited operation history. Zapier's failed task debugging is sparse. When a customer reports you never replied, debugging without proper logs is debugging blind. Managed platforms give you this in the dashboard; on a DIY build you wire it yourself. ## n8n vs Make vs Zapier for WhatsApp: which one fits your case All three can connect to the WhatsApp Cloud API and power a WhatsApp AI agent or chatbot, but they differ in cost, onboarding effort and how much headroom they give you to grow: | Aspect | n8n self-hosted | Make | Zapier | | --- | --- | --- | --- | | Base monthly cost | $6-17 (server included) | $10/mo Core | $19.99/mo Pro | | Onboarding difficulty | High (server + domain + n8n) | Medium (account + modules) | Low (account and you are off) | | Monthly operation/task cap | No cap (server is the limit) | 10,000 ops on Core | 750 tasks on Pro | | Built-in AI? | No, you add OpenAI manually | No, you add OpenAI manually | No, you add OpenAI manually | | Best for | Developer with own server and full control | Small business at medium volume without their own server | Quick tests or tiny volume | In short: **self-hosted n8n** is the most flexible and cheapest long term, but you need to know how to run a server with HTTPS. **Make** is the reasonable middle ground if you do not have a server. **Zapier** ends up expensive because WhatsApp consumes tasks fast and almost always pushes you to a higher plan. If you plan to run multiple bots or your monthly volume goes over 1,500 messages, self-hosted n8n wins on cost every time. If you want to set it up in one afternoon without touching a server, Make is the fastest path. ## When DIY actually makes sense There are cases where DIY is the right call. These: - **Your internal system cannot push events when something changes.** If your ERP or SAP system can only be read by direct query (you make a call and it answers, but it cannot notify anyone when an order changes state), building it yourself on n8n lets you make that call right at the moment of the message with an HTTP node inside the flow. At Bravos AI we have **Custom Webhook** and it runs in real time, but it works the other way around: your system needs to **notify us** when something changes. If your ERP can push, we cover you in real time; if it can only be queried, that is where DIY wins. - **Your volume is tiny and predictable.** 50 messages a month, one conversation type. At that volume, paying a monthly subscription does not pay off. - **You want to learn how the WhatsApp Cloud API works under the hood.** Legitimate reason. Just do not confuse it with a business case. For everyone else — small business with real volume, owner who is not a developer, agency deploying for multiple clients — the math does not work. The headline price of Make or n8n does not offset 15 build hours or 2-4 monthly maintenance hours. ## Closed alternative: WhatsApp AI chatbot ready in 60 seconds At [Bravos AI](https://bravos-ai.com/) we take the nine components off your plate and bill them as a flat fee. We are a **verified Meta Tech Provider**, so connecting your WhatsApp Business number is a 60-second popup where you log in with Facebook and pick the phone number. No Meta for Developers app to create, no Meta webhook to configure, no token to renew, no server to maintain. We handle the 9 components on the list, not you. The bot pulls from the real catalog connected in the dashboard and shares a booking link with the product reference pre-filled. The real catalog is connected once: upload a CSV or sync your Shopify or PrestaShop store. From that point on, the bot replies with real stock, real price and real availability, not invented ones. If it does not know, it says so. And if you have an ERP, an SAP system or an internal database, you get **Custom Webhook**: your system sends us data signed with a secret key and everything you sync becomes available to the bot. We cover this topic in depth in [how to keep your chatbot up to date without Zapier or n8n](https://bravos-ai.com/blog/mantener-chatbot-actualizado). Want to add a PDF with return policies? Upload it and you are done. Cost: **$23/month** Starter (WhatsApp and analytics included) or **$59/month** PRO if you need up to 3 bots or built-in lead-capture forms. Nothing extra from Meta as long as you reply within the 24-hour service window opened by each customer message — which is exactly what a reactive bot does. New accounts start with a **7-day PRO trial** that lets you test everything, WhatsApp included, with no charge until day 7. ## FAQs ### Can I build a WhatsApp AI chatbot with n8n? Yes, technically. n8n has an official WhatsApp Business Cloud node that connects to the WhatsApp Cloud API. The node covers blocks 1 and 8 of the 9 (inbound webhook and outbound send). The other seven you build with extra nodes: the AI brain, conversation memory with persistence, inbound media handling, token renewal and logging. Viable if you have 13-26 hours and a server to host n8n with a proper HTTPS domain. ### Is Make.com cheaper than a managed WhatsApp chatbot platform? On the headline subscription, Make ($10/month Core) looks cheaper. Once you add the required extras (OpenAI calls, vector database for RAG, database for conversation memory), the real monthly cost lands at $30-85, very close to a managed platform that already bundles everything. And it does not cover the 13-26 build hours or the maintenance. Zapier usually ends up more expensive because WhatsApp consumes tasks quickly and forces you to a higher plan. ### How long does it take to build a working WhatsApp AI bot from scratch with n8n? For a developer with REST API experience but no prior Meta or RAG work: 13-26 hours for a serious bot with AI, RAG and memory. Without RAG (system prompt only), about 8-15 hours, but the result is notably weaker for product catalogs. For someone without prior technical experience, the timeline blows up because Meta for Developers has a steep learning curve. ### Why does the WhatsApp token need renewing every 60 days? Meta's standard user access token has a 60-day expiry by design. When it expires, every WhatsApp Cloud API call starts failing and the bot goes silent. To avoid it, configure a System User token (requires being a Meta-verified business) or schedule an automated task that renews the token before it expires. At Bravos AI, since we are a verified Tech Provider, we use a System User token at the platform level: the end customer never touches tokens. ### Can I integrate ChatGPT with WhatsApp using n8n? Yes. n8n has an official OpenAI node that calls the API and returns the model's response. You wire it to the WhatsApp Business Cloud node so the flow is: inbound message → OpenAI call with the customer context → reply back to WhatsApp. It works for one-off questions but, if you want the bot to answer with your real business data (catalog, prices, hours), you also need RAG: upload your content, generate embeddings, query them on every call. That is blocks 5 and 6 of the 9 this article details. ## Bottom line Building a WhatsApp AI chatbot with n8n, Make or Zapier is **technically possible** and, on raw subscription dollars, can look cheaper than a managed platform. But the **nine components** you have to build, the **13-26 hours** it takes the first time, and the **monthly maintenance** on tokens, media files and logs turn the savings into something more nuanced than the headline suggests. If you are a developer and need the bot to query an internal system in real time during the message, self-hosted n8n is a reasonable call. For everyone else, at Bravos AI we take most of the work off your plate: the 9 components come built-in, token renewal included, and if you have internal data you wire it via Custom Webhook. If you want to give us a try with everything included for 7 days, [start here](https://bravos-ai.com/). For more context, check our piece on [how to build a WhatsApp AI chatbot in 2026](https://bravos-ai.com/blog/como-hacer-chatbot-whatsapp) (the four real paths) or [how to auto-reply on WhatsApp](https://bravos-ai.com/blog/responder-whatsapp-automaticamente). ### Want your WhatsApp chatbot live in 60 seconds instead of 15 hours? Build your bot on Bravos AI, connect it to WhatsApp Business and let the AI answer with the real info from your business. Without building the 9 components from scratch. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # WhatsApp Auto Reply in 2026: How to Set It Up (Out-of-Office, Business and AI) **Category:** Practical guide | **Read time:** 13 min | **Date:** May 28, 2026 | **URL:** https://bravos-ai.com/en/blog/whatsapp-auto-reply When people search for «WhatsApp auto reply», they usually mean three very different things hiding under the same name. The away message you can set inside the WhatsApp Business app, the rule-based bots people glue together with Zapier, and an AI agent that actually understands what the customer is asking and answers with your real business data. We walk through all three, when each one is enough, and what to use for an actual business in 2026. Why this matters: a [Harvard Business Review study of 1.25 million sales leads](https://hbr.org/2011/03/the-short-life-of-online-sales-leads) found that companies that reply to a customer **within an hour** are **7x more likely** to have a meaningful conversation with a decision-maker than those who wait even one hour more. WhatsApp customers expect a faster reply than email — closer to minutes than hours. Every hour of silence is a sale cooling off. ## The three real ways to auto reply on WhatsApp One name, three very different levels. Most people only know the first one — that is why so many businesses end up «auto-replying» with a static text and then wonder why they still lose customers. ### Level 1: the static away message — WhatsApp's built-in auto reply The official WhatsApp Business app on your phone lets you set two fixed messages: a greeting (sent on first contact) and an **away message** for outside business hours. Third-party Android apps like WhatsAuto do something similar for personal WhatsApp. It is **free**, takes five minutes to set up and covers the bare minimum so you do not look like a business that ignores its customers. That is it: it does not understand what the customer is asking, it does not pull from your business data, and it sends the same text to everyone. A sign that mails itself. ### Level 2: rule-based replies (keyword triggers) A step up. You wire «if customer types X, reply Y» — usually with n8n, Zapier, Make or platforms like Manychat. If the customer types «hours», the bot answers with your hours; if they type «price», it sends a PDF. **Works** when your customers always ask the same four things using the same four words. Breaks the moment someone phrases it differently: «what time do you open?» is not «hours», and the bot says nothing. Maintenance is heavy and you never cover every case. ### Level 3: an AI that understands and replies with your real business data An AI agent connected to your WhatsApp Business number. It does not reply the same to everyone, and it does not work off keywords. **It reads the question**, retrieves the answer from the business information you uploaded (your website, a PDF, a product CSV), and replies the way a person who had read all your documents would. If someone asks «are you open Saturday evening?», it answers with your actual hours. If they ask «do you have washer-dryers under $800?», it pulls from your catalog and lists the ones that fit. This is no longer «auto reply» in the old sense. It is full customer service, 24/7, without hiring anyone. ## Out-of-office on WhatsApp: what works and what does not A lot of people land on this topic looking for a clean out-of-office message on WhatsApp — the equivalent of the email autoresponder they already use. The good news is there is a native way to do it. The bad news is the native way is very limited, and **there is a much better alternative** once you understand the difference. ### The native away message in WhatsApp Business Inside the free WhatsApp Business app (Android or iPhone) you can set an **away message** that auto-sends when you are outside your business hours. You go to the app settings, open *Business tools*, tap *Away message*, write the text, pick the schedule (always away / custom schedule / outside business hours), and choose who receives it (everyone, only people not in your address book, etc.). It works, and it is free. ### What the native away message cannot do One big limit: it is the **same text for every customer**. Whether they ask about your hours, your pricing, an order or a refund, they all get the same line — «We are away, we will get back to you soon». For a small shop closing at night that is fine. For a business that genuinely wants to look after its customers outside business hours, it is acknowledgment, **not service**. The customer still leaves the chat with no answer. ### The modern replacement: an AI that actually answers The away message exists because, in 2018, the only thing a phone-based business app could do automatically was send a fixed text. In 2026, with a WhatsApp Business AI chatbot, your number is covered 24/7: at 11 PM the customer asks «are you open tomorrow?» and the bot answers with your real hours. Asks «do you ship to Canada?» and the bot answers from your shipping policy. No more «we are away» messages — you actually solve the customer's problem while you sleep. That is the upgrade from out-of-office to always-on. ## Auto reply on personal WhatsApp: the ban risk Worth being upfront about this: some searches for «WhatsApp auto reply» come from individuals who want to automate their **personal** WhatsApp account, not a business one. There is no clean way to do that. Apps like WhatsAuto and AutoResponder for Android sit in the background and send fixed replies to incoming messages on personal WhatsApp. Three problems: - **Not official.** They use Android accessibility tricks; any WhatsApp update can break them. - **Ban risk.** WhatsApp explicitly does not allow automating personal accounts with external apps. If they detect automated bulk sending, your number can be **blocked**. For one-off use (vacation away message, while you are driving, in a meeting) they may be acceptable if you take the risk. For anything resembling a business, the right move is to switch to **WhatsApp Business** (free, separate app) and, if you want a real reply, plug it into an AI platform. ## WhatsApp Business auto reply: from a fixed text to an AI agent For businesses, there are two layers worth knowing: what the free WhatsApp Business app gives you out of the box, and what you can do once you connect the number to the WhatsApp Business Cloud API. ### In the WhatsApp Business app (free, phone-based) Greeting message + away message + quick replies (manual one-tap shortcuts). Fine for a shop closing at night, no customer-service ambition. Setup: 5 min. Cost: $0. ### With WhatsApp Business Cloud API + AI chatbot A different category. Your number moves to the cloud, and a chatbot platform ties it to an AI agent grounded on your business data. Every incoming message gets an automatic reply in seconds with the actual information from your website, PDFs and product catalog. Multilingual, handles photos / audio / PDFs from the customer, 24/7. Setup: ~10 min. Cost: from $23/month, with no extra Meta fees for customer-initiated replies since July 2025. The cleanest way to plug into the Cloud API is through a **Meta-verified Tech Provider** — which is what we are at [Bravos AI](https://bravos-ai.com/) . That means the integration is **direct with Meta**: no Twilio, no third-party BSP middleware, no per-client tokens that expire every 60 days, no surcharge on each message. The whole setup is a single 60-second popup. Most other platforms in the market route through middleware, which adds days of extra configuration and a recurring per-message fee on top of Meta's own pricing. ## Real WhatsApp auto reply conversations (AI in action) Two conversations your bot would handle on its own, no human input needed: Neither of these answers could be produced with a static text or a keyword rule. They need an agent that **understands** the question, has access to your business information (hours, product manuals) and can interpret what the customer sends in a photo. That is the difference. ## How to set up an AI-powered auto reply with Bravos AI (step by step) **1.** Sign up at [app.bravos-ai.com](https://app.bravos-ai.com/register?lang=en) and create your first bot. Load it with your business information: website, PDFs, a product spreadsheet, or connect [Shopify](https://bravos-ai.com/blog/chatbot-shopify) or [PrestaShop](https://bravos-ai.com/blog/chatbot-prestashop) . **The AI learns on its own**, no flow building, no keyword rules. **2.** Activate the Starter plan (**$23/month**), which is where the WhatsApp channel lives. The **7-day PRO trial** also lets you test WhatsApp at no cost before deciding. **3.** Go to **Integrations** in the sidebar, open the **Available** tab and click the WhatsApp card. Pick the bot to wire WhatsApp to, then a Meta popup opens: sign in with Facebook, pick (or create) your Meta Business account, pick the phone number and confirm with an SMS code. Total time: about **60 seconds**. The connection screen before the Meta popup: pick which bot the WhatsApp number is wired to, read the pre-flight notes (Facebook account, clean phone number, no payment method needed) and hit Connect WhatsApp. } /> **4.** Done. From this point on, every message that hits your WhatsApp Business number is answered by the bot using the information you uploaded. You see every conversation from the dashboard —or, if you use Claude or ChatGPT, from the chat itself with [our MCP connector](https://bravos-ai.com/conector-ia-mcp)—; if you ever need to handle a customer yourself, you have to **disconnect the integration**, which returns the number to your WhatsApp Business app on the phone so you can reply from there. ## What it costs and how much time it saves **Monthly cost:** $23/month for the Bravos AI Starter plan. Nothing else on top from Meta: [since July 2025](https://help.twilio.com/articles/30304057900699-Notice-Changes-to-WhatsApp-s-Pricing-July-2025) , replying to a customer-initiated conversation is **free with no monthly cap**. The bot only replies (it does not initiate conversations or send templates), so your bill does not move with volume. **Time saved (real numbers):** a small business handling 100–300 WhatsApp messages a month spends 8–20 hours typing the same answers (hours, pricing, stock, returns). At a fully-loaded labor cost of $20–25/hour, that is **$160 to $500 a month** of staff time burned on repetitive replies. The bot pays for itself the first month and frees that time for higher-value work. For a deeper cost breakdown across [WhatsApp platforms](https://bravos-ai.com/blog/mejor-chatbot-whatsapp), see [our AI chatbot pricing guide for 2026](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa) . ## Common pitfalls ### Settling for the static away message and calling it customer service The most common one. Setting a «Thanks, we will get back to you shortly» inside the WhatsApp Business app and pretending you are now «covered after hours». That is not customer service — it is an **auto-receipt**. If your customer has a concrete question and only gets a placeholder, they move on to a competitor who actually answers. ### Trying to automate personal WhatsApp with unofficial apps Apps like WhatsAuto can get your number suspended if WhatsApp detects automated bulk sending. If it is for a business, the right path is moving to WhatsApp Business (free, separate app) and, for AI replies, plugging it into a serious platform with official Meta integration. ### Loading messy content and blaming the bot Five outdated PDFs and an out-of-date website URL in, vague replies out. **Thirty minutes** cleaning your sources before the upload is the difference between a useful bot and one that hedges. ### Not testing before going live Before connecting the bot to your real WhatsApp Business, test it on your website for a couple of days. Ask it the five things customers ask you most. If it answers well, connect WhatsApp. If not, tweak the content and test again. **Half an afternoon** of testing now prevents months of mediocre replies to real customers later. ## Frequently asked questions ### How do I set up an auto reply on WhatsApp? Two paths depending on what you want. For a static text (away message), install the free WhatsApp Business app, go to *Business tools → Away message*, write the text and pick a schedule. For a real AI reply that understands the customer and answers with your business data, sign up on an AI chatbot platform (with Bravos AI it is a **60-second popup** to connect your WhatsApp Business number directly to Meta), upload your business information, and the bot replies on its own from that point. ### How do I set an away message on WhatsApp? In the WhatsApp Business app (not the regular WhatsApp): open the app, tap the three-dot menu, go to *Business tools → Away message*, toggle it on, write the text, pick the schedule (always sent / custom schedule / outside business hours) and pick the recipients. Important note: if your number is connected to the WhatsApp Business Cloud API (which is what AI chatbots use), the away message setting is no longer available — the bot replaces it. ### Can I auto-reply on my personal WhatsApp? Only with unofficial third-party apps (WhatsAuto, AutoResponder, etc.) and only for static text. WhatsApp **does not allow** automating personal accounts, and your number can be banned if mass automated sending is detected. For business use, switch to WhatsApp Business (free, separate app) and, for AI-powered replies, plug it into a real platform. That is the only safe path. ### What is the best WhatsApp auto responder? Depends on what you mean. For a static text, the WhatsApp Business app itself is fine and free. For personal WhatsApp, third-party Android apps like WhatsAuto (with the risks above). For a business that wants real AI replies, a platform with **direct Meta Tech Provider integration** like Bravos AI is the cleanest path: no Twilio, no tokens that expire every 60 days, and the bot answers with the actual information from your business. ### What is the best auto reply message for WhatsApp Business? If you are stuck with a static away message, two rules: **be specific about when you will reply** («We reply within 4 business hours» beats «We will get back to you soon»), and **point the customer to where they can get an answer right now** (a link to your FAQ, your hours page, your phone number for urgent issues). The default «Thanks for reaching out» with no extra information is the least useful version because the customer leaves the chat with nothing. ### How much does a WhatsApp auto reply cost? The native away message in the WhatsApp Business app is **free**. An AI-powered auto reply starts at around **$23/month** (Bravos AI Starter, unlimited messages). Meta does not charge for customer-initiated conversations since July 2025, so the bill stays flat at $23/month for a typical reply-only use case. Other European-market platforms start higher (mid-tier from ~$49/month, enterprise from ~$99/month per seat). ## Bottom line There are three real ways to set up an auto reply to WhatsApp in 2026, and they are not interchangeable. The **static away message** in the WhatsApp Business app is fine to avoid looking absent, but it is not customer service. **Keyword rules** work when customers always ask the same things the same way, and break the moment they do not. **AI grounded on your real business data** is the only one that actually answers: 24/7, in any language, with real information. If it is for personal WhatsApp, skip the unofficial apps: the ban risk is real and the most you get is a static reply. If it is for a business, the only clean path is WhatsApp Business + a platform with **direct Meta integration**. Bravos AI is **$23/month flat**, no Twilio, no middleware, no tokens that expire, and the bot answers your customers with your actual business data. For the technical detail behind each route (short, long and engineering), see [our step-by-step guide to building a WhatsApp AI chatbot in 2026](https://bravos-ai.com/blog/como-hacer-chatbot-whatsapp) . ### Ready to let WhatsApp answer on its own? Create your bot in Bravos AI, connect it to WhatsApp Business from the Starter plan, and let AI answer with the real information from your business. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # WhatsApp AI Chatbot in 2026: How to Build One Without Code **Category:** Practical guide | **Read time:** 13 min | **Date:** May 26, 2026 | **URL:** https://bravos-ai.com/en/blog/whatsapp-ai-chatbot A WhatsApp AI chatbot in 2026 is not the technical project it used to be. You no longer need a developer, a server or a stack of glue tools. You also do not need to settle for a button-based bot that replies «press 1 for hours, 2 for pricing». There are three real ways to build one today, and only one of them makes sense for a business that just wants better customer conversations on WhatsApp. And it is worth paying attention now. According to [Pew Research](https://www.pewresearch.org/internet/2025/11/20/americans-social-media-use-2025/) , **32% of US adults now use WhatsApp**, up from 23% in 2021. Globally, according to [Meta](https://investor.atmeta.com/investor-events/event-details/2025/Q1-2025-Earnings-Call/default.aspx) , WhatsApp passed **3 billion monthly active users** in 2025. If you sell to Hispanic households, international customers, or anyone who already lives outside iMessage, WhatsApp is where your buyers already are. The only question is whether you let messages pile up or you put an AI chatbot to work. ## The three real ways to build a WhatsApp AI chatbot ### 1. The long way: n8n, Make or Zapier plus a separate AI agent The DIY route. You connect WhatsApp Business to an automation tool (n8n, Make, Zapier) through a middleware provider (Twilio or a Meta-authorized BSP), wire the conversation flow by hand, plug in a call to a language model (OpenAI, Anthropic), and keep the whole thing running. If you are technical and enjoy the build, fine. For a normal business that just wants better customer service, it is **laborious**, **brittle**, and the bill is split across **multiple subscriptions** (the automation tool, the AI model, the WhatsApp middleware, hosting if you need it). ### 2. The technical way: Meta's Cloud API plus your own code The full-engineering route. You register as a developer with Meta, set up the WhatsApp Business Cloud API, run your own server, manage thread state, template approvals, webhooks, the LLM conversation, retries, error handling, the lot. Only makes sense for **large companies** with very specific use cases. Most small and mid-size businesses do not want to be in this lane. ### 3. The short way: a purpose-built WhatsApp AI platform with direct Meta integration (this is what we recommend) This is what we do at [Bravos AI](https://bravos-ai.com/) . You upload your business knowledge (website, PDFs, a product spreadsheet, or you connect Shopify), then plug your WhatsApp Business number into Meta straight from the dashboard. From that moment, any message that hits your number is answered by an AI agent that **actually understands** what the customer is asking and replies with **real data** from your business: pricing, hours, products, return policy, anything you have loaded. And if you use Claude or ChatGPT, with [our MCP connector](https://bravos-ai.com/conector-ia-mcp) the assistant helps you get that bot ready the way our team would: what content is missing, how to word the answers better and what tone to give it. The difference with most of the market is that we are a **Meta Tech Provider**: the integration is **direct**, no Twilio, no third-party fees per message. Plans start at **$23/month** (around €19) with unlimited messages. Step-by-step setup below. ### Quick comparison: time, money, effort | Path | Time to live | Monthly cost | Needs a developer | Answers with your data | | --- | --- | --- | --- | --- | | n8n / Make / Zapier + AI | Days to weeks | Stacked subscriptions (~$40-120/mo) | Yes, or many hours of your own time | Yes, if you wire it well | | Meta Cloud API + custom code | Weeks to months | Engineering time + hosting | Yes, mandatory | Yes, if you code it | | Bravos AI + WhatsApp | ~10 min (no middleware) | From $23/month | No | Yes, with real AI | ## ChatGPT on WhatsApp vs a real WhatsApp AI chatbot for business A lot of people search for «ChatGPT WhatsApp» thinking they can drop the same chatbot you use in a browser into their business number. That is a different product with different limits. ChatGPT on WhatsApp (or a generic LLM plugged into WhatsApp through Zapier) **does not know your business**: it does not have your prices, your stock, your delivery policy, your opening hours, or your last terms of service. It can chat in a friendly tone and that is about it. A real WhatsApp AI chatbot for business does three things a vanilla ChatGPT cannot: - **Grounded answers.** Every reply is based on the content you loaded (website, PDFs, product feed). Less hallucinations, no «I do not know about that company». - **Live product data.** If you sell, the chatbot can answer «do you have running shoes under $80 in size 10?» with real, in-stock results from your Shopify or your CSV. - **Lead capture inside WhatsApp.** Native forms (Meta calls them Flows) so a buyer can leave name, phone and message without ever leaving the chat. ChatGPT on WhatsApp is fine for personal use. For a business, you want the conversational quality of ChatGPT plus the business knowledge of your CRM. That is the gap purpose-built platforms fill. ## Is there a free WhatsApp AI chatbot, really? Short answer: **not really**. The official WhatsApp Business app on your phone is free and lets you set a static greeting and an out-of-hours auto-reply. That covers the bare minimum politeness, but it is **not a chatbot**: it does not understand the question, it does not pull from your knowledge, and it sends the same text to everyone. Anything beyond that has a cost. We dug into this in another piece where we [asked Tidio, Crisp and Freshchat's own chatbots](https://bravos-ai.com/blog/tidio-crisp-freshchat-plan-gratis) whether their free plans include real AI. Their own bots admitted: no, or very little and very capped. The same applies to WhatsApp: today, **no serious AI chatbot platform for WhatsApp is genuinely free in continued use**. One more thing worth knowing: **Meta itself does not charge you when you reply to a customer**. Since [July 2025](https://help.twilio.com/articles/30304057900699-Notice-Changes-to-WhatsApp-s-Pricing-July-2025) , service conversations (when the customer messages you first) are free with no monthly cap. Meta only charges if **you message first** using an approved template (reminders, shipping updates, marketing). For a business that just answers incoming messages — which is the typical use case — your Meta bill is **$0**. ## What you need before you start ### A dedicated phone number A number that can receive SMS or a voice call for verification, and that is **not already active** in the regular WhatsApp app. Most businesses buy a new number for this (a cheap second SIM, an eSIM, or a Twilio-style virtual number); your business landline also works if it has never been on WhatsApp. ### A Meta Business account The hub where Meta groups your Facebook, Instagram and WhatsApp accounts. It is **free** at [business.facebook.com](https://business.facebook.com). If you have a Facebook business page, you already have one linked. Worth knowing: you do **not** need to verify your business with Meta to start replying to customers — verification is only required for advanced features (in-chat lead forms or high-volume marketing sends). ### Content for the AI to learn from This is the step most people skip. A WhatsApp AI chatbot is only as good as the information you feed it. If you want it to answer questions about pricing, hours, products, services or returns, you need to upload that information: your website, a PDF, a product CSV, an FAQ. In Bravos AI you drop it all in the dashboard and **the AI learns on its own** — no flow builder, no manual Q&A pairs. ### A product source if you sell things The cleanest path is connecting your store directly. Bravos AI has native integration with [Shopify](https://bravos-ai.com/blog/chatbot-shopify) and [PrestaShop](https://bravos-ai.com/blog/chatbot-prestashop) : one click and your catalog syncs daily. That lets the chatbot answer real, in-stock questions like «do you have running shoes under $80 in size 10?». No store? Upload a CSV or Google Sheet — see [the different sync paths we support](https://bravos-ai.com/blog/mantener-chatbot-actualizado) . ## How to set it up with Bravos AI, step by step **1.** Sign up at [app.bravos-ai.com](https://app.bravos-ai.com/register?lang=en) and create your first bot. Load it with your business information: website, PDFs, a product spreadsheet, or connect Shopify/PrestaShop. **The AI learns on its own**, no flow building, no scripted answers. At this point you already have a working chatbot for your website. **2.** Activate the Starter plan (around **$23/month**), which is where the WhatsApp channel lives. The **7-day PRO trial** also lets you test WhatsApp at no cost before deciding. **3.** Go to **Integrations** in the sidebar, open the **Available** tab and click the WhatsApp card. Pick the bot you want WhatsApp wired to, then a Meta popup opens: you sign in with your Facebook account, pick (or create) your Meta Business account, choose the phone number, and confirm with an SMS code. The whole flow takes about **60 seconds**. Here is the key part: we are a **Meta Tech Provider**, so the integration is direct and single-step. **No Twilio**, no middleware, no extra per-message fees. Most other platforms in the market do go through middleware — that means days of extra setup and a surcharge on every message. WhatsApp connection screen inside the Bravos AI dashboard: pick the bot, read the Meta pre-flight notes, then click Connect WhatsApp. } /> **4.** Done. From this point on, every message that hits your WhatsApp Business number is answered by **the same bot** you have on your website, with the same knowledge. One agent, **two channels**: update prices in the dashboard once and they update everywhere. And in either channel the customer can send photos (a label, an error code, a product picture), voice notes or a PDF — the bot understands them and answers about them. ### What you get out of the box **Lead capture inside the chat.** When someone shows real interest (a property, a service quote, a treatment), the bot can pop a native WhatsApp form right there (Meta calls them Flows) with the fields you choose (name, phone, message). The customer fills it in **without leaving WhatsApp** and the lead lands in your dashboard and, if you want, in your inbox. Zero external form, zero copy-paste. **Multilingual answers, no retraining.** You load your content **once** (in any language) and the bot [replies in the customer's language](https://bravos-ai.com/blog/chatbot-multiidioma) automatically. Spanish in, Spanish out. German in, German out. No duplicate bots, no translated documents. **Automatic catalog sync.** Connect Shopify or PrestaShop and your catalog syncs daily. New product live this morning, the bot answers about it on WhatsApp **the same morning**. **Every conversation visible from one dashboard.** You see every conversation (web and WhatsApp) in a single panel and can read them at any time. If at some point you want to handle a customer yourself, you need to **disconnect the integration**, which returns the number to your WhatsApp Business app on the phone so you can reply from there. There is also a **pause** option, but it only silences the bot: any messages that arrive while paused stay unanswered until you re-enable the bot or disconnect. ## How Bravos AI compares to Manychat, Wati and others A lot of people land here after searching for «Manychat WhatsApp» or «Wati alternative», so a quick reality check: - **Manychat**: strong on Instagram and Facebook Messenger; on WhatsApp it goes through Meta as well but the bots are **flow-based**, not AI-grounded on your knowledge by default. Plans from around $15/month for basic, more for AI tools. - **Wati / AiSensy**: WhatsApp-first platforms popular in India and the Middle East. Strong on marketing broadcasts. AI conversational quality varies; plans typically $40-80/month. - **Respond.io**: multichannel inbox plus its own AI agent. Aimed at mid-market, plans from around $99/month. - **Bravos AI**: AI-first, grounded on your business content (RAG over your PDFs, website, catalog), direct Meta Tech Provider integration (no middleware fees), same bot on web and WhatsApp. From $23/month. The honest summary: if you want flow-builders and broadcast campaigns, Manychat or Wati fit that brief. If you want an AI chatbot that answers like ChatGPT but with your business knowledge, on WhatsApp, without paying middleware fees, that is what we built. ## Common mistakes (and how to dodge them) ### Meta's business verification (only when you actually need it) To start and reply to customers, **you do not need to verify your business with Meta**. You only need it for advanced features (in-chat lead forms, high-volume marketing sends). When the time comes, most rejections happen because of the wrong document or a bad scan. Upload a **clean, high-res PDF**, and make sure the legal name on it matches exactly what you put in Meta Business. If they keep rejecting, open the Meta Business support chat and ask for the specific reason. ### Loading messy content and blaming the bot This is the **most common mistake**, by far. You upload five outdated PDFs, a half-organized spreadsheet and a website URL that is also out of date, and then complain the bot «answers vaguely». The bot is **only as good as the information you feed it**. If your current pricing lives only in your sales rep's spreadsheet, upload that spreadsheet. If your catalog changes weekly, plug Shopify or PrestaShop in so the sync is automatic. **Thirty minutes** of cleaning your sources before the upload makes the difference between a useful bot and one that hedges. ### The 24-hour window nobody warns you about This is the most surprising Meta rule. You can reply freely to a customer for **24 hours** after their last message. After that window closes, if you want to message them again you have to use an approved template and Meta bills you as a marketing or utility conversation. It is by design — Meta does not want WhatsApp turning into **spam**. If your bot answers instantly (which is normal when AI is doing the work), you will rarely notice this rule, but it pays to know. ### Settling for static auto-replies The WhatsApp Business app gives you a static greeting and a static away message. That is fine as a courtesy, but it is **not customer service**, it is an **auto-receipt**. The difference between a static reply and an AI chatbot is the difference between a customer getting the answer they need to buy from you and getting nothing. If you are going to set up WhatsApp at all, set it up so it does something useful. ## Frequently asked questions ### What is a WhatsApp AI chatbot? A system that automatically replies to messages sent to your WhatsApp Business number. It can be basic (a **static text**) or it can use **artificial intelligence** to understand what the customer is asking and answer with information from your business — pricing, hours, products, returns. The gap between the two is **huge**. ### How do I build a WhatsApp chatbot without code? Sign up on a purpose-built platform, connect your WhatsApp Business number through Meta's direct flow (in Bravos AI, a **60-second popup** where you log in with Facebook and pick the number), upload your business information, and from that moment the bot answers on its own. There is nothing to install on your phone. ### Is there a free WhatsApp AI chatbot? Only if you settle for a static greeting and an away message — the two free features in the official WhatsApp Business app. A real AI chatbot, with answers grounded on your business data, has a monthly cost starting around **$23/month** for the platform, plus whatever Meta charges per conversation (which is **$0 if you only reply** to customers, not initiate). Watch out for platforms that advertise as «free» — usually a 50-message cap, a trial that bills after 7-14 days, or locked-out AI features. ### Can I use ChatGPT on WhatsApp for my business? You can plug ChatGPT into a WhatsApp number, but for a business it falls short. ChatGPT **does not know your prices**, your stock, your hours or your terms. You want the chat quality of an LLM **plus** grounded answers from your own business data — that is what a purpose-built WhatsApp AI chatbot gives you and a vanilla ChatGPT does not. ### How long does it take to go live? In Bravos AI it is **5-10 minutes** because the Meta integration is direct, no Twilio or middleware. Platforms that route through middleware usually add 1-3 days of extra setup. If you also need Meta business verification (only for advanced features like marketing sends or in-chat lead forms), add another 2-7 business days. Most small businesses can be answering messages **the same day** they sign up. ## Bottom line Building a WhatsApp AI chatbot in 2026 is not the multi-week engineering project it used to be. With a purpose-built platform that integrates directly with Meta, you can be live in **10 minutes** at a predictable monthly cost — **$23/month** with Bravos AI if you are only answering inbound messages, which is the case for most small and mid-size businesses. The other two paths (assembling it yourself with n8n, Make or Zapier; or coding straight against Meta's API) still exist and still make sense if your team is technical and has the bandwidth. For everyone else, they cost more, take longer and break more often. What actually matters when you pick: that the bot answers with the **real information from your business** (not invented), that the Meta connection is **direct** (no Twilio fees), and that the **same agent** works on your website and on WhatsApp with the same knowledge. If that fits your case, the next step is trying it on your own business. ### Ready to let WhatsApp answer on its own? Create your bot in Bravos AI, connect it to WhatsApp Business from the Starter plan, and let AI answer with the real information from your business. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # 5 ways to keep your chatbot up to date without Zapier or n8n: how we built them at Bravos AI **Category:** Practical guide | **Read time:** 15 min | **Date:** 15 May 2026 | **URL:** https://bravos-ai.com/en/blog/chatbot-data-sync A chatbot is exactly **as good as the data sitting behind it**. If prices changed last week, if a SKU went out of stock yesterday, if a product was discontinued three days ago — that update has to reach the bot without anyone having to remember it. If it doesn't, the bot keeps promising things that no longer exist. Most chatbot platforms solve this halfway: they let you upload a file or paste a URL and, for everything else, they push you to wire up the plumbing yourself with Zapier, n8n or Make. The typical result is **a chatbot advertised at $30/month that ends up costing a lot more** once you add the middleware subscription, the extra task tiers and the maintenance. This article takes that apart. There are many possible ways to keep a chatbot's data current without Zapier or n8n, and what separates one vendor from the next is **how many they ship inside the product and how many they push you to bolt on from outside**. At [Bravos AI](https://bravos-ai.com/en) we've built four of them running in production so far — the ones that cover the vast majority of practical use cases — and we walk through exactly how they work so you can compare them with whatever platform you're evaluating. ## The real cost calculation nobody runs before picking a chatbot Picture a store with 500 products. Prices rotate with weekly promotions, stock moves daily, a few SKUs get retired every month. For a chatbot to answer with current data, you need to sync that catalog at least twice a day. If your chatbot platform doesn't ship native sync, the usual workaround is to wire the connection outside with an automation tool like Zapier or n8n. They're external platforms that trigger a workflow every time something changes in one tool and replicate it in another: they spot a new row in a spreadsheet, push that data to your chatbot, and wait for the next change. Sounds neat on paper. The problem shows up on the invoice. And that's the calculation almost nobody runs when comparing chatbot prices. Let's run the numbers on Zapier, the most widely used option. Its free plan allows **100 tasks per month**, two-step Zaps max ([zapier.com/pricing](https://zapier.com/pricing)); a task is each individual action a Zap executes. Syncing 500 products twice a day means **1,000 tasks per month** just to move the catalog — before counting leads, errors or retries. You're kicked out of the free plan **by day two of the month**. The next step is Zapier's Professional plan, which starts at **$30/month (or $19.99 billed annually) for 750 tasks**. Need more? Professional scales internally with a task-tier selector: more tasks, higher monthly cost. In other words, the 1,000 tasks from our example already push you past the entry tier. A chatbot advertised at $30/month ends up with **a combined bill that nobody includes in their public price comparison**, because the math depends on your volume of changes and the exact tier each company picks. The right question before picking a platform isn't **how much does the chatbot plan cost**, it's **how much does it cost to actually keep it current**. What lives inside the product is paid for. What needs Zapier in between is paid for too. ## Why a chatbot with stale data is worse than no chatbot A chatbot without fresh data doesn't stay neutral — **it starts doing damage**. One customer asks about a product that's no longer sold and the bot eagerly recommends it at an expired price. Another asks if a specific size is in stock and the bot confirms it is, when in reality it sold out a week ago. A third looks for a color the bot still offers, even though it's been two months since the supplier last shipped it. These aren't hypotheticals. Air Canada lost a 2024 lawsuit because its chatbot promised a bereavement-travel discount that wasn't valid; the court ruled that the airline was liable for what its own chatbot said, not the customer who believed it. We cover this in detail in [chatbot legal liability](https://bravos-ai.com/blog/chatbot-responsabilidad-legal). The uncomfortable part is that most companies don't find out. **The customer doesn't write to complain — they just don't come back.** The company thinks it has a working automated tool and meanwhile the bot has been recommending a discontinued product for three months. That's why sync isn't a technical detail, it's a decision that hits the P&L directly. There are three reasons a chatbot serves old information. First: the knowledge base was loaded once as a static file and nobody touched it again. Second: sync existed but quietly broke — the classic Zapier chain that dies without notifying anyone. Third: the bot answers from its trained model instead of querying the live data at answer time. All three are solvable, and all three depend on the plumbing the platform ships. We unpack this in [why your chatbot makes up answers](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). ## The 5 native paths we ship today at Bravos AI There's no one-size-fits-all here. Every vendor decides what they solve inside the product and what they leave to you; **most fall short and push the customer toward Zapier or n8n** the moment data changes with any frequency. These are the five paths we've built into Bravos AI so that external plumbing isn't needed, with their real-world usage notes. Which one fits depends on where your data lives today. ### Path 1 — Google Sheets or Excel **The most universal path.** Every company has a spreadsheet with something in it: product catalog, prices, service rates, branch-by-branch listings, basic inventory, B2B contact list. Connecting that spreadsheet to the chatbot solves the majority of practical use cases without writing a line of code. How it works at Bravos AI: you paste the public URL of the sheet into the dashboard, choose how often you want the bot to check for changes (daily or weekly), and the system handles the rest. **Edit a cell today and the bot reflects it at the next sync.** Don't touch anything and nothing happens — the bot stays as it was. Where it fits: a dental clinic with three locations and different price lists, a restaurant with seasonal pricing, a training academy with a course catalog, a small store with fewer than 200 SKUs, any business whose product or service catalog changes regularly. Maintenance boils down to editing a cell; the bot finds out the next day without anyone having to ping it. Where it doesn't fit: catalogs with thousands of SKUs that change daily (use your e-commerce platform's native integration if one exists), data that must be live in real time (use a webhook), or sensitive information that shouldn't live in a shared spreadsheet. ### Path 2 — Shopify (real-time sync) If your store runs on Shopify, you don't need spreadsheets or files: the store itself is the source of truth, and the native integration brings it across. Bravos AI loads your entire catalog upfront (products, variants, prices, stock, descriptions) and from then on, Shopify notifies us every time something changes. **In seconds, the bot knows about the change.** The difference compared to periodic sync matters for stores with a moving catalog. **A product that goes on sale at 9:00 is reflected in the chatbot at 9:00, not the next day.** If a variant runs out of stock, the bot stops recommending it that same morning. This matters especially for Black Friday campaigns, limited drops and flash promotions, where a 24-hour delay equals a missed offer. The integration also covers the small details most platforms skip: translated product fields when the store uses Shopify Markets, optional attributes, automatic grouping by option (size, color, material) so the bot understands variants instead of treating them as separate products. Full breakdown in [Shopify AI chatbot guide](https://bravos-ai.com/blog/chatbot-shopify). ### Path 3 — PrestaShop For stores built on PrestaShop the mechanics are similar but the cadence is different: full initial catalog load and, from then on, **a daily check that updates only what changed since the last run**. No manual upkeep. Why it's not real-time like Shopify: PrestaShop doesn't emit webhooks as a standard, so the integration runs in pull mode (Bravos asks every night) rather than push (PrestaShop notifying instantly). In practice, daily cadence usually suffices for PrestaShop stores because catalog rotation tends to be less aggressive than on Shopify. The integration covers products, categories, variants, prices, descriptions and translated URLs for multi-language stores. No theme code changes, no marketplace modules to install, no Zapier in the middle. Full breakdown in [PrestaShop AI chatbot guide](https://bravos-ai.com/blog/chatbot-prestashop). ### Path 4 — WooCommerce (real time + nightly check) If your store runs on WooCommerce, the mechanics combine the best of the two above: a full initial catalog load (products, variations with their own price and stock, descriptions already cleaned of the leftover codes page builders like Elementor leave behind) and, from then on, **your store notifies us whenever something changes and the bot reflects it within seconds**. With a safety net that isn't optional in WordPress: WooCommerce notifications depend on an internal mechanism that can fail on low-traffic stores, so **the full catalog is checked against the store every night** in case any notification got lost. Changes arrive instantly when your store sends them, and the nightly check guarantees nothing stays behind for more than a day. If the API key you generate is read-only, the integration runs in nightly-only mode and the dashboard tells you so. The integration understands WooCommerce's quirks: the three ways it expresses stock (exact quantities, in stock / out of stock, inherited from the parent product), sales quoted with the before-and-after price, prices with or without tax depending on how your store displays them, and multilingual stores running WPML or Polylang. Full breakdown in our [WooCommerce chatbot guide](https://bravos-ai.com/blog/chatbot-woocommerce). ### Path 5 — Custom webhook (when your data lives somewhere else) The fifth path is for anything that isn't a spreadsheet or a Shopify, WooCommerce or PrestaShop store. **Data living in a homegrown ERP, a corporate PIM, a custom database, an industry-specific tool of record.** Plugging directly into the source isn't an option because every source is different. The right pattern is the source pushing changes to a generic endpoint whenever they happen. That's exactly what Bravos AI's custom webhook does. Your system pushes changes in JSON or XML to a unique signed URL of your bot, and the bot processes them instantly. The initial bulk load handles up to 50 MB. Each subsequent call can carry an insert, an update or a delete. And the data doesn't have to be flat: if your object has sub-lists (say, a real estate listing with several photos, or a service with several pricing tiers), the bot also knows how to use them in its answers. Real use cases are the ones that don't fit any standard template. A real estate firm that doesn't use Resales Online and has its own database. A travel agency with an internal fare engine. A manufacturer with a white-label PIM. A professional services company with a bespoke CRM. They all share one trait: nobody's going to build a dedicated connector for them, so they need a generic, well-documented path. Full reference, with payload examples, HMAC signatures, accepted formats and limits, lives on the [developer documentation page](https://bravos-ai.com/developers). For companies without internal technical staff, the person maintaining the ERP or an external vendor usually handles the connection — **once it's wired up, it runs for years untouched**. ## The external plumbing trap: Zapier, n8n and friends Zapier and n8n are excellent tools for plenty of things. The problem isn't the tools — it's using them for something the chatbot platform should handle internally. When the sync between your data and your bot routes through an external service, three things happen that are worth seeing in numbers. ### The cost compounds With our 500-product store syncing twice a day, that's 1,000 tasks per month just for the catalog. Add a flow that pushes leads captured by the bot into your CRM and you're another 500-1,000 tasks deep. Errors and retries count too. Zapier's Free plan is out (100 tasks/month). Professional starts at $30/month (or $19.99 billed annually) for 750 tasks and scales internally: more tasks, higher monthly cost. The chatbot advertised at $30/month stops being $30/month and grows with your change volume. n8n isn't hugely different. The n8n Cloud Starter plan is €20/month (around $24) for 2,500 executions, and Pro is €50/month (around $54) for 10,000, per their [official pricing page](https://n8n.io/pricing). An execution covers multiple steps, so the per-unit cost is lower than Zapier, but the underlying logic is the same: you're paying every time something moves between your data and your bot. ### And it's slow to notice Another number rarely scrutinized on the pricing page changes the whole story: polling cadence. On Zapier, the Free plan checks sources **every 15 minutes**. The Professional plan drops to 2 minutes, Team or Enterprise to 1 minute. In other words, on Zapier Free, a change you make in your sheet at 10:00 can take until 10:15 to reach the chatbot, even though the technical task itself takes half a second. Compared to the native Shopify integration at Bravos AI, where the store itself notifies us when something changes and the bot updates in seconds, those are two different worlds. The gap matters especially when we're talking about stock, flash offers and limited-time promotions: a 15-minute delay can be the difference between recommending an available product and recommending one that just sold out. ### The chain breaks silently Zapier disables Zaps that repeatedly fail and emails you about it, but that email lands in an inbox nobody monitors. n8n behaves similarly: the workflow stalls and the notification depends on whether the error monitor was configured, which most of the time it wasn't. **One of the most common stories we hear from companies switching to Bravos AI after some time on another platform: the chatbot had been quoting old prices for months and nobody noticed because the middleware had quietly died.** ### Maintenance accumulates Every change to the source format means going back to the automation. Add a new column to your spreadsheet — you have to reconfigure the flow. Shopify tweaks a field format — same drill. It's maintenance debt that builds up over time, never appears in the initial estimate, and tends to land with whoever is least prepared to handle it. ### The self-hosted n8n case n8n has one variant Zapier and Make don't: a free, self-hosted Community Edition under the FairCode license. For technical teams it can look attractive because the per-execution cost disappears. What doesn't disappear is the real cost: a server to maintain ($5-20/month depending on capacity), updates to apply when they're released, backups, monitoring, security patches and someone to take care of all that. **If the company doesn't have that someone, the initial saving translates into technical debt.** ### When external plumbing does make sense Two honest cases. One, when the source is so unusual or so rarely used that no platform will give it native support, and there's no headroom for a custom webhook integration. Two, when changes happen very infrequently (once a month or less) and the per-task cost stays inside the free tier. For everything else, **having sync inside the product is cheaper, more reliable and more maintainable**. Zapier and n8n prices cited here are publicly available on their official sites. Real numbers depend on each company's actual change volume, so it's worth running the math with your real catalog and frequency before picking a platform. ## How to pick: a four-question decision tree If, after the five paths above, you still aren't sure which one is yours, work through these four questions in order. The answer to the last one points you to the path. - **Do you run an online store on Shopify, WooCommerce or PrestaShop?** If yes, the native integration covers your case. If no, keep going. - **Does your data live in a spreadsheet today (Google Sheets, Excel) or could it reasonably fit in one?** If yes, Sheets sync. If no, keep going. - **Do you have someone on your team, your agency or an external vendor who can configure a signed webhook (a developer, a sysadmin, someone who maintains your ERP)?** If yes, custom webhook. If no, keep going. - **Are you willing to add Zapier, n8n or a similar tool in the middle, with its subscription cost or server-and-maintenance cost?** If yes, that's your path. If no, the sensible move is to prepare a spreadsheet with your essential data and start with Sheets sync. The fourth question is the one most often skipped. A well-organized spreadsheet with the 100-200 most critical data points of your business (top products, prices, variants, references and availability) **typically covers 80% of real customer queries**. You don't need to dump your entire ERP to start getting value. ## Comparison: Bravos AI vs. Tidio, Crisp and Intercom All four platforms exist and all four advertise AI chatbots. What changes is how much each one ships natively to keep data current, and starting on which plan. Comparison built from each vendor's public documentation, verified in May 2026. | Mechanism | Bravos AI | Tidio | Crisp | Intercom | | --- | --- | --- | --- | --- | | Native Google Sheets / Excel sync | Yes, from Starter (€19/month) | Via Zapier | Via Zapier | Via external tool | | Shopify integration (official app) | Yes, real-time catalog sync | Yes, native actions from Growth plan ($59/month) | Yes, official app on Shopify App Store | Yes, from Essential plan ($39/seat/month) | | PrestaShop catalog sync to chatbot | Yes, incremental daily sync | Chat widget plugin only | Chat widget plugin only | Chat widget plugin only | | Documented inbound webhook for arbitrary data | Yes, HMAC-SHA256 + JSON/XML | Not documented | Not documented | Not documented | | Re-sync that preserves bot customizations | Yes | No (re-sync wipes attached Q&A) | N/A | Yes (Fin resyncs every 24h) | | Lowest plan with automatic bot sync | €19/month | Plus ($749/month) for URL auto-sync | Not documented publicly | Essential $39/seat + Fin from $0.99/resolution | A few caveats that don't fit in the table and are worth knowing. **Tidio Lyro: re-syncing wipes your manual Q&A.** Their own documentation confirms it: when you re-sync a data source, the attached question-and-answer pairs are deleted and replaced with the new ones. If your team spent time correcting specific bot answers, those corrections vanish every time you re-sync. **All three vendors have PrestaShop modules, but only for the chat widget.** Tidio, Crisp and Intercom all have modules in the PrestaShop Addons Marketplace, but per public documentation they exist to install the chat widget on the store; we didn't find automatic catalog sync from PrestaShop into the chatbot in any of them. At Bravos AI, catalog sync is the whole point of the integration: the bot knows products, prices, stock and variants. **Intercom Fin syncs external sources, but not Google Sheets.** Fin accepts Confluence, Notion and Guru as external content sources with a 24-hour resync. To get a spreadsheet into its knowledge base, you have to route it through a third-party tool. Prices change, so double-check each vendor's pricing page before deciding. Verification links at the end of the article. The comparison focuses on data sync; each platform has strengths in other dimensions (unified inbox, surveys, marketing) that fall outside the scope of this piece. ## In short A chatbot is only as good as its data. The difference between platforms isn't the price of the plan or how slick the dashboard looks — **it's how many sync paths they ship inside and how many they push you to wire outside**. And when you wire it outside with Zapier, n8n or anything similar, that middleware has three known problems: a cost that scales with volume, delays of up to a quarter of an hour before changes reach the bot, and silent failures that leave your chatbot quoting old prices for months without anyone noticing. At Bravos AI we've built five native paths so far that cover the vast majority of practical use cases: Google Sheets / Excel sync, real-time Shopify integration, real-time WooCommerce integration with a nightly check, daily-sync PrestaShop integration, and a fully documented custom webhook for anything that doesn't fit the above. **All five ship from the Starter plan at $20/month**, with no tasks to count and no chains breaking in silence. There is one more way to keep it current, this one from the other side. If you work with Claude or ChatGPT, with [our MCP connector](https://bravos-ai.com/conector-ia-mcp) you ask the assistant itself to go through the loaded content, spot what has fallen out of date and fix it in the same conversation. It doesn't sync an external source like the five above: it helps you audit and repair what the bot already has, without opening the panel. The suggestion is the same one that opens the article: before comparing chatbot prices, **open a spreadsheet and add up what it actually costs to keep the bot current**. That's the cost that matters. ## Frequently asked questions ### How often should I update my chatbot's knowledge base? It depends on the type of information. Catalog data (prices, stock, availability, references) changes daily and is what needs automatic sync — daily or real-time. The rest of the bot's content (service descriptions, informational text, hand-edited FAQs) is updated by a person when something changes, no automatic sync involved. The rule of thumb we use at Bravos AI: **if a customer could make a buying decision based on a piece of data, that data shouldn't be more than 24 hours old**. ### Can I connect my chatbot to Google Sheets without Zapier? Yes, if the chatbot platform supports it natively. At Bravos AI you connect a Google Sheet or Excel file directly from the dashboard: paste the public URL, pick the sync frequency (daily or weekly), and the bot updates itself whenever you change something in the sheet. No middleware, no per-task billing, no broken chain when Zapier goes down. And if you use Claude or ChatGPT, you can even review and update the bot's content straight from the chat with [an MCP](https://bravos-ai.com/blog/que-es-un-mcp). ### Why does my chatbot quote outdated prices even after I changed the catalog? Usually three reasons. One, you uploaded the catalog as a static file (PDF, manual CSV) and there's no sync. Two, sync exists but is silently broken — the classic Zapier or n8n chain that dies without notifying anyone. Three, the chatbot reads from the trained model instead of querying the live data at answer time. We unpack this in [why your chatbot makes up answers](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). ### How much does it really cost to keep a chatbot updated with Zapier or n8n? Zapier Free caps at 100 tasks/month, two steps max per Zap. Professional starts at $30/month (or $19.99 billed annually) for 750 tasks and scales internally: passing the 750-task ceiling forces you up a tier within Professional, with the matching monthly cost. n8n Cloud Starter is €20/month (about $24) for 2,500 executions; Pro is €50/month (about $54) for 10,000. Self-hosted Community is free under FairCode, but adds server and maintenance cost. In every case, it's an invoice separate from your chatbot, and worth pricing out before picking a platform. ### What if my data isn't in Sheets or in Shopify, WooCommerce or PrestaShop? That's what the custom webhook is for. Your system (ERP, PIM, in-house database, custom tool) pushes changes to a unique URL of your bot in Bravos AI, signed with HMAC-SHA256 and in JSON or XML. It supports inserts, updates, deletes, multiple datasets per bot and objects with sub-lists (for example, photos of a real estate listing or pricing modalities of a service). The initial bulk load handles up to 50 MB. Full reference on the [developer documentation page](https://bravos-ai.com/developers). ## Verification and sources Data and quotes verified in May 2026. Pricing pages and vendor documentation change frequently: check the official sources before making decisions. - **Zapier** — Plans and tasks per tier: [zapier.com/pricing](https://zapier.com/pricing). - **n8n** — Cloud pricing and Community self-hosted notes: [n8n.io/pricing](https://n8n.io/pricing). - **Tidio Lyro** — Data source documentation: [help.tidio.com](https://help.tidio.com). - **Crisp** — Pricing and AI credits: [crisp.chat/en/pricing](https://crisp.chat/en/pricing/). - **Intercom Fin** — External source sync: [intercom.com/help](https://www.intercom.com/help). - **Bravos AI** — Developer documentation (custom webhook): [bravos-ai.com/developers](https://bravos-ai.com/developers). ### All 5 paths, inside the product, from €19/month Synced spreadsheet, real-time Shopify, real-time WooCommerce, native PrestaShop and custom webhook — all five ship with Bravos AI's Starter plan. No Zapier or n8n in between, no tasks to count, no chains breaking in silence. Try it with 7 days free on the full PRO plan. [Try PRO free for 7 days](https://app.bravos-ai.com/register?lang=en) --- # We asked Tidio, Crisp and Freshchat\ **Category:** Research | **Read time:** 8 min | **Date:** May 12, 2026 | **URL:** https://bravos-ai.com/en/blog/tidio-crisp-freshchat-free-plan The most popular AI chatbot platforms market themselves the same way: **chatbot with artificial intelligence**. Some of them also offer a permanent free plan. The natural question for anyone evaluating options is straightforward: is that AI actually included in the free plan, or is it gated behind a paid tier? Instead of signing up for each platform to find out, we asked the official chatbots they themselves run on their own websites. Those chatbots are their premium products, running in real conditions. If they admit how the free plan works, that's the most authoritative source you'll get. Here's what Lyro (Tidio), Hugo (Crisp) and the Freshchat pricing page had to say, and what happened when we tried to repeat the experiment with the Bravos AI assistant. To keep the experiment symmetrical: in all three cases we asked the premium chatbot of each platform about its own free plan. The fourth case is our own — [Bravos AI](https://bravos-ai.com). ## Tidio: 50 AI conversations «forever» — but they don't renew Tidio has a permanent free plan. Their pricing page highlights that this plan includes **50 Lyro AI Agent conversations**. What isn't highlighted, and what most visitors miss, is that those 50 conversations are **a one-time lifetime total**, not monthly. Once you burn through them, they don't renew. To confirm this, we asked Lyro — Tidio's official assistant on their own website — directly. The answer was unambiguous: The 50 free Lyro AI conversations are a one-time lifetime total, not a monthly allowance. This means that on the Free, Starter, and Growth plans, you get 50 Lyro AI conversations in total, and they do not reset or renew each month. And on how to get a renewable monthly quota, Lyro made it explicit: To get a renewable monthly quota of Lyro conversations, you'd need the Lyro AI plan, which offers between 50 and 1,000 conversations per month (resetting every 30 days). In short: Tidio's free plan gives you 50 AI conversations. Once they're gone, they're gone. To get a renewable monthly allowance you have to pay for the Lyro AI plan separately. ## Crisp: the free plan does not include AI. And their own chatbot confirms it. ### What the pricing table shows Crisp has four plans: Free (€0), Mini (€45/month), Essentials (€95/month) and Plus (€295/month). On the «Compare our plans» table there's a row called **«AI credits»** showing how much each plan includes: **€0 on Free**, €5 on Mini, €25 on Essentials and €75 on Plus. On the free plan, AI credits are zero. Not a single euro. ### In English, Hugo doesn't answer To understand what this actually means in practice, we asked Hugo — Crisp's official assistant on their own website — in English. Here's what came back: Sorry, I don't know how to answer that question. You have been transferred to our support team. Then it asked for an email so someone could reply later. Crisp's official assistant, in English, couldn't answer a question about its own pricing model and escalated to human support. ### But in Spanish, Hugo answers in detail Out of curiosity, we asked the exact same question in Spanish. The response was the opposite — Hugo answered with three concrete data points that aren't obvious from the pricing page (translated from Spanish): The Free plan doesn't include AI credits, so Hugo (our AI agent) isn't available. You can only use human agents. AI credits are token-based. The average cost of a conversation handled by Hugo is around $0.05 to $0.10. With the $5 included in Mini, that's roughly 50-100 automated conversations per month (the pricing page estimates around 90 conversations). The documentation doesn't specify exactly what happens when they run out. Three numbers the homepage doesn't show: on Free, zero AI; on Mini, the €5 buys around 50-100 automated conversations per month; and no official clarity on what happens when credits run out mid-month. The same question, in a different language, returns the entire answer — which is what happens when a chatbot's knowledge base only covers some of the languages it's supposed to support. ## Freshchat: AI starts at the paid tier. And there's no chatbot on their own website. Freshchat also offers a permanent free plan: €0/month for up to 10 agents. But Freshchat's AI, branded Freddy, **is not included on the free plan**. On the pricing table, under the «Boost your results with Freddy AI» section, AI is sold as two separate products: - **Freddy AI Agent**: only available from the Growth plan upwards (€21/agent/month on monthly billing). The first 500 sessions are included; after that, €45 per additional 100 sessions. - **Freddy AI Copilot**: only on the Pro and Enterprise plans. Even there, it costs an extra €35/agent/month. We wanted to ask Freshchat's own official chatbot the same question, the way we did with Lyro and Hugo. But **Freshchat doesn't have a chatbot installed on its own website**. To get in touch, you have to fill out a form and wait for an email reply, or schedule a demo. The platform that sells AI chatbots doesn't use AI chatbots for its own customer support. ## Bravos AI: we decided not to play the "free plan with fine print" game Transparency note: we are [Bravos AI](https://bravos-ai.com/en). When we first published this article, we did have a permanent free plan (200 monthly messages, renewable, with GPT-4o-mini). It worked well as a hook, but after months of auditing how other platforms package theirs — and seeing what you've just read — we realized we were playing the same game. 200 monthly messages sound like a lot until 50 visitors land over a weekend and the bot stops mid-conversation. The visitor has no idea what happened. You don't find out until next month. And the promise of "real AI, free" turns into "real AI for the first 3 days of the month". So we changed the model. We removed the permanent free plan and replaced it with **7 days free on the full PRO plan**: 3 chatbots, unlimited messages, all integrations (Shopify, PrestaShop, Resales Online, WhatsApp), lead capture, and the most advanced AI model. No asterisks. After 7 days, if it's worth it, plans start at €19/month (Starter, also with unlimited messages); if it isn't, you cancel before the first charge and pay nothing. It's less "attractive" on a pricing page than a "Free Forever" badge. But it's honest: for 7 days you get the real, full version — not a stripped-down one that breaks on day 7 of the month. And from day 8 onwards, you know exactly what you pay. And there is one thing none of the three platforms in this article offer: managing the bot from your own Claude or ChatGPT. With [our MCP connector](https://bravos-ai.com/conector-ia-mcp) you review the conversations that are going unanswered, fix the content and adjust the tone by asking the assistant you already use, without opening the panel. Neither Tidio, nor Crisp, nor Freshchat lets you touch the bot outside their own tool. The full pricing comparison (ours and the rest of the market) is in our [AI chatbot pricing guide for 2026](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa). ## What anyone who actually asks learns When a platform says «free AI chatbot» on its homepage, the detail of what the free plan actually includes isn't always where the visitor looks first. It's there, but it's in comparison tables, footnotes or FAQs. Sometimes you have to ask the platform's own chatbot directly to get a concrete number. Two questions worth answering before picking a platform: - Is the free-plan AI renewable each month, or does it get burned once and that's it? - Is the AI included in the plan, or is it sold as a paid add-on? This experiment can be repeated with any other platform that runs its own chatbot on its website. The answers the bot gives are often clearer and more specific than the pricing page itself. ## Frequently asked questions ### Does Tidio's free plan include AI? Yes, but only 50 Lyro AI conversations **as a one-time lifetime total**. They are not monthly and they don't renew. Once they're gone, you need to pay for the separate Lyro AI plan to get a renewable monthly quota, which ranges from 50 to 1,000 conversations per month depending on the tier. ### Does Crisp's free plan include AI? No. Crisp's free plan has €0 in AI credits, so Hugo (their AI agent) isn't available on that tier. Only human agents can be used. AI starts on the Mini plan (€45/month) with €5 in credits, equivalent to roughly 50-100 automated conversations per month according to Crisp's own official chatbot. ### Does Freshchat's free plan include AI? No. Freshchat's AI is sold as two separate products (Freddy AI Agent and Freddy AI Copilot), both available only from the Growth plan (€21/agent/month) and Pro plan respectively, with extra charges per additional session. The free plan only includes up to 10 human agents with no AI. ### Which AI chatbot platform offers real AI on a renewable monthly free plan? After auditing the main platforms, the honest answer is that none of them offer a useful free plan with real AI: it's either spent once and gone (Tidio) or not included at all (Crisp, Freshchat). That's why at Bravos AI we chose a different model: instead of a free plan with fine print, we offer 7 days free of the full PRO plan with unlimited messages, AI included and all integrations. After 7 days you decide whether it's worth paying (plans start at €19/month). ### How can I check if a chatbot's free plan includes AI before signing up? The most direct way is to ask the platform's own official chatbot installed on their website (Lyro on Tidio, Hugo on Crisp). It's their premium product running live, and it usually responds with concrete data about the free plan. If the platform doesn't have an official chatbot on its site (like Freshchat), the detailed pricing table usually has the answer in a row called «AI credits», «AI usage» or similar. ## Notes - Intercom, Zendesk, Drift and Ada are not in this comparison because they don't offer a permanent free plan — only limited trials or enterprise demos. - HubSpot and Landbot also lock their AI behind paid plans, although they do have a basic free tier. - Pricing verified in May 2026. Platforms change their conditions frequently; check the official pricing page before making decisions. ## Verification: how to reproduce this experiment Every number and quote in this article was checked in May 2026. Any reader can reproduce the experiment by asking the official chatbots directly and checking the pricing pages: - **Tidio** — Pricing: [www.tidio.com/pricing](https://www.tidio.com/pricing). Lyro chatbot at [www.tidio.com](https://www.tidio.com). - **Crisp** — Pricing: [crisp.chat/en/pricing](https://crisp.chat/en/pricing/). Hugo chatbot at [crisp.chat](https://crisp.chat). - **Freshchat** — Pricing: [freshworks.com/live-chat-software/pricing](https://www.freshworks.com/live-chat-software/pricing/). No official chatbot on their own site as of publication. - **Bravos AI** — Pricing: [bravos-ai.com/en/pricing](https://bravos-ai.com/en/pricing). Official chatbot at [bravos-ai.com](https://bravos-ai.com). ### Want to try one without the fine print? 7 days free on the full Bravos AI PRO plan: 3 bots, unlimited messages, all integrations. Cancel before the first charge and pay nothing. [Try PRO free for 7 days](https://app.bravos-ai.com/register?lang=en) --- # Dental Chatbot: AI Receptionist for Treatments, Pricing & Insurance in 13 Languages **Category:** Use case | **Read time:** 12 min | **Date:** Apr 30, 2026 | **URL:** https://bravos-ai.com/en/blog/dental-clinic-chatbot At [Bravos AI](https://bravos-ai.com/en) we've built the dental chatbot we wished existed when we tried to book a Sunday-night cleaning: a virtual assistant that answers questions about treatments, pricing, accepted insurance and opening hours using your practice's real information — and hands the conversation to a human when it should. Most chatbots for dental practices are either rigid online-booking widgets that only let you pick a time slot, or generic AI that makes up prices on the fly. The first one doesn't answer questions. The second one is dangerous: a wrong quote either turns away a patient or makes them show up angry. Bravos uses a technique called RAG that searches your documents before answering — we explain it in [this article on why chatbots hallucinate](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). This guide covers what an AI dental chatbot can do for your practice, what it shouldn't do (especially with health data under GDPR and HIPAA), how it compares to Tidio, Intercom and online booking platforms like Doctolib or Zocdoc, and how much it actually costs. ## The problem: a phone that no one picks up after hours The [American Dental Association](https://www.ada.org/resources/research/health-policy-institute/dentist-workforce) counts 202,485 working dentists in the US in 2024 — 59.5 per 100,000 people. Add to that around 45,000 registered dentists in the UK and tens of thousands more across Europe and Australia, and you get an extremely competitive market where most practices win or lose patients on the same battleground: how the very first inquiry is handled. And that's where money leaks. The phone only answers during office hours, patients ask on WhatsApp at night, social media DMs pile up unread, and contact forms get a reply "tomorrow morning, when reception is free." That's why more practices are looking at AI receptionists and dental virtual assistants to automate the first conversation — including the [WhatsApp channel](https://bravos-ai.com/blog/chatbot-whatsapp), where most after-hours inquiries land. A widely cited study published in [Harvard Business Review](https://hbr.org/2011/03/the-short-life-of-online-sales-leads) found that **responding within 5 minutes makes you 21x more likely to convert an inquiry**. After 30 minutes, the probability drops by 90%. If your prospective patient is comparing you against the practice three blocks away, they're not waiting long. Then there's the chronic problem of the sector: missed appointments. Studies summarised in [SciELO](http://www.scielo.org.pe/scielo.php?script=sci_arttext&pid=S1019-43552024000100063) place dental no-show rates between 5% and 25% in developed countries, and [Hellomatik (2024)](https://hellomatik.com/es/news/que-dice-realmente-tu-tasa-de-inasistencias-sobre-tu-clinica) reports averages of **18% to 22%** — the highest among medical specialties. SMS or WhatsApp reminders cut that by 30% to 50%, but only if you have the patient's contact in the first place. A cold web form rarely gets it. ## What an AI dental chatbot does in a real practice ### Answers about treatments, pricing and accepted insurance Upload your treatment list (PDF, text or Excel), the insurance plans you accept and your financing terms. The chatbot can answer "how much is a root canal?", "do you accept Delta Dental / Cigna / Bupa?" or "do you offer 12-month interest-free financing?" using the information you gave it. If a question isn't covered by your documents, it says so and tells the patient to call — it doesn't invent a price. ### Captures appointment requests with full context When a patient asks about an implant or wants an orthodontics quote, the chatbot collects their name, phone number and preferred time slot. The whole conversation lands in your inbox: what they asked, which treatment they care about, whether they have insurance. Reception calls back with that context already on screen instead of starting from zero. We dig into this in [how to use a chatbot for lead capture without being pushy](https://bravos-ai.com/blog/chatbot-captar-clientes). ### Handles 13 languages out of the box Tourist destinations, expat communities, and patients who simply prefer their mother tongue are everywhere. Per [CSA Research](https://csa-research.com/Blogs-Events/CSA-in-the-Media/Press-Releases/Consumers-Prefer-their-Own-Language), **76% of people prefer to buy in their own language**. A visitor types in German, Russian or French; the chatbot replies in that language using a single set of content — no manual translation. More on that in our [multilingual chatbot guide](https://bravos-ai.com/blog/chatbot-multiidioma). ### Updates instantly when prices or treatments change Add a new treatment (clear aligners, whitening, conscious sedation), change a price, replace your insurance list. Update the document in the Bravos panel and the chatbot is in sync. No developer, no support tickets, no waiting. And if you use Claude or ChatGPT, with [our MCP connector](https://bravos-ai.com/conector-ia-mcp) the assistant itself reviews what the bot has loaded and, with our team's experience inside, tells you what's missing or worth rewriting. ### Email alerts when an appointment request comes in When someone leaves their details to be called back, reception gets an email with the full conversation. They don't need to keep the dashboard open or check a tab — it shows up like any other email. ## Real things a dental virtual assistant can resolve These are concrete questions the chatbot can answer with no human intervention: - **First visit:** "How much is the first exam?" → answers whether it's free or paid, what's included, and how to book. - **Cosmetic treatments:** "How much is teeth whitening?" or "Do you do Invisalign?" → returns your prices and options. - **Implants and orthodontics:** "How many appointments does an implant take?" → uses your protocol and timeline. - **Insurance and financing:** "Do you accept Delta Dental, Aetna or Bupa?", "Any zero-interest plans?" → answers from the list you configured. - **Emergencies:** "I have severe pain, what should I do?" → tells the patient whether you offer same-day emergencies, when, and what to do next. - **The basics:** "Are you open Saturdays?", "Is there parking?" → answers in the visitor's language with your real data. These are questions reception ends up answering on repeat. When an AI receptionist handles them 24/7, the phone stops ringing for the obvious stuff and the team can focus on the patients who are actually in the chair. ## What options exist for a dental practice (and why the difference matters) If you're thinking about automating that first conversation, today there are three common paths. Each one solves a different problem: | | Online booking (Zocdoc, Doctolib, Doctoralia) | Button bot or web form | AI trained on your data (Bravos AI) | | --- | --- | --- | --- | | Answers questions in natural language | No | No | Yes | | Knows your real prices | No (booking only) | Only if you script every option | Yes, from your list | | Knows your accepted insurance | Sometimes | Manually | Yes | | Auto multilingual | Limited | No | 13 languages | | Captures contact with context | Booking-only | Cold form | Yes, in-conversation | | Says "I don't know" when it doesn't | N/A | Falls back to the tree | Yes | The third column is what Bravos AI does. The technique is called RAG (Retrieval-Augmented Generation): instead of improvising an answer, it searches your documentation and replies with what it finds. We unpack it in [this article on getting a chatbot to answer with your data](https://bravos-ai.com/blog/chatbot-catalogo-productos). The first two columns aren't competitors: they coexist with the chatbot. Online booking brings you appointments; the chatbot answers questions and captures the patients who don't feel ready to call. ## GDPR, HIPAA and health data: what you need to know before deploying This matters and many vendors gloss over it. Under [GDPR](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32016R0679) (Article 9), health data is a **special category** and demands stricter safeguards than ordinary personal data. In the US, HIPAA imposes its own constraints on protected health information (PHI) and on any vendor that handles it (covered entities and business associates). For that reason a dental chatbot **should not** be used for the patient to send their medical history, intraoral photos, X-rays, or to receive symptom advice. Conversations must stay within: - General information about the practice (hours, location, team). - Treatments, descriptions and indicative pricing. - Accepted insurance and financing options. - Basic contact data (name, phone, email) so the practice can call back. Anything clinical needs to flow through secure channels: your EHR, your encrypted practice management software, a phone call, or an in-office visit. The chatbot helps with intake and information — it's not a replacement for the clinical team. Bravos AI encrypts data in transit and at rest, runs on European infrastructure (Hetzner) and signs a Data Processing Agreement (DPA) with any practice that asks. But configuring the chatbot within legal limits and informing the patient about how their data is handled remains the practice's responsibility as the data controller. We dig into that side in [what legal liability you carry when your chatbot gets it wrong](https://bravos-ai.com/blog/chatbot-responsabilidad-legal). ## How much a dental chatbot costs in 2026 Verified pricing as of April 2026: | Platform | Base price | AI included | Notes | | --- | --- | --- | --- | | Tidio | Free (50 conv.); Starter $29/mo | Lyro AI: from $39/mo extra | Real cost with AI: $68+/mo | | Intercom (Fin) | $29/mo | $0.99 per resolved conversation | 100 conversations/month = $99 extra | | Zocdoc / Doctolib | Not published | Online booking, not AI chatbot | Subscription + per-booking commissions | | Bravos AI | 7-day free PRO trial; Starter €19/mo | AI included in every plan | Pro: €49/mo (3 bots, lead capture). No per-conversation fees | Detailed pricing comparison in [how much an AI chatbot costs in 2026](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa). ## What a dental chatbot shouldn't do **It doesn't diagnose.** If a patient describes symptoms ("I feel pain when chewing on the right side"), the chatbot acknowledges the message and asks them to book an appointment or call. It doesn't say "sounds like an endodontic infection." That's clinical advice, not chat content. **It doesn't book directly into your practice management software.** Out of the box it won't connect to Dentrix, Eaglesoft, Open Dental, Curve, Software of Excellence or your in-house calendar. It captures the request and emails it to the practice so reception can close the appointment in your system. **It doesn't carry sensitive clinical data.** It shouldn't be used for the patient to send medical history, X-rays or intraoral photos. Those flow through secure channels approved by the practice. What it does, well: answer the questions your practice fields every single day — pricing, insurance, hours, financing, emergencies — 24/7, in 13 languages, capturing contact with full context so you close the booking by phone. For most practices that means swapping a half-day on the phone for a team that's actually with patients. ## How to set up a dental chatbot with Bravos AI - **Create your account** at [app.bravos-ai.com](https://app.bravos-ai.com/register?lang=en) and start the 7-day PRO trial. - **Upload your content:** treatments and pricing, accepted insurance, financing terms, hours, location and your website URL. PDF, text, Excel or CSV. - **Set the tone** (a warm-professional voice tends to work well in healthcare) and the welcome message. - **Enable lead capture** for higher-ticket treatments: implants, orthodontics, cosmetic dentistry, sedation. - **Paste one line of code** on your site. The Bravos AI dashboard ships with step-by-step instructions for WordPress, Squarespace, Wix, Webflow, Shopify, PrestaShop and direct HTML. If you're on WordPress, we wrote a quick walkthrough in [how to add a chatbot to WordPress in 5 minutes](https://bravos-ai.com/blog/chatbot-wordpress). - **That's it.** The chatbot answers 24/7 in 13 languages and emails any appointment request to your team. ## Frequently asked questions about dental chatbots ### Is it legal to use a chatbot in a dental practice under GDPR or HIPAA? Yes, provided you respect data protection law. You must inform users about data processing as the conversation begins, never collect sensitive clinical data via chat (history, photos, X-rays), and have a Data Processing Agreement (DPA) signed with the chatbot provider. In the US, any vendor handling protected health information should sign a Business Associate Agreement under HIPAA. The practice remains the data controller for patient information. ### Can a dental chatbot give medical advice or diagnose? It shouldn't, and a properly configured chatbot won't. If a patient describes symptoms, the right behaviour is for the chatbot to acknowledge the query, recommend booking an appointment or calling the practice, and explain emergency procedures if you offer them. Diagnosing through chat is a bad idea legally and clinically: the patient's description is always incomplete and clinical responsibility stays with the licensed practitioner. ### Can a chatbot book appointments directly in Dentrix, Open Dental or my practice management software? Most AI chatbots on the market — Bravos AI included — don't connect out of the box with Dentrix, Eaglesoft, Open Dental, Curve, Software of Excellence or similar systems. What they do is capture the request (name, phone, treatment, preferred slot) and email it to the practice so reception can close the booking. If your software accepts incoming webhooks, custom integrations are possible. ### Can a chatbot replace the front-desk receptionist? No. A chatbot handles the repetitive questions (pricing, insurance, hours, first visit, emergencies) and captures qualified contacts with full context. The front desk is still essential for closing appointments in the software, attending to patients on-site and resolving anything that needs human judgement. ### How much does a dental chatbot cost? It depends on the vendor. Tidio starts at $29/month plus $39/month for the AI add-on (around $68/month combined). Intercom charges $0.99 per AI-resolved conversation, which scales fast on a busy site. Bravos AI offers a 7-day free trial of the full PRO plan, Starter at €19/month (1 bot, unlimited messages) and Pro at €49/month (3 bots, unlimited messages, lead capture), with AI included in every plan and no per-conversation fees. ### Want to see it working with your practice? Connect your chatbot to your treatment list and see how it answers. After that, plans start at €19/month. [Try PRO free for 7 days](https://app.bravos-ai.com/register?lang=en) --- # AI Agent or Chatbot: Which One Does Your Business Actually Need? **Category:** Analysis | **Read time:** 11 min | **Date:** Apr 16, 2026 | **URL:** https://bravos-ai.com/en/blog/ai-agent-or-chatbot In 2026, everything has become an "AI agent." Tools that were chatbots yesterday are now marketed as intelligent agents. But the difference between an AI chatbot and an [AI agent](https://bravos-ai.com/blog/que-es-un-agente-de-ia) is real, technical, and has direct implications for what you pay and what you get. [Gartner](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) predicts that over 40% of agentic AI projects will be canceled before 2027. The main reason: many companies are buying something they don't need — or that simply isn't what they were sold. This article breaks down what each one actually is, what it costs, and how to figure out which one your business needs. ## What is an AI agent, really? Gartner defines AI agents as **autonomous or semi-autonomous software entities** that use AI to perceive their environment, make decisions, take actions, and achieve goals. The key word is *take actions*: an agent doesn't just answer — it acts. [Oracle](https://blogs.oracle.com/developers/what-is-the-ai-agent-loop-the-core-architecture-behind-autonomous-ai-systems) puts it more concretely: what separates an agent from a chatbot is the **agent loop — a cycle of "think → act → evaluate."** An AI model decides what step to take, executes it using external tools (a CRM, a payment system, an email service), checks the result, and decides the next step. It repeats until the task is complete or it determines it can't continue. That bridge between the AI and external tools now has a standard of its own: [an MCP](https://bravos-ai.com/blog/que-es-un-mcp). A concrete example: you tell the agent "handle the complaint for customer #4521." The agent, on its own, reviews the customer's history in the CRM, checks the order status in the management system, determines that a partial refund applies based on company policy, processes it, and sends a confirmation email. All without human intervention. ### Where they work and where they don't AI agents already work in real production environments, but only in very specific contexts. [Salesforce](https://www.salesforce.com/news/stories/agentforce-customer-success-stories/), for example, has a product called Agentforce with 18,500 customers, reporting that only 4% of conversations need a human. Fisher & Paykel, an appliance manufacturer, increased their self-service rate from 40% to 70% with the tool. But general autonomy is still far off. OpenAI Operator — OpenAI's agent designed to browse the web and complete tasks — achieved only a **38.1% success rate** in its benchmark tests. In a Washington Post test, it made an unauthorized $31 purchase on an online store. As [an analysis on the Stack Overflow blog](https://stackoverflow.blog/2026/03/20/was-2025-really-the-year-of-ai-agents/) put it: "Agents work well in narrow verticals, not as a general-purpose solution." ## What is an AI chatbot? An AI chatbot uses a language model combined with your business data to answer questions naturally. The concept is straightforward: when a customer asks something, the chatbot searches your information (website, documents, product catalog) and generates an answer based on what it finds. This technique is called **RAG** (Retrieval-Augmented Generation). Here's the thing: **a chatbot is only as good as its data**. If the information is outdated, poorly structured, or incomplete, the chatbot will fail regardless of how good the AI is. The difference between a chatbot that works and one that frustrates customers usually isn't the language model — it's the quality and freshness of the data it works with. The numbers back this up: according to [ChatBot.com](https://www.chatbot.com/blog/chatbot-statistics/), AI chatbots resolve **87% of queries without human intervention**. [Freshworks](https://www.freshworks.com/How-AI-is-unlocking-ROI-in-customer-service/) puts the return at **$3.50 for every $1 invested** in customer service AI. And they deploy in days, not months. If you're curious about the difference with button-based and scripted chatbots, we have a [dedicated article on rule-based vs AI chatbots](https://bravos-ai.com/blog/chatbot-reglas-vs-chatbot-ia). ## AI chatbot vs AI agent: the real differences At [Bravos AI](https://bravos-ai.com/) we build AI chatbots, so we're not neutral here. But that's precisely why we know where a chatbot's limits are and when you actually need something more. Here's the comparison that matters: the difference between a chatbot that uses AI to answer and an agent that uses AI to act. | | AI Chatbot | AI Agent | | --- | --- | --- | | What it does | Answers questions using your data | Makes decisions and executes actions across systems | | Example | "Yes, that jacket is available in size M at €45" | Processes a refund, updates the CRM, and sends a confirmation email | | Technical team required | No | Yes — integrations, maintenance, monitoring | | Time to go live | Days | Weeks or months, depending on complexity | | Typical cost (SMB) | €200–5,000/year | Tens of thousands of €/year (implementation + usage) | | Best for | Customer support, product catalogs, lead capture | Complex multi-system operations | For a small or mid-sized business, the most relevant takeaway is the gap in cost and complexity. An AI chatbot can be set up in days without a technical team. An agent requires integrations with your systems, a team to maintain it, and a budget that often exceeds what an SMB spends on technology in an entire year. For specific platform pricing, we have a [chatbot pricing comparison for 2026](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa). ## The agent washing problem Gartner places AI agents at the **Peak of Inflated Expectations** in their 2025 Hype Cycle. In plain terms: this is the moment of maximum noise and minimum clarity about what actually works. The result is a phenomenon the industry calls **agent washing** — companies rebranding their products as "agents" without changing anything real. Scripted chatbots, call recorders, basic CRM integrations — everything is sold as an "AI agent" now. [SDxCentral](https://www.sdxcentral.com/analysis/was-2025-really-the-year-of-the-ai-agent/) summed it up: "The word *agent* replaced *copilot*, which had replaced *chatbot*." A label swap, not a technology change. The data points in the same direction: - **Over 40% of agentic AI projects will be canceled** before 2027 due to runaway costs and unclear business value (Gartner) - **80.3% of AI projects fail to deliver** expected business value ([RAND Corporation](https://www.rand.org/pubs/research_reports/RRA2680-1.html)) - **Fewer than 10% of companies** have managed to scale AI agents with tangible results ([McKinsey](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai)) How do you tell if what you're being sold is a real agent? Ask three questions: can it execute actions in external systems (not just answer)? Does it make intermediate decisions on its own? Does it operate in that "think → act → evaluate" loop we described earlier? If the answer to all three isn't a clear yes with concrete examples, you're probably looking at a chatbot with a marketing layer on top. ## When you actually need an AI agent There are cases where a chatbot — no matter how good — isn't enough. An AI agent makes sense when your process requires: - **Multi-system coordination.** The task involves reading from the CRM, writing to the management system, sending an email, and updating a ticket. That's not a conversation — it's a chain of coordinated actions - **Autonomous decisions within guardrails.** Approving a refund, classifying an insurance claim, assigning a resource — decisions that follow business rules but require evaluating context - **Complex troubleshooting.** Beyond FAQs: the agent diagnoses the problem, tries solutions, and escalates to a human with full context if it can't resolve it - **Multi-step workflows.** Processing a full complaint, managing employee onboarding, coordinating an approval chain These are real needs, typically from mid-sized or large companies with a dedicated technical team and a five-figure implementation budget. If your business fits here, a chatbot won't solve the problem. ## When an AI chatbot is the right choice For most small and mid-sized businesses, the real need is simpler than it seems: someone to answer customers well when you can't. And for that, an AI chatbot trained on your own data already solves the problem. - **Your main need is answering questions.** About products, services, pricing, hours, shipping policies. The same questions your customers ask over and over - **You want to capture leads after hours.** 97% of your website visitors leave without a trace. A chatbot can [capture leads naturally](https://bravos-ai.com/blog/chatbot-captar-clientes) when nobody is available - **You sell products with a catalog.** An AI chatbot can help your customers find what they're looking for, answer product questions, and guide them through the buying decision - **You need to serve customers in multiple languages.** With content in just one language, an AI chatbot can respond in whatever languages the platform supports. [Here's how we do it at Bravos AI](https://bravos-ai.com/blog/chatbot-multiidioma) - **You don't have a dedicated technical team.** A chatbot platform like Bravos AI, Tidio, or Intercom can be set up without writing code. An AI agent requires integrations, maintenance, and monitoring - **Your budget is hundreds, not thousands.** From €19/month you can have a working chatbot. A real agent starts in a very different range The data supports this: 64% of SMBs plan to adopt a chatbot by 2026, and 91% of those already using AI report increased revenue (Salesforce). For a deeper dive, we have a [complete guide to chatbots for small businesses](https://bravos-ai.com/blog/chatbot-para-pymes). **On product catalogs:** most chatbots rely on text-based search and [can't filter by real attributes](https://bravos-ai.com/blog/chatbot-catalogo-productos) like price, size, or stock. At Bravos AI we solve this by combining AI with structured queries, and automatically syncing the catalog with your store (Shopify, PrestaShop) so the data is always up to date. ## Five questions to help you decide Not sure which one your business needs? Answer these five questions: - **Does your process need to touch 3 or more different systems?** If yes (CRM + order management + email + payments), you're in agent territory. If you just need to answer questions based on your data, an AI chatbot is enough. - **Do you need the AI to execute actions or to answer?** Processing a payment, modifying an order, sending a follow-up email = agent. Answering about your menu, your catalog, your services = chatbot. - **Do you have a five-figure budget and a technical team?** A real AI agent requires an implementation that can take weeks or months, integrations with your systems, and a team to maintain it. If you don't have both, an agent isn't viable today. - **Do you need to be live in days or can you wait?** An AI chatbot can be configured in an afternoon. An AI agent requires an integration project that, depending on complexity, can take considerably longer. - **Is your main problem that you're not responding fast enough?** If most of your lost opportunities come from not responding — after hours, on weekends, in another language — an AI chatbot solves exactly that. You don't need an autonomous agent to answer questions. ## Frequently asked questions ### Can an AI chatbot evolve into an AI agent? Technically yes, but they're different architectures. An AI chatbot answers based on your data. To turn it into an agent, you need to add the decision-action-evaluation loop, connections to external systems, business logic, and error handling. It's not a natural evolution — it's a different engineering project. ### Will AI agents replace chatbots? Not in the short term. The chatbot market (roughly $10 billion in 2025) is still larger than the AI agent market ($7.8 billion). Agents are growing faster, but for most business use cases — customer service, sales, support — a well-trained chatbot remains the most practical and cost-effective solution. ### How much does an AI agent cost compared to a chatbot? The gap is significant. An AI chatbot for an SMB costs between €200 and €5,000 per year. A real AI agent (not a rebranded chatbot) requires a substantial upfront investment in implementation and integrations, plus per-use costs that are many times higher. For detailed chatbot pricing, we have a [cost guide updated for 2026](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa). ### How do I know if what I'm being sold is a real agent or agent washing? Ask three questions: can it execute actions in external systems (not just answer)? Does it make intermediate decisions on its own? Does it operate in a loop where it evaluates results and decides the next step? If the answer to all three isn't a clear yes with concrete examples, you're probably looking at a chatbot with a marketing layer on top. ### Need to answer your customers, not an autonomous agent? Bravos AI is an AI chatbot that learns from your data and answers with precision. No months of implementation, starting at €19/month. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Restaurant Chatbot: Answer Menu, Allergen & Reservation Questions 24/7 with AI **Category:** Use case | **Read time:** 11 min | **Date:** Apr 13, 2026 | **URL:** https://bravos-ai.com/en/blog/restaurant-chatbot At [Bravos AI](https://bravos-ai.com/en) we've built an AI chatbot that reads your restaurant's menu, allergen sheet, and reservation policy — and answers your customers based on that information. When someone asks "does the pad thai contain peanuts?" at 11 PM, the chatbot responds with what your allergen sheet says, not with what it makes up. Most restaurant chatbots don't do this. They're either button-based bots ("press 1 to book, 2 to see the menu") or generic AI that doesn't know what dishes you serve. We use a technique called RAG that searches your documents before answering — we explain how it works in [this article about why chatbots hallucinate](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). This article covers what our restaurant chatbot does, what it can't do (honestly), how the pricing compares to Tidio and Intercom, and how to set it up in 30 minutes. ## The problem: most restaurants have zero after-hours support Customer inquiries at restaurants cluster at the worst possible times: the two hours before service and after closing. According to [Yelp for Business](https://business.yelp.com/resources/articles/cost-missed-calls/?domain=restaurants), a restaurant that misses 20–30 calls per week could be losing **$15,600 to $23,400 per year** in unrealized bookings and revenue. And there's a data point from the [MIT/Harvard Business Review](https://hbr.org/2011/03/the-short-life-of-online-sales-leads) that keeps holding up: **responding within 5 minutes makes you 21x more likely to convert an inquiry into a customer**. After 30 minutes, the probability drops by 90%. ## What the Bravos AI restaurant chatbot does ### Answers questions about your menu and allergens Upload your menu as a PDF or paste it as text, add your allergen information, and the chatbot can answer "do you have gluten-free options?" or "how much is the tasting menu?". It doesn't make things up: it searches your information and responds with what it finds. If it doesn't have the answer, it says so. This is especially relevant for allergens. [EU Regulation 1169/2011](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32011R1169) requires all food establishments in the EU to disclose the 14 major allergens. A chatbot trained on your allergen data can fulfill this obligation 24 hours a day, even when no one is around to answer. In the US and UK, similar regulations apply (FDA Food Code, Natasha's Law). ### Responds in 13 languages automatically If your restaurant is in a tourist area, a significant portion of your customers don't speak the local language. According to [CSA Research](https://csa-research.com/Blogs-Events/CSA-in-the-Media/Press-Releases/Consumers-Prefer-their-Own-Language), **76% of consumers prefer to buy in their own language**. With Bravos AI, the visitor writes in German, French, English or Japanese, and the chatbot responds in their language. With a single set of content, no manual translation needed. We explain how this works in [our multilingual chatbot guide](https://bravos-ai.com/blog/chatbot-multiidioma). ### Updates when you change the menu New menu every week? Seasonal dishes? Update the document in the Bravos panel and the chatbot adapts. No coding, no waiting, no one to call. ### Captures potential customer contacts When someone asks about a private event for 30 guests, you probably want their email. The chatbot can ask for it during the conversation, at the right moment, and saves it with the context of what they asked about. No cold forms. More on this in [how to capture leads with a chatbot without being pushy](https://bravos-ai.com/blog/chatbot-captar-clientes). ## Three types of restaurant chatbot (and why the difference matters) Not all chatbots do the same thing, even if they use the same words to sell themselves: | | Button bot | Generic AI | AI trained on your data | | --- | --- | --- | --- | | Understands natural language | No | Yes | Yes | | Knows your menu | No | No | Yes | | Answers allergen questions | No | Hallucinated | Accurate | | Multilingual | No | Partial | 13 languages | The third column is what Bravos AI does. The technique is called RAG (Retrieval-Augmented Generation): instead of generating answers from scratch, it **searches your information and responds with what it finds**. We explain the details in [this article about why chatbots fail with structured data](https://bravos-ai.com/blog/chatbot-catalogo-productos). ## What else you can do with a restaurant chatbot on Bravos AI **Customize the look and tone.** Choose between friendly, professional, or custom tone. Change the widget colors to match your brand, add your restaurant's logo as the avatar, and configure the welcome message. All from the dashboard —or from the Claude or ChatGPT chat itself with [our MCP connector](https://bravos-ai.com/conector-ia-mcp)—, no code. **Upload content in multiple formats.** PDF of the menu, text with your reservation terms, URLs from your website with information about the chef or the restaurant's history. The chatbot extracts text from everything you upload and uses it to answer. **Works on any website.** WordPress, Squarespace, Wix, Webflow, custom-built — doesn't matter. One line of JavaScript you paste wherever you want. You can also restrict which domains it runs on so nobody copies your widget. **Usage analytics.** See how many conversations the chatbot handles, what customers are asking, and how many contacts it captures. Useful for spotting questions people ask that you haven't covered yet. A [Deloitte study (Q4 2024)](https://www.deloitte.com/us/en/about/press-room/deloitte-how-ai-is-revolutionizing-restaurants.html) surveyed 375 restaurant executives across 11 countries and found that **82% plan to increase their AI investment**, while 63% already use AI daily for customer experience. But 48% still don't know which use case to prioritize. Answering repetitive questions 24/7 is probably the simplest one with the highest return. ## How much does a restaurant chatbot cost in 2026 Real prices, verified April 2026: | Platform | Base price | AI included | Notes | | --- | --- | --- | --- | | Tidio | Free (50 conv.); Starter $29/mo | Lyro AI: from $39/mo extra | Real cost with AI: $68+/mo | | Intercom (Fin) | $29/mo | $0.99 per resolution | 100 queries/mo = $99 extra | | Bookline | Not published | Voice AI (phone) | Specialized in phone calls. 1,500 restaurants in Spain | | Bravos AI | 7-day free PRO trial; Starter €19/mo | AI included in all plans | Pro: €49/mo (3 bots, lead capture). No per-resolution fees | For a more detailed pricing comparison, see [How much does an AI chatbot cost for your business in 2026](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa). ## What a restaurant chatbot can NOT do **It doesn't manage reservations directly.** It doesn't connect to OpenTable, Resy, or your reservation book. It can answer questions about your booking policy and provide the link or phone number to reserve, but it doesn't block tables. **It doesn't take complex orders.** "Two pizzas, one without cheese and the other with extra ham, for pickup at 8" requires POS integration. It doesn't do that. **It doesn't handle emotional complaints.** For that you need a person. What it does: answer the questions your staff gets asked every single day — menu, allergens, hours, location, private events — 24/7, in 13 languages, with information you control. For most restaurants, that's already a significant step up. ## How to set up a restaurant chatbot with Bravos AI - **Create your account** at [bravos-ai.com](https://app.bravos-ai.com/register?lang=en) and start the 7-day PRO trial - **Upload your menu** as PDF, text, Excel, or CSV - **Add your allergen info**, opening hours, reservation policy, and anything else you want the bot to know - **Paste one line of code** on your website — works on WordPress, Squarespace, or any platform. Step-by-step guide in [how to add a chatbot to WordPress in 5 minutes](https://bravos-ai.com/blog/chatbot-wordpress) - **Done** — the chatbot answers about your menu, allergens, hours, and everything you've taught it. In 13 languages. 24/7 ## Frequently asked questions about restaurant chatbots ### Can a restaurant chatbot answer allergen questions? Yes, if it's trained on your allergen data. Bravos AI reads your documentation and answers accurately. It doesn't make things up: if the information isn't in your data, it says so. ### Can the chatbot handle reservations? Not directly. It doesn't connect to OpenTable, Resy, or similar systems. But it does answer questions about your reservation policy and provides the booking link or phone number. ### How much does a restaurant chatbot cost? Bravos AI offers a 7-day free trial of the full PRO plan. Starter: €19/month (1 bot, unlimited messages). Pro: €49/month (3 bots, unlimited messages, lead capture). AI is included in all plans, no per-resolution fees. ### Does it work in multiple languages for tourist areas? Yes. It detects the visitor's language and responds automatically. 13 languages from a single set of content, no manual translation. ### How long does it take to set up? About 30 minutes. Upload your menu, allergen info, hours, and policies. Paste one line of code on your website. Done. ### Want to see how it works with your menu? Connect your chatbot to your restaurant's menu and see how it answers. After that, plans start at €19/month. [Try PRO free for 7 days](https://app.bravos-ai.com/register?lang=en) --- # PrestaShop AI Chatbot: Connect Your Catalog and Filter Products **Category:** Use case | **Read time:** 10 min | **Date:** Apr 9, 2026 | **URL:** https://bravos-ai.com/en/blog/prestashop-chatbot You run a PrestaShop store and your customers keep asking the same things: "Is this available in size M?", "How much is shipping?", "Do you have this in black?" You want an AI chatbot that handles these questions 24/7 — and actually understands your product catalog. We built a direct integration between [Bravos AI](https://bravos-ai.com/en) and PrestaShop for exactly that. The chatbot connects to your catalog, syncs automatically, and filters products by any attribute: price, size, color, material, stock — whatever you have. This article covers what it does, how to set it up, and how the pricing compares to the alternatives. ## What AI for PrestaShop actually means, and why stores are adopting it AI on a PrestaShop store comes down to an assistant that lives on your website and answers customer questions automatically, in natural language. Not a button menu or a contact form — the customer types what they need and the chatbot responds like a person would. Modern chatbots use AI models (from OpenAI, Google, etc.) that can understand open-ended questions, keep track of the conversation, and respond with real information about your business. That's why more and more online stores use them to handle after-hours inquiries, answer repetitive questions, and take pressure off their support team. But there's an important distinction for ecommerce: most AI chatbots are built to answer text-based questions (opening hours, policies, FAQs). If you have a product catalog with prices, sizes, and stock levels, you need a chatbot that can work with structured data — not treat your catalog like loose text. That's what our PrestaShop integration is built for. ## What the Bravos AI PrestaShop chatbot does ### Connects to your store without modules PrestaShop comes with a built-in system that lets external apps read your store data (the Webservice). Bravos AI connects through it: you enter your store URL and an access key you generate from your PrestaShop back office. No modules to install, no code to touch, and if you ever disconnect Bravos AI, your PrestaShop stays exactly as it was. ### Imports your full catalog and keeps it updated All your products are imported with their combinations: sizes, colors, prices, images, and descriptions. If you sell t-shirts with Color and Size variants, the chatbot understands these are independent filters and groups them automatically. When you update a price or add a product in PrestaShop, the chatbot picks it up on the next daily sync. No file exports, no manual work. ### Filters products for real, not by approximation When a customer asks "leather bags under €90 in black," the chatbot applies exact filters: material leather, color black, price under 90. It only shows what matches everything. If a bag costs €95, it won't appear. If nothing matches, it says so — it doesn't make things up. This works with any combination of attributes, all at once. Most chatbots can't do this because they treat your catalog as text, not as structured data with real filters. We explain this in detail in [our article on why AI chatbots fail with product catalogs](https://bravos-ai.com/blog/chatbot-catalogo-productos). The same technology we use in our [Shopify chatbot](https://bravos-ai.com/blog/chatbot-shopify), adapted for PrestaShop. ### You control what stock information customers see You choose how the chatbot handles availability: - **Hidden:** doesn't mention availability at all - **Available / Sold out:** just says if it's in stock or not - **Exact stock:** "3 units left" ### Responds in 13 languages automatically It detects the visitor's language and responds in that language. One catalog, no translations needed. Especially useful if you sell across Europe: [76% of consumers prefer buying in their own language](https://csa-research.com/Blogs-Events/CSA-in-the-Media/Press-Releases/Consumers-Prefer-their-Own-Language) according to CSA Research. More in our [multilingual chatbot guide](https://bravos-ai.com/blog/chatbot-multiidioma). ### Captures leads with context Not every visitor buys on the first visit. The chatbot can ask for an email during the conversation and saves it along with what the customer was looking for. If someone asks about standing desks under €400 in walnut and leaves their email, you know exactly what they want. More in our article on [how to capture leads with a chatbot without being invasive](https://bravos-ai.com/blog/chatbot-captar-clientes). ## How to set up your PrestaShop chatbot in 4 steps No technical background needed. The Bravos AI dashboard walks you through it: **1. Connect your store.** In your PrestaShop back office, go to Advanced Parameters → Webservice and create a new key. Bravos AI tells you exactly which permissions to enable. Copy the key, paste it along with your store URL, and click "Test connection." **2. Upload company information.** Return policy, shipping costs, business hours, locations, contact details, company history — anything you want the chatbot to know beyond products. Upload it as documents or free text. **3. Customize the chatbot.** Pick a name, tone (formal, friendly, technical...), colors, and welcome message. On advanced plans you can write a full system prompt to control exactly how it behaves. Language detection is automatic based on the visitor. **4. Add the widget to your site.** One line of code in your PrestaShop theme. The chatbot is now live on your store, answering questions about products and your business, 24/7. The whole process takes less than half an hour. ## How much does a PrestaShop chatbot cost in 2026 Search "PrestaShop chatbot" and you'll find marketplace modules, generic SaaS platforms, and a few specialized options. The advertised price and what you actually end up paying are very different things: ### AI modules from the PrestaShop marketplace Modules like [FME Modules](https://www.fmemodules.com/en/prestashop-modules/248-prestashop-advance-ai-chatbot-module-chatgpt-and-gemini.html) (€119) or [Webkul](https://store.webkul.com/prestashop-ai-chatbot-using-llm.html) ($299). They install directly in your PrestaShop. The module price looks cheap, but all of them require you to create your own OpenAI or Google Gemini account and pay per conversation. Depending on traffic, that's an extra €20–100/month with unpredictable costs. And setting up an OpenAI account and connecting it to the module isn't straightforward if you're not technical. ### Tidio **Advertised:** Starter at $29/month. **What you pay:** the AI (Lyro) is billed separately, from $39/month for 50 conversations. Plus Flows (visual chatbot builder): from $29/month. Real minimum with AI: $97/month. Doesn't connect to your PrestaShop catalog. Source: [Tidio Pricing](https://www.tidio.com/pricing/), [Desk365 analysis](https://www.desk365.io/blog/tidio-pricing/). ### Zendesk **Advertised:** from $19/agent/month. **What you pay:** each AI resolution costs $1–2. The advanced AI add-on: $50/agent/month. 500 resolutions/month = $500–1,000 extra. Built for large enterprises, not small businesses. Source: [Zendesk Pricing](https://www.zendesk.com/pricing/). ### Crisp **Advertised:** Mini at €45/month. **What you pay:** AI is limited to 50 uses/month on the Essentials plan (€95). For unlimited AI you need the Plus plan: €295/month. Has a PrestaShop module, but doesn't filter products from your catalog. Source: [Crisp Pricing](https://crisp.chat/en/pricing/). ### Bravos AI 7-day free trial of the full PRO plan to test it. Starter: €19/month (1 bot, unlimited messages, direct PrestaShop integration). Pro: €49/month (3 bots, unlimited messages, lead capture). AI is included in every plan — no per-conversation charges, no external accounts to set up. | Platform | Advertised price | Actual cost with AI | Filters your catalog | | --- | --- | --- | --- | | PS modules (FME, Webkul) | €119–$299 (one-time) | + €20–100/mo (OpenAI account) | Partial | | Tidio | $29/mo | $97/mo minimum | No | | Zendesk | $19/agent/mo | $219/agent + $1–2/resol. | No | | Crisp | €45/mo | €295/mo (unlimited AI) | No | | Bravos AI | €19/mo | €19/mo (everything included) | Yes (auto-sync) | For a broader comparison (not just PrestaShop), we have a [full AI chatbot pricing breakdown for 2026](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa). ## Limitations of this PrestaShop chatbot - **No cart management or checkout.** It recommends products and links to the product page in your store, but doesn't add to cart. - **No order or return processing.** It can explain your return policy if you've given it that information, but doesn't access your order system. - **Syncs once a day, not in real time.** If you change a price at 10am, the chatbot will reflect it on the next sync. You can trigger a manual sync anytime. - **Doesn't replace human support.** Complaints, damaged shipments, and complicated situations still need a person. ## Frequently asked questions ### Do I need to install a module on PrestaShop? No. Bravos AI connects to the Webservice that comes built into PrestaShop. You just need your store URL and an access key. To show the chatbot on your site, you paste one line of code in your theme. ### Does it search products or only answer general questions? Both. If a customer asks "I need a wooden desk for a small space, max €300," the chatbot filters your catalog and shows only what matches. And if you've also given it information about shipping or your company, it answers those questions too. ### How much does it cost? You get a 7-day free trial of the full PRO plan to test it. Starter (€19/month) includes direct PrestaShop integration, 1 bot and unlimited messages. Pro (€49/month) adds 3 bots and lead capture (also with unlimited messages). AI is included in every plan. ### Does it support multiple languages? Yes. It detects the visitor's language automatically and responds in that language. Supports 13 languages without you having to translate anything. ### Does it update when I change prices or add products? Yes. It syncs your catalog automatically once a day. If you need it updated sooner, you can trigger a manual sync from the dashboard. ## Conclusion If you're looking for a PrestaShop chatbot that actually understands your catalog, Bravos AI connects directly to your store, syncs your products automatically, and filters by any attribute. It takes less than half an hour to set up, starts at €19/month with AI included, and you can [start with 7 days free on the full PRO plan](https://app.bravos-ai.com/register?lang=en) with the PrestaShop integration included. Not on PrestaShop? We also have a native integration and a dedicated guide for [Shopify](https://bravos-ai.com/blog/chatbot-shopify) and [WooCommerce](https://bravos-ai.com/blog/chatbot-woocommerce). ### Want to see how it filters your catalog? Connect your PrestaShop store and test the filtering on your own catalog. After that, plans start at €19/month. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Shopify Chatbot That Actually Filters Your Catalog: Real Prices, Real Comparison **Category:** Use case | **Read time:** 10 min | **Date:** Apr 7, 2026 | **URL:** https://bravos-ai.com/en/blog/shopify-chatbot We built a Shopify chatbot that works like your store's own filter: a customer asks "waterproof jackets under $80 in size L" and only sees products that match all three criteria. If nothing matches, it says so. It doesn't guess, doesn't hallucinate, doesn't show a $150 jacket because it's "also a jacket." Most Shopify chatbots don't do this. They work more like a Google search: you give them some words and they return whatever seems closest. Sometimes they get it right. Often they don't — especially when a customer combines price, size, color, or material in the same question. This article breaks down how our Shopify integration works, how it compares to Tidio, Gorgias, Rep AI, and other options, and what each one actually costs — with pricing verified in April 2026, not the number they put in large font on their homepage. ## The problem: FAQ bots disguised as shopping assistants Search "chatbot" in the Shopify App Store. You'll find dozens of options promising the same things: "AI-powered," "boost your sales," "24/7 support." Install any of them and ask something specific about your catalog: "Do you have waterproof jackets under $80 in size L?" Most return vague or irrelevant results. A $150 jacket. A vest that isn't waterproof. Or worse: a confident-sounding answer that has nothing to do with your actual inventory. Why? Because they're not built to search products. They're support bots in disguise. They can tell customers your return policy or shipping times — and they do that well. But when someone wants to combine filters (price + size + material + availability), they fall short. The technology behind most of them is RAG (Retrieval-Augmented Generation): it converts your products into text fragments and finds the most "similar" ones to the customer's question. The problem is that "similar" isn't the same as "exact." $95 is "similar" to $80 for an algorithm, but not for your customer. We explain this in detail in our article on [why AI chatbots fail with product catalogs](https://bravos-ai.com/blog/chatbot-catalogo-productos). This isn't a niche problem. According to [Grand View Research](https://www.grandviewresearch.com/industry-analysis/chatbot-market), the global chatbot market reaches $11.775 billion in 2026, with ecommerce as the largest segment (~30% of spend). There's a lot of money in chatbots — and a lot of smoke to filter through. ## Two types of Shopify chatbot (and why the difference matters) Not all chatbots do the same thing, even when they use the same words to sell themselves. **Support chatbot (FAQ):** Answers general questions. Shipping policy, business hours, returns, payment methods. Searches relevant text fragments and generates a natural response. Gorgias, Zendesk, and Tidio fall here. They work well for what they're designed to do. **Catalog chatbot (real product search):** Searches products by applying combined filters on structured data. Price, size, color, material, availability — all at once. Doesn't search for similar text; runs real queries. The difference between "finding something close" and "finding exactly what the customer asked for." | | Support chatbot | Catalog chatbot | | --- | --- | --- | | "What's your shipping policy?" | Yes | Yes | | "Waterproof jackets, size L, under $80" | No (or imprecise) | Yes, exact filtering | | Combines 3+ filters at once | No | Yes | | Real-time stock | No (or outdated) | Yes | | Price accuracy | Approximate | Exact ($81 won't show if you ask for < $80) | | Technology | RAG (semantic) | SQL + RAG combined | Most options in the Shopify App Store are the first type. They're not bad — they're just designed for a different problem. If you need a bot that answers "how long does shipping take?," any of them will do. If you need one that actually recommends the right products, you need the second type. ## How a Shopify chatbot that actually understands your catalog works When we built the [Bravos AI](https://bravos-ai.com/en) integration with Shopify, the goal was clear: the chatbot queries your real store data, not an approximate text copy. ### 1. Direct connection via OAuth You connect your store in one click. OAuth — the same system used by Klaviyo or Mailchimp — asks for authorization and you're done. No CSV exports, no copy-pasting. The chatbot has direct access to your Shopify catalog. ### 2. Your full catalog, always up to date All your products are imported with every variant: sizes, colors, prices, SKUs, stock, images, descriptions. If you use Shopify metafields (material, weight, certifications), those too. If you change anything in Shopify — a price, a new product, stock running out — the chatbot updates itself in seconds. You don't have to touch anything. ### 3. Real filtering, not guessing When a customer asks "organic cotton t-shirts in size M under $30," the chatbot doesn't look for similar things. It filters: organic cotton, size M, price under $30. Only shows what matches everything. If a product costs $31, it doesn't show up. If nothing matches, it says so — it doesn't hallucinate. This works with any combination of filters in your catalog: price, size, color, material, availability — whatever you have. All at once. ### 4. You control what customers see You can choose whether the chatbot shows exact stock ("3 units left"), just availability ("in stock" / "sold out"), or nothing about stock at all. Your call. ## What Shopify chatbots actually cost: verified pricing (April 2026) Chatbot pricing for Shopify is a minefield. Nearly every platform advertises an entry price that doesn't include AI, or charges per resolution so the bill spikes with volume. We verified all these prices directly on official pricing pages: ### Tidio **Advertised price:** Starter at $29/month. **What you actually pay:** The AI chatbot (Lyro) is billed separately: from $39/month for 100 AI conversations. Need 500? That's $79/month. 1,000 conversations: $149/month. AI conversations don't count toward your base plan limit — you pay for both. The unlimited plan (Plus) starts at $749/month. Source: [Tidio Pricing](https://www.tidio.com/pricing/) and [Featurebase analysis](https://www.featurebase.app/blog/tidio-pricing). ### Gorgias **Advertised price:** Starter at $10/month (50 tickets). **What you actually pay:** AI is always an add-on: $0.90–$1.00 per conversation resolved. 500 resolutions/month = $450–$500 extra. Plus overages if you exceed included tickets ($0.36–$0.40 per extra ticket). One Shopify App Store merchant reported paying [over $13,500/year](https://www.eesel.ai/blog/gorgias-ai-for-shopify) on the Advanced plan. Source: [Gorgias Pricing](https://www.gorgias.com/pricing). ### Zendesk **Advertised price:** From $19/agent/month. **What you actually pay:** AI resolutions cost $1.50 each. 500 resolutions/month = $750 extra, on top of per-agent licenses. Source: [Zendesk Pricing](https://www.zendesk.com/pricing/). ### Rep AI **Price:** From $99/month. Built for Shopify Plus stores with high traffic. Works as a proactive "sales assistant." Good for large stores, but the entry price is already high for most merchants. ### Bravos AI **Actual price:** 7-day free trial of the full PRO plan to test it. Starter: €19/month (1 bot, unlimited messages). Pro: €49/month (3 bots, unlimited messages, lead capture). Direct Shopify integration included in all paid plans. No per-resolution charges. No AI add-ons. What you see is what you pay. For a broader comparison (not just Shopify), we have a [full AI chatbot pricing breakdown for 2026](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa). ### The hidden cost: keeping data up to date Beyond the monthly fee, there's a cost nobody mentions: the time you spend keeping chatbot data current. If your platform doesn't auto-sync with Shopify, every price change, new product, or stock update means manually updating the chatbot. With a direct OAuth integration and real-time updates, that maintenance cost disappears. ## What to look for in a Shopify chatbot If you're comparing options, ask yourself these questions: **Does it actually filter or just guess?** Test it: ask for products under a specific price. If it returns more expensive ones, it's not filtering — it's guessing. **Does it update automatically?** If you change a price in Shopify and the chatbot still shows the old one, you have a problem. Look for one that syncs on its own. **Is pricing predictable?** Per-resolution pricing can spike during Black Friday or holiday season. If your store has traffic peaks, a flat rate is safer. Stores that respond quickly can increase conversion by up to 69%, according to [Shopify Inbox](https://apps.shopify.com/inbox). If your chatbot gets throttled when you need it most, you lose sales. **Does it support multiple languages?** If you sell internationally, this isn't optional. [76% of consumers prefer buying in their own language](https://csa-research.com/Blogs-Events/CSA-in-the-Media/Press-Releases/Consumers-Prefer-their-Own-Language) according to CSA Research, and 40% won't buy at all if the site isn't in their language. We go deeper in our [multilingual chatbot guide](https://bravos-ai.com/blog/chatbot-multiidioma). **Does it capture leads with context?** Not every visitor buys on the first visit. A chatbot that captures an email along with what the customer asked about ("Gore-Tex jackets, size L, under $150") gives you a qualified lead, not a cold email. More in our article on [how to capture leads with a chatbot without being invasive](https://bravos-ai.com/blog/chatbot-captar-clientes). ## Honest limitations It would be hypocritical to criticize other platforms' marketing and not be transparent about our own. Here's what the Bravos AI Shopify chatbot does **NOT** do: - **No cart management or checkout.** It's a product discovery and recommendation assistant, not a checkout system. It can link directly to the product in your store, but it doesn't add to cart. - **No returns or order processing.** It can explain your return policy (if you've given it that information), but it can't access your order system. - **No replacement for human support in complex cases.** Complaints, damaged shipments, emotional situations — those need a person. - **No aggressive upselling.** It doesn't chase visitors with pop-ups or push products they didn't ask about. It answers what it's asked. If anyone promises a Shopify chatbot that does all of this, be skeptical. ## Frequently asked questions about Bravos AI for Shopify ### Do I need technical skills to connect my store? No. You authorize your Shopify store in one click and Bravos AI imports your catalog automatically. To install the widget, you paste one line of code in your Shopify theme (Online Store > Themes > Edit code > theme.liquid). No coding required. ### Does the chatbot recommend products or just answer questions? Both. If someone asks "I need a gift for a hiker, budget $50–$80," Bravos AI filters your catalog and shows only what fits. These aren't generic recommendations — they're filtered results from your actual catalog. ### How much does it cost? 7-day free trial of the full PRO plan to test it. Starter: €19/month (1 bot, unlimited messages). Pro: €49/month (3 bots, unlimited messages, lead capture). Direct Shopify integration is included in all paid plans. No extra per-resolution fees or AI add-ons. ### Does it support multiple languages? Yes. Bravos AI automatically detects the visitor's language and responds without you having to translate anything. It supports 13 languages. ### Is it better than Shopify Inbox? They're complementary. Shopify Inbox is live chat — it needs someone from your team to be available. Bravos AI responds 24/7 automatically, including product search. The ideal setup is using both: Bravos AI for automated queries and Shopify Inbox for what needs a human. ## Conclusion If your Shopify store sells products with filterable attributes — price, size, color, material, availability — you need a chatbot that actually filters, not one that guesses. And you need it to update automatically when you change something in Shopify, with predictable pricing, and multilingual support if you sell internationally. We built the [Bravos AI](https://bravos-ai.com/en) Shopify integration for exactly this. Direct integration is available from the Starter plan (€19/month). Want to test the filtering first? You get [7 days free on the full PRO plan](https://app.bravos-ai.com/register?lang=en) with the Shopify integration included. Not on Shopify? We also have a native integration and a dedicated guide for [WooCommerce](https://bravos-ai.com/blog/chatbot-woocommerce) and [PrestaShop](https://bravos-ai.com/blog/chatbot-prestashop). ### Want to see how it filters your catalog? Connect your Shopify store and test the filtering on your own catalog. After that, plans start at €19/month. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # How to Capture Leads with a Chatbot Without Being Invasive **Category:** Practical guide | **Read time:** 10 min | **Date:** Apr 1, 2026 | **URL:** https://bravos-ai.com/en/blog/chatbot-lead-capture 97% of your website visitors leave without a trace. No form filled, no email left, no call made. They just vanish. A chatbot can intercept the ones with questions and turn them into real leads — but only if you ask for their data at the right moment, in the right way. At [Bravos](https://bravos-ai.com/), we've spent months fine-tuning exactly that: when to surface the form, what fields to include, and how to weave lead capture into the conversation without breaking it. This article distills what we've learned, backed by data and real conversation examples. ## Your visitors aren't anonymous by accident The average website conversion rate sits between 2% and 3%. But those visitors who leave aren't leaving because they're not interested. Many had questions they didn't know where to ask, or didn't want to commit to a form. Others simply didn't find what they were looking for fast enough. The goal isn't to convert 100% — that's impossible. The goal is to give those undecided visitors a channel where they can talk without commitment. If out of those 97 visitors who leave, 5 end up talking to a chatbot and 2 leave their email, you've just multiplied your contacts. ## Why people ignore your contact form A form asks for effort in exchange for nothing immediate. The visitor has to decide what to write, hand over their data to a company they don't know, and wait for a response that might take hours or never come. The average conversion rate for contact forms is 1% according to Formstack (665,000 forms analyzed). The root problem is reciprocity: the form asks before giving. A chatbot flips the equation: first it solves something (a question, a recommendation, a comparison) and then it asks. We wrote a [full analysis of chatbots vs forms](https://bravos-ai.com/blog/formulario-contacto-vs-chatbot) with data from Gartner, HubSpot and Formstack. If you want to dig into the numbers, it's there. Here we focus on something else: **how to ask for the visitor's data without them closing the tab.** ## What 90% of chatbots get wrong If you've ever landed on a website and had the entire chat window pop open after 2 seconds with a "Hi! How can I help you?" taking up half your screen, you know what we're talking about. You didn't have any question yet. You closed it automatically. Like everyone else. These are the most common mistakes: **Asking for data before providing any value.** "Leave your email so I can help you better" is the conversational version of the contact form. The visitor sees no value in giving you their email if they don't even know if you can help. **Giving generic responses.** The visitor asks "Do you ship to Hawaii?" and the bot responds "For shipping inquiries, contact our team." If the chatbot can't answer specific questions, it's not doing its job. **Being overly enthusiastic.** "Great question!! So glad you asked!" when someone asks about your hours. Artificial tone breeds distrust. **Blocking the conversation with a form.** Some chatbots force you to fill in name, email and phone before letting you type. That's not a chatbot, it's a form with a bubble design. A bad chatbot is worse than no chatbot at all. Because it trains your visitors to ignore the widget. Next time they see the chat icon on your site, they won't even open it. One thing that often gets overlooked: the bot that gives generic answers isn't generic because it's bad — it's generic because **it doesn't have your actual business data.** Prices, stock levels, shipping policies, opening hours, service descriptions. Without that information, any chatbot (no matter how smart) ends up responding with empty phrases. The fix is feeding the bot real content from your business: product sheets, catalogs, FAQ pages, internal docs. Platforms like Bravos let you upload documents, spreadsheets, or connect your product catalog directly — and the bot gives concrete answers because it has concrete data. ## The exact moment to ask for data The rule is simple: help first, ask later. But "help first" doesn't mean answering one question and dropping the form. It means generating enough value for the visitor to feel the conversation is worth continuing. There are clear signals that someone is ready: - **They've asked 2 or more questions.** They're engaged in the conversation. There's real interest. - **They ask about price, availability or shipping.** These are buying signals. They're evaluating whether to purchase. - **They're comparing options.** "What's the difference between model A and B?" — High intent. - **They ask something the bot can't answer.** Perfect opportunity: "Someone from the team can help with that. Want me to connect you?" Here's what a natural transition looks like in a real chatbot: The bot answered two real questions with concrete data. It provided value. And when it detected buying intent (asking about shipping = evaluating whether to purchase), it offered the form as a bridge to the human team. It didn't ask for a bare email — it offered to connect the visitor with someone who can help with what they need. The form appears as the logical next step, not a formality. Now compare with this: The visitor hasn't received anything yet. They don't know if the bot can actually help. And they're already being asked for data. They close the chat. ## What to ask, how much, and who decides when According to HubSpot, reducing a form from 4 to 3 fields increases conversion by 50%. In a chatbot the principle is the same: **the less you ask for, the more people respond.** For a first touch, email is usually enough. You already have the conversation context (what product they were looking at, what they asked, what doubts they had). With that and an email you can do personalized follow-up that's worth more than a 6-field form full of generic data. How you ask matters as much as when. And here's the key difference between basic chatbots and ones that actually work: **who decides when to show the form.** Most chatbots either rely on rigid rules ("show the form after the third message") or let the bot decide on its own. Both fail: fixed rules ignore the conversation context, and an unguided bot drops the form at awkward moments. What works is a middle ground: **you tell the bot when it's appropriate, and the bot judges the moment.** For example, you can set something like "offer the form when the user asks for a quote or shows real interest in a specific product." The bot reads the conversation and decides whether that moment has arrived. If it hasn't, it keeps helping without forcing anything. Real configuration in Bravos: you complete the instruction "Offer the contact form when..." and the bot judges the moment. } /> This has several advantages: the form appears when it makes sense to the visitor (not mechanically after 3 messages), the bot never asks for data in the chat itself (the form handles collection), and if the visitor doesn't fill it out, the conversation continues normally. On top of that, the system automatically controls how often the form is offered so it never feels pushy. And if you use Claude or ChatGPT, with [our MCP connector](https://bravos-ai.com/conector-ia-mcp) you ask the assistant to fine-tune when the form is offered, and it applies our best practices to capture without scaring the visitor off, without you touching the panel. Don't ask for a phone number unless absolutely necessary. Phone fields alone can reduce conversion by up to 5% — people fear sales calls. For a first touch, name and email are usually enough. You already have the conversation context as a complement. ## The first 5 minutes after contact You capture the lead perfectly. The visitor leaves their email after a productive conversation. And then... you take 3 days to respond. You've lost the sale. The MIT study on response times is clear: **contacting a lead within the first 5 minutes makes you 21 times more likely to qualify them**, compared to waiting 30 minutes. Wait 10 minutes instead of 5, and the odds already drop 4x. Harvard Business Review analyzed 2,241 companies and found that only 37% responded within the first hour. The average response time was 42 hours — nearly two days. Companies that responded within one hour were 7 times more likely to qualify the lead. The chatbot gives you an advantage that a form doesn't: context. You don't receive a cold email saying "I'm interested in information about your services." You receive a notification with everything the visitor discussed: what product they were interested in, what questions they asked, what doubts they had. With that you can respond like: "Hi Maria, I'm Jake from HomeGear. I saw you were interested in the Cosori Pro II for dehydrating — it's one of the best in that price range. Shipping to Austin takes 2-3 days and it's free since it's over $40. Want me to reserve one for you?" That's not a generic response. It's a response that shows you read the conversation, understand what they need, and you're solving their problem. For this to work you need three things: - **Instant notification.** Email or webhook that arrives the second a lead is captured. If you see it 8 hours later, the visitor already bought elsewhere. - **Conversation context.** What they asked, what products they looked at, what doubts they had. If you respond without context, you lose the advantage the chatbot gave you. - **Clear expectations.** Have the chatbot say something like "someone from the team will write to you in the next few hours." So the visitor knows something will happen. If you can't respond within an hour, at least make sure the chatbot sets expectations in the conversation itself: "Someone from the team will reach out within a few hours." The bot doesn't need to send emails — the notification to the business owner is what matters. What counts is that the visitor doesn't leave wondering what happens next. ## Summary: the complete sequence - **Don't interrupt.** Let the visitor explore. The chatbot should be visible but not jump on them after 2 seconds. - **Answer with real data.** Prices, availability, specs. If the chatbot can't answer specific questions about your business, it's not doing its job. - **Detect interest.** Multiple questions, questions about price or shipping, comparisons — those are the signals. - **Ask for little.** Email, at most. Framed as something useful for the visitor, not a formality. - **Respond fast.** A lead is 21 times more valuable if you contact them in 5 minutes vs half an hour. - **Use the context.** The chatbot already knows what the visitor wants. Use it in your response. ## Frequently asked questions ### When should a chatbot ask for the visitor's email? After providing real value — never before. The best signals are: the visitor has asked 2 or more questions, they ask about price, availability or shipping (buying signals), they're comparing options, or they ask something the bot can't resolve. Never ask for data in the first message. ### What data should a chatbot ask for to capture a lead? Email only, in most cases. According to HubSpot, reducing a form from 4 to 3 fields increases conversion by 50%. Phone fields can reduce conversion by up to 5% because people fear sales calls. You already have the conversation context — with an email and that context, you can do effective personalized follow-up. ### How fast should I respond after a chatbot captures a lead? Within 5 minutes, ideally. The MIT Lead Response Study found that contacting a lead within 5 minutes makes you 21 times more likely to qualify them compared to waiting 30 minutes. Harvard Business Review found that only 37% of companies respond within the first hour, and the average response time is 42 hours. ### How does the chatbot know when to show the form? It depends on the platform. Basic ones use fixed rules (after X messages). The better ones let you give the bot instructions about when it's appropriate (e.g., "when they ask for a quote or show real interest in a product") and the bot reads the conversation to judge whether that moment has arrived. That way the form appears when it makes sense, not mechanically. ### Want to capture leads while you sleep? Create your chatbot for free. Lead capture built into the conversation. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Ecommerce Chatbot: How to Stop Losing Sales Because No One Is Answering **Category:** Practical guide | **Read time:** 12 min | **Date:** Mar 25, 2026 | **URL:** https://bravos-ai.com/en/blog/ecommerce-chatbot It's Friday at 11pm. Someone is browsing your online store. They have a product in their cart. But they have a question: is shipping free above €50? No one's there to answer. They leave. You never know they existed. An ecommerce chatbot solves exactly that. This happens every single day in thousands of online stores. And it's not just shipping questions. Sizing, compatibility, returns, payment options. Questions that, if someone answered them, would turn into sales. An [MIT study](https://cdn2.hubspot.net/hub/25649/file-13535879-pdf/docs/mit_study.pdf) found that responding to a web inquiry within 5 minutes — instead of 30 — makes you 21 times more likely to qualify that lead. And according to [InsideSales.com](https://www.ringly.io/blog/ecommerce-customer-support-statistics-2026), 78% of sales go to the first company that responds. Your store is open 24/7. Your customer support probably isn't. In this guide, we'll cover what a chatbot for online stores can actually do (no hype), how it works with Shopify, WooCommerce, and PrestaShop, and how you can set one up in under 5 minutes with [Bravos AI](https://bravos-ai.com/). ## The real cost of not responding The [Baymard Institute](https://baymard.com/lists/cart-abandonment-rate) has been tracking cart abandonment data for years. Their current average, calculated from over 50 studies: **70% of online shopping carts are abandoned**. Not all of that 70% is recoverable. Many people are "just browsing." But a significant portion has concrete questions that no one answered. According to [Contentsquare](https://contentsquare.com/guides/cart-abandonment/stats/), 55% abandon because of unexpected costs — shipping fees, taxes, and charges they didn't see coming. A chatbot that explains shipping terms before checkout removes that friction entirely. European ecommerce hit **€842 billion in 2024**, growing 7% year-over-year. The market is expanding, competition is intensifying, and customer expectations keep rising. 75% of customers already expect 24/7 service ([Freshworks, 2025](https://www.freshworks.com/customer-service-statistics/)). But the average response time in online stores is 4 to 6 hours. Customers expect under 30 minutes. The gap is massive. ## What a chatbot can do for your store (no hype) An AI chatbot won't magically transform your business. But it can solve one very specific problem: **answering the repetitive questions your team already handles every day**, except at any hour and with zero wait time. ### What you can do with Bravos AI - **Answer product questions** — sizing, materials, compatibility, dimensions, variants - **Search products with real filters** — "I need an iPhone 15 Pro case under €25 in black" → exact results, not approximate - **Make all your business information accessible** — shipping policies, returns, size guides, differences between products... your bot knows everything you've taught it and explains it directly. No digging through pages — not every visitor is good at navigating websites. - **Capture contacts** — the chatbot asks for email or phone at the moment you choose, along with the full conversation context - **Work 24/7** — no holidays, sick days, or night shifts - **Respond in 13 languages** — auto-detects the visitor's language, no translation needed ### What no chatbot does (let's be honest) This applies to every chatbot on the market, including ours: - **Doesn't process payments or manage carts** — it's not a checkout assistant - **Doesn't replace human support** — complex complaints still need a real person - **Doesn't turn a bad offer into sales** — but it does convert visits that would leave because they couldn't find the information they needed to make a decision - **Doesn't fix a bad website** — if your site is slow, prices aren't competitive, or photos are poor, a chatbot won't compensate ## The catalog problem: why most chatbots fail here This is where most ecommerce chatbot platforms fall short. A customer asks: "Do you have organic cotton T-shirts in size M under €30?" The chatbot needs to filter your catalog by material, size, and price. Three filters at once. Sounds simple. But most chatbots treat your catalog as if it were a text document. They search for similar words, not exact data. It's like asking someone to search through a book — they might find something close, but they can't filter precisely. If a customer asks for "under €30," they might return products at €35 or €50 because the word "T-shirt" appeared nearby. We cover this in depth in [our article about chatbots and product catalogs](https://bravos-ai.com/blog/chatbot-catalogo-productos) , but here's the summary: the leading chatbot platforms — Chatbase, Intercom, Zendesk, Tidio — can't filter your catalog the way a database would. They treat your products as text, not as data. [Bravos AI](https://bravos-ai.com/) works differently. Upload your catalog as a CSV and the chatbot filters with precision: if a customer asks for "organic cotton T-shirts in size M under €30," it only shows products that match all three criteria. If one costs €31, it doesn't show up. If there are no results, it says so — no making things up or suggesting "something similar." ## Shopify, WooCommerce, PrestaShop: works with all of them Shopify has 6.5 million active stores globally. WooCommerce has between 4.5 and 6 million. PrestaShop remains strong in Southern Europe with over 17,500 stores in Spain alone ([StoreLeads](https://storeleads.app/reports/prestashop/ES/top-stores)). Whatever platform you use, Bravos AI works with it. For the big three there's a **native integration**: connect your store and the catalog syncs itself, with variations, stock and sale prices up to date. We have a dedicated guide for each platform: [Shopify chatbot](https://bravos-ai.com/blog/chatbot-shopify), [WooCommerce chatbot](https://bravos-ai.com/blog/chatbot-woocommerce) and [PrestaShop chatbot](https://bravos-ai.com/blog/chatbot-prestashop). Is your store built on Wix? There's a guide too — there the catalog runs on a synced sheet, for now: [Wix chatbot](https://bravos-ai.com/blog/chatbot-wix). And if you use another platform, the universal method still works: export your catalog as a CSV and upload it to Bravos AI. Every ecommerce platform supports CSV export. On top of the catalog, you can upload documents with all your additional information — shipping policies, returns, FAQs, size guides — so the chatbot has the full context of your business. ## Multilingual support: if you sell internationally, this isn't optional [CSA Research](https://www.newswire.com/news/survey-of-8-709-consumers-in-29-countries-finds-that-76-prefer-21174283) surveyed 8,709 consumers across 29 countries. The result: **76% prefer buying products with information in their native language**. And the hardest-hitting number: 40% *never* buy from websites in other languages. Never. If your store serves international customers — whether you sell across Europe, attract tourists, or operate in a multilingual market — multilingual customer support isn't a nice-to-have. It's what separates stores that convert from those that don't. Bravos AI automatically detects the visitor's language and responds accordingly. It supports 13 languages without you having to translate anything. Upload your content in English (or whatever language you use) and the chatbot responds in Spanish, French, German, Portuguese, or Arabic depending on who's asking. We explain it in depth in [our multilingual chatbot guide](https://bravos-ai.com/blog/chatbot-multiidioma). ## Lead capture: every conversation is a contact A visitor who talks to your chatbot and asks about a specific product is a warm lead. They have real purchase intent. But if they leave without sharing their contact details, they're gone. Built-in lead capture lets your chatbot ask for name, email, or phone at a specific point in the conversation — you decide when. Without blocking the chat: the visitor can fill in the form or keep browsing without interruption. The best part: you receive the contact **along with the full conversation history**. It's not a cold email from a generic web form. It's a contact who asked about "Bluetooth noise-cancelling headphones under €100." Your sales team knows exactly what they want before picking up the phone. ## How to set up a chatbot for your online store (3 steps, really) No fluff: 1 #### Create your bot [Sign up](https://app.bravos-ai.com/register?lang=en) and start the 7-day PRO trial. Create your first chatbot: give it a name, choose the tone, and set the primary language. No technical knowledge required. 2 #### Upload your content Upload your product catalog as CSV for exact filtering. Add documents with your shipping, returns, size guides, and FAQ policies. The more information you give it, the better it responds. 3 #### Install the widget Copy one line of code and paste it on your site. The Bravos AI app shows platform-specific installation instructions for Shopify, WooCommerce, and WordPress. For other platforms, just paste the snippet into your site's HTML. ## How much does it cost? You start with 7 days free on the full PRO plan to test it with your store and see if it works. To continue, the Starter plan costs €19/month (1 bot, unlimited messages) and Pro is €49/month (3 bots, unlimited messages, lead capture). For perspective: €19 per month is less than one hour of outsourced customer support. And the chatbot works all 720 hours of the month. There's also an Enterprise plan for stores that need more. It's not as expensive as it sounds, and the key difference is access to premium AI models like GPT-4o or Claude — which produce noticeably better responses than the standard model (GPT-4o-mini). If your catalog is large or your customers expect high-quality answers, it's worth asking us about. Want to compare with other platforms? [We have a detailed chatbot pricing analysis for 2026](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa). ## Conclusion: is an ecommerce chatbot worth it? A chatbot isn't the solution to every problem in your online store. But it solves one specific issue that costs you money every day: not having anyone available when a customer has a question. What separates a useful ecommerce chatbot from a useless one isn't the AI model underneath — they almost all use GPT. It's **how it accesses your data**. If it treats your catalog as text, it'll fail at the questions that matter. If it runs real queries on your data, it'll work. And if you sell internationally, supporting multiple languages without you having to translate anything isn't a bonus — it's a requirement. With [Bravos AI](https://bravos-ai.com/) you can set up a chatbot for your online store in under 5 minutes, with real catalog filtering and automatic multilingual support. [Start with 7 days free on the full PRO plan](https://app.bravos-ai.com/register?lang=en). ## Frequently asked questions ### Does it work with my ecommerce platform? Yes. Bravos AI has native integrations with Shopify, WooCommerce and PrestaShop (the catalog syncs itself, no exports needed), and works with Magento and any standard HTML website by uploading the catalog as a CSV. The app includes step-by-step installation instructions for each platform, and in no case do you need to install plugins on your store. ### Can the chatbot handle returns or complaints? It can explain your returns policy, deadlines, conditions, and the step-by-step process. But it can't process a return for you — that requires access to your order management system. For complex complaints, the chatbot collects the customer's details and notifies you by email so your team can handle it. ### Does it work on mobile? Yes. The widget is fully responsive and works on all browsers: Chrome, Safari, Firefox, Edge. Customers can even send product photos or record voice questions. ### How much does an ecommerce chatbot cost? You can try the full PRO plan free for 7 days. The Starter plan costs €19/month with unlimited messages and covers most mid-size stores. There's also an Enterprise plan with premium AI models for stores that need more. ### Is your online store losing sales because no one's answering? Set up your chatbot with catalog filtering and multilingual support in under 5 minutes. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Chatbot Legal Liability: What Happens When Your AI Chatbot Gets It Wrong (and How to Protect Yourself) **Category:** Analysis | **Read time:** 7 min | **Date:** Mar 17, 2026 | **URL:** https://bravos-ai.com/en/blog/chatbot-legal-liability Your chatbot just told a customer they can return a product past the deadline. Or that the treatment includes a free follow-up. Or that the apartment comes with a parking space. None of it is true. But the customer already made a decision based on that information. Who bears the legal liability? The chatbot? The company using it? The platform that built it? There's already case law that answers this question. And Air Canada didn't like the answer. ## The Air Canada chatbot case: wrong information, real consequences In 2022, Jake Moffatt needed to fly urgently to his grandmother's funeral. Before buying the ticket, he asked Air Canada's chatbot whether he could apply for the bereavement fare after travelling. The chatbot said yes: he had 90 days after the flight to request a partial refund. That was wrong. Air Canada's actual policy requires bereavement fares to be requested before travel, not after. But the chatbot got it wrong. Moffatt trusted that information, bought the ticket at full price, and when he applied for the refund, Air Canada denied it. He took the case to a civil tribunal in British Columbia. And he won. ## Who is liable when an AI chatbot gives wrong information? What the judge ruled Air Canada's defence was striking: they argued the chatbot was "a separate legal entity responsible for its own actions." The tribunal member didn't buy it. The ruling was clear: "While a chatbot has an interactive component, it is still just a part of Air Canada's website. It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or from a chatbot." Air Canada had to pay CAD $812 in refund and costs. The amount was small. The precedent was not. ## Your chatbot, your liability If you have an AI chatbot on your website answering customer questions, legally it's you who's answering. Not the chatbot platform. Not the AI. Not a "separate entity." You. That's why in sensitive sectors the bot is configured not to overstep: a [law firm chatbot](https://bravos-ai.com/blog/chatbot-para-abogados), for instance, informs and refers, but doesn't give legal advice. If the chatbot says a treatment costs €200 when it actually costs €350, the liability is yours. If it says you ship to the Canary Islands and you don't, the liability is yours. If it confirms an appointment that doesn't exist, the liability is yours. This isn't theory — there's already case law backing it up. ## EU AI Act: chatbot regulation taking effect August 2026 As if the Air Canada case weren't enough, on August 2, 2026, most provisions of the [EU AI Act](https://artificialintelligenceact.eu/) take effect — the most ambitious AI regulation in the world. Some obligations have been in force since February 2025, and full enforcement is completed by August 2027. If your business operates in Europe or sells to European customers, this applies to you. The EU AI Act classifies AI systems by risk level. A customer service chatbot falls under "limited risk," which comes with one specific obligation: transparency. Your customers must know they're talking to an AI, not a person. This isn't optional. From August, it's law. And in Spain there's already an agency dedicated to enforcing it: the [AESIA](https://aesia.digital.gob.es/) (Spanish Agency for the Supervision of Artificial Intelligence), the first of its kind in the European Union, headquartered in A Coruña. This isn't a piece of paper gathering dust in Brussels. ### What are the EU AI Act fines for non-compliance? Penalties can reach €35 million or 7% of global annual turnover, whichever is greater. That sounds terrifying, but there's an important nuance for [small businesses](https://bravos-ai.com/blog/chatbot-para-pymes) : fines are calculated on the lower of the two amounts. A company with €2 million turnover would pay a maximum of €140,000 for the most serious infringement, not €35 million. Still a lot of money. But it's not the apocalyptic figure you read in the headlines. ### Is my chatbot "high risk"? It depends on the use case, not the technology. The same AI model can be minimal risk if you use it to draft internal emails, limited risk as a customer service chatbot, or high risk if it screens job candidates. If your chatbot simply answers questions about your products or services, the obligations are reasonable: transparency and accurate data. You don't need a legal department to comply. ## How to protect yourself from chatbot legal liability Now for the practical part. ### 1. Make sure your chatbot doesn't hallucinate Air Canada's problem wasn't having a chatbot. It was that the chatbot gave wrong information. If your chatbot is trained on your real, up-to-date data, the chance of it making things up [drops dramatically](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo) . Chatbots with RAG (Retrieval-Augmented Generation) search your documents before answering, instead of improvising. What if your data changes? If you raise prices, change your terms of service, or stop shipping to a certain area, your chatbot needs to know. Outdated data is what gets you in trouble. If you use a chatbot with automatic data sync, this problem disappears: you update your spreadsheet and the chatbot already knows. ### 2. Disclose that it's an AI Keeping your chatbot accurate protects you from lawsuits like Air Canada's. Disclosing that it's an AI protects you from EU AI Act fines. These are two separate risks, and you need to cover both. The EU AI Act will require it, but it's good practice regardless. A clear notice that they're talking to a virtual assistant. You don't need a 47-paragraph legal disclaimer. It's enough if the chatbot's name, the welcome message, or a label on the widget makes it clear that it's a bot. ### 3. Provide a path to a human A chatbot shouldn't be a wall. If the query is complex, if the customer insists, if there's a real issue, there needs to be a way to reach a person. An email, a phone number, a contact form. The chatbot handles 80% of repetitive queries. The remaining 20% needs someone from your team. ### 4. Don't let it make promises Air Canada's chatbot "confirmed" that Moffatt could request the refund after travelling. If instead of confirming, it had said "according to our policy, check the details at [link] or contact our team," there probably wouldn't have been a lawsuit. Clear instructions in the chatbot's prompt: don't confirm bookings, don't guarantee prices that might change, don't promise deadlines you can't keep. Inform, don't commit. ## Don't be Air Canada A chatbot on your website is a useful tool. But it's not an intern you can ignore. What your chatbot says, you say. The good news is that protecting yourself isn't hard: up-to-date data, transparency, a path to a human, and a well-configured prompt. These are things that improve the customer experience anyway, with or without regulation. In fact, that's exactly what we do at [Bravos AI](https://bravos-ai.com/) . Our chatbots only respond with the data you provide and sync automatically when you update it. And you control the name, the welcome message, and the tone — just enough to make it clear it's a bot, without scary disclaimers. Air Canada tried to blame their chatbot. The judge said no. Don't be Air Canada. ## Frequently asked questions ### Who is liable if an AI chatbot gives wrong information? The company that deploys it. The [Moffatt v. Air Canada](https://www.mccarthy.ca/en/insights/blogs/techlex/moffatt-v-air-canada-misrepresentation-ai-chatbot) case established that a chatbot is part of a company's website, and the company is responsible for all information on its site, whether it comes from a static page or from a chatbot. ### Do AI chatbots have to disclose they are AI? Yes, from August 2026 under the EU AI Act. AI systems that interact directly with people must inform users that they're talking to an AI, not a human. ### What if my chatbot is classified as "high risk"? It depends on the use case. A customer service chatbot is "limited risk" (transparency only). But if your chatbot screens job candidates, evaluates credit applications, or makes decisions that directly affect people, it's "high risk" and requires impact assessments, technical documentation, and human oversight. ### How much can EU AI Act non-compliance cost? Fines can reach €35 million or 7% of global annual turnover. For SMEs, the fine is calculated on the lower of the two amounts. A company with €2M turnover would pay a maximum of €140,000 for the most serious infringement. ### Want a chatbot that won't get you in trouble? Build your chatbot with up-to-date data, auto-sync, and full transparency. In under 5 minutes. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Rule-Based Chatbot vs AI Chatbot: Key Differences and How to Choose **Category:** Analysis | **Read time:** 14 min | **Date:** Mar 12, 2026 | **URL:** https://bravos-ai.com/en/blog/rule-based-chatbot-vs-ai-chatbot Most chatbots you see on websites are rule-based: buttons, menus, and scripted flows that work until someone asks a question that wasn't in the script. AI chatbots are a different thing entirely: they understand what the user writes and generate answers based on your actual data. But not all AI chatbots are equal, and not all are good. Here's an honest breakdown so you can choose the right one. ## What is a rule-based chatbot (and how does it work) A rule-based chatbot — also called a flow-based chatbot, scripted chatbot, or decision tree chatbot — follows a predefined script. You design every possible conversation path: "If the user clicks button A, show option B. If they choose C, respond with D." They work like the classic "press 1 for sales, press 2 for support" phone system, but in chat form. Platforms like ManyChat, Landbot, or Chatfuel let you build these flows with drag-and-drop visual editors. **Strengths:** - **Predictable.** You know exactly what it will respond in every case - **Easy to build.** Drag blocks, connect arrows, done - **Cheap.** Many platforms have free plans - **No hallucinations.** Since they don't generate text, they can't make up information **Limitations:** - **No natural language understanding.** The user can't ask what they want — they can only choose from the buttons you've set up. If their question doesn't fit any option, there's no way forward - **Doesn't scale.** 10 products with 5 variants each = 50 branches. 100 products = maintenance nightmare - **Single language.** Each additional language is another complete decision tree to maintain - **No memory.** Each conversation turn starts from scratch - **No search capability.** Can't search a catalogue or documents. Only shows what was hardcoded into the flow ## What is an AI chatbot (and why most aren't what they claim) An AI chatbot — sometimes called a conversational AI chatbot — uses language models (like the ones behind ChatGPT) to understand what the user writes and generate responses. There are no flows to design: you give the bot your information (documents, website, product catalogue) and it answers questions about it. But "AI chatbot" is a very broad term. Some platforms just bolt ChatGPT onto a text input and call it an AI chatbot. Technically it is, but the result is very basic: generic answers, no access to your actual data, no conversation memory. We can't speak for every platform, but we can speak for ours. At [Bravos AI](https://bravos-ai.com/) we build AI chatbots, so here's what a well-built one can actually do: **Understands what the user means.** No buttons needed. If someone writes "something affordable with good reviews," the bot understands what they're looking for. If they write "I want to return an order" or "my order hasn't arrived," same thing. It understands intent, not just exact keywords. **Searches your information, not the internet.** You upload your documents, product catalogue, or FAQs in a single language, and the bot answers based on that — in whatever language the visitor uses. If someone asks "how many days do I have to return a product?" it answers with your actual returns policy. If someone else asks the same thing in Spanish, it translates on the fly. Without you having programmed any of those questions. **Responds in the visitor's language.** Someone writes in Polish, the bot answers in Polish. In Arabic, in Arabic. Without translating your content or maintaining separate versions for each language. For businesses with [international customers](https://bravos-ai.com/blog/chatbot-multiidioma) , this is a massive shift from scripted chatbots, where each language is another flow to build and maintain. **Remembers the conversation.** If you say "Nike running shoes" and then ask "something cheaper?", the bot knows you're still looking for Nike shoes. It doesn't start from scratch with every message. **Filters catalogues like a person would.** "Laptop with 16GB RAM under €800" — the bot extracts those filters (category, memory, max price) and searches your actual catalogue. It doesn't get confused or return random results. This is [much harder than it sounds](https://bravos-ai.com/blog/chatbot-catalogo-productos) , and most platforms don't do it well. **Captures leads when it makes sense.** Instead of showing a forced pop-up at 30 seconds (which almost nobody fills in), the chatbot offers the contact form when the conversation calls for it — when the user shows genuine interest. Chatbots with this approach [convert up to 3x better than traditional forms](https://bravos-ai.com/blog/formulario-contacto-vs-chatbot) (Drift). **Easy to update.** Change a document, upload a new spreadsheet, or edit an FAQ — the chatbot uses the updated information. No flows or branches to touch. With a rule-based chatbot, every change means a new branch to create manually. **Limitations (being honest):** - **Can hallucinate.** If poorly configured, the model can generate information that sounds convincing but is completely false. It can be minimised significantly (using only your content as source, filtering unreliable results, giving the model clear instructions), but the risk is never zero. [More on why chatbots make up answers](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo) - **Less predictable.** Two identical questions might generate slightly different responses - **Cost per message.** Each response consumes AI model resources. At high volumes this can add up, though [it's still far cheaper than a human agent](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa) ## Rule-based chatbot vs AI chatbot: side-by-side comparison Resolution rates come from Gartner and Intercom. The rest are inherent technical differences between each type. | | Rule-based chatbot | AI chatbot | | --- | --- | --- | | Resolution without escalation | 20-40% | 51-75% | | User understanding | Predefined buttons only | Natural language | | Languages | 1 per flow | Any the model supports | | Memory | No | Yes (varies by platform) | | Searches your data | No | Yes (varies by platform) | | Setup | Visual editor, fast | Upload content, configure | | Maintenance | Manual (each branch) | Update the source content | | Scales with products | Poorly (>50 = chaos) | Thousands, no problem | | Hallucination risk | None | Low-medium (depends on setup) | The resolution data is key: Gartner documented how Solo Brands jumped from 40% to **75% resolution** after switching from a rule-based chatbot to a generative AI one. Intercom reports **51% resolution** with their Fin agent, with 99.9% accuracy. Rule-based chatbots stay in the 20-40% range because they can only resolve what was explicitly programmed. ## When a rule-based chatbot is enough You don't always need AI. A scripted chatbot works well when: - **You have fewer than 10-15 FAQs** that cover 90% of enquiries - **The flow is linear and closed:** appointment booking, service selection, satisfaction survey - **No product catalogue** or database to search - **You only operate in one language** - **The goal is to guide**, not to answer. Example: "Would you like to book an appointment? Yes / No" If your business fits this profile, a rule-based chatbot at €0-30/month gets the job done. No need to overcomplicate things. ## When you need a conversational AI chatbot AI makes the difference when the conversation isn't predictable: **Large catalogues.** 500 products with different prices, features, and categories don't fit in a button flow. With AI, the user describes what they're looking for and the bot filters through everything. A real estate agency with hundreds of listings, an online shop with thousands of SKUs — same principle. **International customers.** 76% of consumers prefer to buy in their native language (CSA Research, 8,709 respondents across 29 countries). 40% won't buy at all if the website isn't in their language. An AI chatbot responds in the visitor's language without you having to translate anything. **Extensive documentation.** Technical manuals, company policies, product guides. Upload the documents and the chatbot answers about any of them, without designing 200 flow branches. **Qualification conversations.** A [proactive chatbot](https://bravos-ai.com/blog/chatbot-proactivo-vs-reactivo) asks questions to understand the customer ("What's your budget?", "How many people is it for?") and adapts the conversation based on their answers. With rules, every possible combination is another branch. **Ad campaigns.** Every [Google Ads or Meta Ads](https://bravos-ai.com/blog/chatbot-campanas-publicidad) click you don't respond to is wasted money. An AI chatbot handles that traffic 24/7 and qualifies leads automatically, without visitors leaving because they couldn't find what they needed among the available buttons. ## Cost of a rule-based chatbot vs AI chatbot **Rule-based chatbot:** €0-50/month. ManyChat has a free plan. Landbot starts at €40/month. The cost is fixed because they don't use AI models — they only serve static content. **AI chatbot:** self-service platforms where you set it up yourself. Chatbase from $19 USD/month, [Bravos AI](https://bravos-ai.com/) with 7 days free on the full PRO plan and plans from €19/month. You can have them running in an afternoon without depending on anyone. [Full pricing comparison and how to calculate ROI](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa) . **Full support platforms:** solutions like Intercom or Zendesk include an AI chatbot, but they're entire customer support platforms (ticketing, inbox, knowledge base, agent teams). Starting at €100-300/month. If all you need is a chatbot, it's like buying a car just to use the GPS. In all cases, the chatbot cost is a fraction of human support costs. A chatbot interaction costs **€0.50-0.70** versus **€6-15** for a human agent interaction — 10 to 12 times less. McKinsey reports that companies implementing AI in customer care see **15-40% cost reductions in the first year**. Gartner predicts that by 2029, AI will autonomously resolve 80% of common customer service issues. ## So, which one should you choose? Don't overthink it. Trying an AI chatbot is free or nearly free on most platforms. It's not an irreversible decision or a huge investment. At [Bravos AI](https://bravos-ai.com/) you can create a bot, upload your content, and test it on your website in an afternoon; and if you use Claude or ChatGPT, fine-tune it from the chat itself with [our MCP connector](https://bravos-ai.com/conector-ia-mcp), which applies our judgment to tell you what to improve. If it works, great. If not, you cancel before the first charge and pay nothing. What does have a real cost is having nothing at all — or sticking with a button chatbot that can't handle what your customers actually need while your competitors are already serving clients in 10 languages at 3 in the morning. 85% of customer service leaders are already exploring conversational AI (Gartner, 2024). It's not a trend. It's because it works. ## Frequently asked questions ### Can an AI chatbot fully replace a rule-based chatbot? In most cases, yes. A conversational AI chatbot can do everything a scripted chatbot does (answer predefined questions, guide the user) and also understands natural language, searches your data, and maintains context. The exception would be a very specific flow where you need total control of the path (for example, a step-by-step form with strict validations). ### Do AI chatbots make up answers? They can, if poorly configured. Hallucinations happen when the model generates text without a basis in your data. A well-built chatbot minimises this by only answering based on your content, filtering unreliable results before generating a response, and using a [well-written system prompt](https://bravos-ai.com/blog/system-prompt-chatbot-empresa) that tells it what to do when it doesn't know something. The risk is never zero, but with proper setup it's very low. ### Which chatbot type is best for my business? If your business has fewer than 15 FAQs, no product catalogue, operates in a single language, and wants a simple flow like "book an appointment," a rule-based chatbot is enough. For everything else — catalogues, extensive documentation, international customers, complex conversations — an AI chatbot will deliver better results. For small businesses specifically, see our [complete chatbot guide for SMBs](https://bravos-ai.com/blog/chatbot-para-pymes) . ### Is there a hybrid chatbot that combines rules and AI? Yes. Some platforms let you define guided flows for specific cases (like booking an appointment) and use AI for open-ended questions. In practice, the market trend is moving towards pure AI with proper guardrails, because maintaining both systems adds complexity without a clear benefit. ### How long does it take to set up each type? A simple rule-based chatbot can be set up in a few hours. A SaaS AI chatbot like [Bravos AI](https://bravos-ai.com/) can be configured in under 30 minutes: create the bot, upload your content (website, documents, catalogue), paste the script on your site, and you're live. The most time-consuming part is preparing your content well, not the tool itself. ### Ready to try an AI chatbot? Create your chatbot for free on Bravos AI. Upload your content, customise the tone, and test it on your website in under 5 minutes. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Multilingual AI Chatbot: 13 Languages Without Translation **Category:** Practical guide | **Read time:** 12 min | **Date:** Mar 8, 2026 | **URL:** https://bravos-ai.com/en/blog/multilingual-chatbot Your website is in five languages. Your chatbot speaks one. A real multilingual chatbot solves this — but most platforms force you to duplicate content, maintain separate bots per language, or manually activate each one. The result: a carefully localized website with a monolingual chatbot. In this article we look at why language directly impacts sales, what the major chatbot platforms actually offer for multilingual support, and how [Bravos AI](https://bravos-ai.com/) handles it differently — with automatic URL-based detection and a single knowledge base (the documents, FAQs, and pages that feed the bot) for 13 languages. ## 40% of buyers won't purchase if the website isn't in their language CSA Research's *Can't Read, Won't Buy* study — 8,709 consumers across 29 countries — remains the most cited research in the localization industry. Here's what it found: - **76%** of online shoppers prefer to buy in their native language - **40%** will not buy from websites that aren't in their language - **60%** rarely or never buy from English-only websites - **75%** are more likely to repurchase when post-sales support is in their native language - Fortune 500 companies that invest in localization are **1.5x more likely** to see revenue growth In Europe the picture is even starker. The EU has 24 official languages. According to the 2024 Eurobarometer survey (26,523 respondents across 27 member states), only half of Europeans speak English and **32% speak no foreign language at all**. If your chatbot only answers in English, you're cutting off a huge segment of your addressable market. ## Why hiring multilingual staff doesn't scale **Multilingual customer support teams.** If you have the budget for a large team, this can work. For a small business with 2–5 people, it's not viable. And it doesn't cover nights or weekends — which is exactly when international tourists and overseas buyers are browsing. **Translating all your content.** If your website has 50 pages, translating into 5 languages means 250 pages. Every price change, every updated policy, gets replicated five times. Maintenance costs grow exponentially. **One chatbot per language.** Some platforms force you to do this. Three bots means three times the configuration, content, and upkeep. You change an opening hour and have to update it in three places. ## How a multilingual AI chatbot detects language automatically A multilingual chatbot — sometimes searched as a "chatbot in different languages" — is a single bot, with a single knowledge base, that answers in the visitor's language. You don't need to translate anything: you feed it your information in one language and the AI handles the rest — responding in German, French, Russian, or whatever language the visitor writes in. The key is how it detects the language. There are three approaches: **Message content detection.** The chatbot analyses what the user types. This is the most common method, but it fails with short messages ("Hi", "OK", "Price?") that could belong to multiple languages. **Browser language detection.** The bot reads the Accept-Language header. It's instant, but many users have their browser set to English even though their native language is different. **URL-based language detection.** If the URL contains /en/ , the visitor speaks English; /de/ means German; /fr/ means French. It also works with subdomains ( en.yoursite.com ) or separate domains ( yoursite.co.uk ). This is the most reliable method because it reflects an explicit decision: "I'm browsing in this language." The ideal setup combines URL and content detection: the URL sets the initial language (greeting, placeholder, bot name) and if the visitor switches language mid-conversation, the chatbot adapts automatically by detecting message content. ## What the major chatbot platforms actually offer Most major multilingual chatbots claim to support "multiple languages." But when you read the technical documentation, the fine print matters: **Tidio (Lyro):** 48 languages with automatic detection. Solid multilingual support overall. No per-language widget customisation (greeting, placeholder, bot name) and no URL-based detection. **Intercom (Fin):** 45+ languages with real-time translation. Their documentation states that knowledge base content **is not automatically translated** — if you want documentation in German, you create it yourself in German. No URL-based detection. **Zendesk:** Multilingual bot with automatic detection. The knowledge base must be provided separately in each language. No URL-based detection. **Chatbase:** 80+ languages. When you update content, it requires manual retraining. Multilingual knowledge bases can exceed character limits on some plans. **Crisp:** Widget UI in 60+ languages. Chatbot flows require separate logic branches per language — you build them manually. They all handle conversational translation well. The differences are in the details: none of them offer URL-based language detection or full per-language widget customisation out of the box. And one warning on the more recent visual RAG proposals: if you've seen the headlines around [PixelRAG](https://bravos-ai.com/blog/pixelrag-chatbot-empresarial) — which renders documents as images instead of text — keep in mind it is **only validated in English**. The authors themselves acknowledge the language bias in the paper's limitations appendix. For a real multilingual chatbot, text-based RAG is still what does the job today. ## How Bravos AI handles multiple languages with one knowledge base [Bravos AI](https://bravos-ai.com/) combines URL-based and message content detection, and adds something none of the platforms above offer: **full per-language customisation**. **URL-based detection.** You set up language rules based on your website URL: path prefix ( /en/ , /de/ ), subdomain ( en.yoursite.com ), or domain ( yoursite.co.uk ). When a visitor opens the chat on the German version of your site, the chatbot greets in German because you configured it that way. **Content detection.** If the visitor switches language mid-conversation, the bot adapts automatically. No additional setup. **Per-language customisation.** For each language you configure the bot name, welcome message, placeholder, and tooltip. A German visitor sees "Fragen Sie mich..."; a French visitor sees "Posez-moi une question...". All from a single panel —or, if you use Claude or ChatGPT, from the chat itself with [our MCP connector](https://bravos-ai.com/conector-ia-mcp). **One knowledge base, every language.** You upload your content in one language. The chatbot translates responses to the visitor's language. You update a price once and it applies across all languages. No manual retraining, no content duplication. This is how a RAG-based multilingual conversational AI works: it retrieves information from your documents and generates the answer in whatever language the customer needs. It supports **13 languages** out of the box: Spanish, English, German, French, Italian, Portuguese, Dutch, Russian, Arabic, Polish, Chinese, Japanese, and Korean. Available from the Starter plan (€19/month). You can test it with 7 days free on the full PRO plan. [Full pricing comparison with other platforms](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa) . ## Where a multilingual chatbot makes the biggest difference **Tourism and hospitality.** Spain received 94 million international tourists in 2024 (INE). The Costa del Sol alone welcomed 13.8 million. Top source markets: the UK (17 million), France (12.9 million), Germany (11.9 million). A hotel or rental company that only responds in Spanish leaves most potential guests without answers. **International real estate.** In areas like [Marbella and the Costa del Sol](https://bravos-ai.com/blog/chatbot-marbella) , 70% of buyers are foreign nationals. They browse listings at 10 PM from Amsterdam, Warsaw, or Munich. A chatbot that filters your catalogue and replies in Dutch, Polish, or German without human intervention closes deals while the office is shut. If you use Resales Online as your CRM, [there's a direct integration with automatic daily sync](https://bravos-ai.com/blog/chatbot-inmobiliaria-resales-online) . **Cross-border ecommerce.** European cross-border ecommerce reached €275.6 billion in 2024. Language remains one of the primary purchase barriers. If you sell to Germany, France, and Poland from Spain, a multilingual chatbot that handles customer questions in those languages reduces friction at every step. **Clinics and training centres.** Health tourism (dental, cosmetic) attracts patients from across Europe. Language academies serve Erasmus students. In both cases, the chatbot answers about pricing, treatments, and availability in whatever language the visitor writes in. ## How to set up a multilingual chatbot in 10 minutes **1.** Create your bot at [app.bravos-ai.com](https://app.bravos-ai.com/register?lang=en) and upload your content (website, documents, product catalogue). In whichever language you prefer. **2.** Open your bot and go to the **Multilingual website** tab. Add the languages you need. For each one, set the URL detection rule and customise the greeting, placeholder, and bot name. **3.** Install the widget on your site with a single line of code. If you use WordPress, [there's a step-by-step tutorial](https://bravos-ai.com/blog/chatbot-wordpress) . And if your site is on Wix, there's a [Wix chatbot](https://bravos-ai.com/blog/chatbot-wix) guide. For other platforms like React, Shopify, or static HTML, you'll find installation instructions inside the app. When a visitor lands on yoursite.com/en/ , they see the chatbot in English. On yoursite.com/de/ , in German. If there's no language prefix, the chatbot greets in the bot's default language, but responses adapt to whatever language the visitor writes in. ## Frequently asked questions ### Do I need to translate my content for a multilingual chatbot? No. A RAG-based multilingual chatbot — like Bravos AI — retrieves information from your documents in whatever language you uploaded them in, then generates the answer in the visitor's language. You write your FAQ, product catalogue, or help docs once. The AI handles the translation at response time. ### Can a chatbot detect the visitor's language automatically? Yes, primarily through two methods: analysing the message text or detecting the language from the URL structure of your website. URL-based detection is the most reliable because it reflects an explicit user choice. The best multilingual chatbots combine both. ### Can a chatbot switch languages mid-conversation? Yes. If a visitor writes in French, the chatbot detects the language and responds in French. This is especially useful if your website isn't translated into many languages — even if you only have English and Spanish versions, the chatbot can respond in whatever language the visitor uses. ### How many languages can an AI chatbot support? Modern LLM-based chatbots can technically respond in any language the underlying model supports (GPT-4 covers 50+). In practice, accuracy varies. Bravos AI supports 13 languages with full per-language widget customisation. For most businesses serving European or international customers, this covers well over 95% of visitors. ### What is the difference between a translated chatbot and a multilingual chatbot? A translated chatbot requires you to provide content separately in each language — you maintain parallel knowledge bases, FAQ pages, or conversation flows. A truly multilingual chatbot uses a single knowledge base and translates dynamically at response time. The difference matters: with a translated chatbot, updating one piece of information means updating it in every language. With a multilingual chatbot, you update once. ### Do your customers speak more than one language? Create a multilingual chatbot for free. One knowledge base, 13 languages, no translation needed. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # AI Chatbot for Your Marbella Business: Skip the Agency, Build It Yourself **Category:** Practical guide | **Read time:** 8 min | **Date:** Mar 4, 2026 | **URL:** https://bravos-ai.com/en/blog/marbella-chatbot Google "chatbot Marbella" and you get a wall of agencies. Custom solutions, bespoke development, AI consulting. None of them publish pricing. If you run a business on the Costa del Sol, you already know the challenge: clients arrive from Germany, the UK, the Netherlands, Scandinavia. They browse your website at 10 PM their time, have questions in their own language, and expect instant answers. You don't need a six-month agency project for that. You need a tool you can set up in an afternoon. ## When You Need an Agency (and When You Don't) Agencies make sense for genuinely complex projects: integrating with internal systems, building features that don't exist yet, strategic consulting. That's their value, and that's why they charge what they charge. But if you need a chatbot that answers questions about your business, filters your property listings or product catalog, and captures leads when you're offline — that's not a custom project. That's a platform you sign up for. Today there are platforms where you create an AI chatbot in a single afternoon. No coding, no meetings, no waiting for proposals. You feed it your content — your website URL, PDFs, a spreadsheet with your catalog — and the AI learns to answer questions based on that information. These aren't rule-based bots where you program "if X, then Y." This is AI that understands context, adapts to how each person phrases their question, and says "I don't know" instead of making things up. Pricing is fixed and transparent. You know the cost before you start. No lock-in contracts, no dependencies. [Bravos AI](https://bravos-ai.com/) is one of these platforms, and it has specific features that matter when your business serves an international clientele. ## Multilingual by Default: 13 Languages, Zero Configuration On the Costa del Sol, Spanish alone won't cut it. Your customers are German retirees, British expats, French holidaymakers, Dutch investors, Scandinavian second-home buyers. A chatbot that only speaks English or Spanish is leaving money on the table. Bravos AI detects the visitor's language and responds in kind. Automatically. It [supports 13 languages out of the box](https://bravos-ai.com/blog/chatbot-multiidioma) : Spanish, English, German, French, Italian, Portuguese, Dutch, Russian, Arabic, Polish, Chinese, Japanese, and Korean. No configuration needed. For a real estate agency in Marbella where 70% of buyers are foreign nationals, or a beachfront restaurant serving tourists from half of Europe, multilingual support isn't a nice-to-have — it's the baseline. ## Smart Catalog Search (Not Keyword Guessing) If your business has a catalog — properties, products, services — the chatbot filters it with precision. It doesn't guess based on similar words. It understands that "under €300,000" is a number and filters accordingly. "Apartments in Nueva Andalucía, 2 bedrooms, pool, under €400,000" — the chatbot returns exactly the listings that match every criterion. If nothing matches, it says so. No hallucinations, no "you might also like." This is where [most AI chatbots fall short](https://bravos-ai.com/blog/chatbot-catalogo-productos) . They treat catalogs as unstructured text instead of queryable data. ## Lead Capture That Works at 11 PM on a Sunday When a visitor shows genuine interest, the chatbot offers to collect their details. You get an email with their name, phone number, email — and the full conversation. Not a cold form submission with zero context. Picture this: Sunday night in August. A German buyer is browsing properties from Munich. Your office is closed. The chatbot responds in German, filters listings to match their criteria, and sends you the lead with full context. Monday morning, you call them before any competitor does. For example, [AlquiConfort](https://alquiconfort.com) — an HVAC equipment company — was receiving technical inquiries via [WhatsApp](https://bravos-ai.com/blog/chatbot-whatsapp) that their team couldn't always answer quickly enough. Within a month of adding the chatbot, over 280 inquiries were handled automatically. Their staff now only gets involved when a customer is ready to buy. ## Which Businesses on the Costa del Sol Benefit Most Any business in Marbella or the Costa del Sol that handles repetitive inquiries — especially across time zones and languages — will see immediate value. **Real estate agencies.** A buyer in Munich searches at 9 PM for apartments in Nueva Andalucía — 2 beds, pool, under €400K. The chatbot filters your listings, shows what fits, and captures their details. If you use [Resales Online](https://bravos-ai.com/blog/chatbot-inmobiliaria-resales-online) , there's a direct integration with automatic daily sync. **Hotels and holiday rentals.** Parking, pool, check-in times, airport transfers — the same questions, all day, every day. The chatbot handles them in the guest's language, freeing your team for what actually matters. **Restaurants and beach clubs.** Menu, dietary options, group bookings — answered in the tourist's language, around the clock. **Dental and aesthetic clinics.** Treatment details, pricing, availability. The chatbot informs and captures the lead. **Car hire and dealerships.** Vehicle type, pricing, features — filtered instantly. **Language schools and training centres.** Courses, timetables, fees, requirements — particularly useful with international students. ## Transparent Pricing (No Surprises) You get a 7-day PRO trial to see it working with your own content. Plans start at €19/month and include automatic data sync, support, and everything described above. [See how it compares to other platforms](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa) ## How to Get Started Sign up at [app.bravos-ai.com](https://app.bravos-ai.com/register?lang=en) , upload your content (website, documents, catalog), set your brand colours, and install it on your site with a single line of code. If you use WordPress, [there's a step-by-step guide](https://bravos-ai.com/blog/chatbot-wordpress) . If you use any other platform (HTML, React, Shopify, Framer...) the instructions are built into the app. One afternoon. Thirteen languages. Answers based on your actual business information. No developer required. ### Running a business in Marbella or the Costa del Sol? Create your multilingual AI chatbot for free and start handling inquiries 24/7. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Real Estate Chatbot: Connect Resales Online with AI to Handle Client Inquiries 24/7 **Category:** Use case | **Read time:** 10 min | **Date:** Mar 2, 2026 | **URL:** https://bravos-ai.com/en/blog/real-estate-chatbot-resales-online We just shipped our first CRM integration: **Resales Online**. If you run a real estate agency on the Costa del Sol, you can now connect your property catalog to an AI chatbot in minutes — and have an assistant on your website that answers inquiries, filters listings, and captures leads around the clock. In this article, we'll walk you through how the integration works, what data gets synced, and how it can change the way your agency handles international clients. ## Why We Started with Resales Online [Resales Online](https://www.resales-online.com) is **the most widely used CRM for real estate agencies on the Costa del Sol**. It centralizes property catalogs, manages inter-agency collaborations, and publishes to portals like Idealista, Kyero, and Rightmove. If you work in real estate in southern Spain, chances are you're already using it. That was exactly the point: we wanted our first integration to solve a real problem for a specific industry. Real estate agencies on the Costa del Sol face a unique challenge: **international clients browsing from different countries and time zones**. A German buyer who sits down at 9pm to browse properties on their laptop is active, comparing, sending inquiries to multiple agencies at once. If your website doesn't respond in that moment, the lead goes to another agency. In real estate, where the average sale is hundreds of thousands of euros, every unanswered lead is real money lost. ## How the Integration Works The concept is straightforward: **connect Resales Online to Bravos AI, and your chatbot has access to your up-to-date property catalog**. No manual CSV exports, no keeping anything in sync yourself — the system handles it automatically every day. Setup process: - Create a **new API key in Resales Online** for Bravos AI — each key is bound to an IP, so you need a specific one for the integration. The platform guides you through it - Enter your **credentials** (P1 and P2) in the Bravos AI dashboard - Select which **filters** to sync: **Sale, Short term rental, Long term rental, and/or Featured** — configured to match the ones you already have on your website (from your other API key) - Select the **data language** (English or Spanish) — this determines which language the property descriptions are imported in. The chatbot responds in any language on the fly, regardless of this setting - Set up **URL templates** so each property links to your website: using your **AgencyRef** for your own listings and the **Reference** for shared listings from the network - Click **"Sync"** — your listings are imported as structured data From that point on, sync is automatic. Every day the system updates your catalog. If you add, modify, or remove properties in Resales, the chatbot reflects it. Resales Online integration in Bravos AI — example from [Garcia Navarro Real Estate](https://www.garcianavarro.pro) } /> ### What Data Gets Synced This isn't just text. It's **structured data** that the chatbot can filter with precision: - **Reference** number - **Type:** apartment, villa, townhouse, penthouse, country house... - **Area / location:** Marbella, Estepona, Benahavís, Nueva Andalucía... - **Bedrooms and bathrooms** - **Built area and plot size** - **Features:** pool, parking, terrace, sea views, garden - **Price** - **Full description** (in your chosen language) - **Main image** Because the data is structured, the chatbot doesn't "search text" — it **filters with exact criteria**. If someone asks "apartments in Estepona, 2 bedrooms, pool, under €400,000", the chatbot returns exactly the properties that match. For more on how this works technically, see our article on [chatbots for product catalogs](https://bravos-ai.com/blog/chatbot-catalogo-productos). ## A Real Example: How It Looks on Your Website This is what a visitor would see on your website with the chatbot active. A prospect asks about apartments for rent and the chatbot responds with real properties from your catalog: In this example, the chatbot did three things: - **Filtered the catalog** with the visitor's exact criteria (rental, Marbella/Estepona, 2 bed, pool) - **Showed real properties** with photo, details, and a direct link to your website - **Offered the contact form** to capture the lead with name, email, and phone You receive an email notification with the lead and the full conversation: you know what they're looking for, which properties they're interested in, and what language they speak. We also offer the option to set up a **webhook** (though you may need your developer for this) to send each lead automatically to your CRM or any tool you use. When you call, you go straight to the point. ## Multilingual: Your Clients Speak, the Chatbot Responds The Costa del Sol attracts buyers from all over the world. Germans, Brits, Dutch, Swedes, French, Russians, Arabs, Polish... Each expects to communicate in their own language. **The chatbot automatically detects the visitor's language and responds accordingly**. Property descriptions from Resales are imported in English or Spanish, but that doesn't limit the chatbot. The AI model understands and responds in [any language](https://bravos-ai.com/blog/chatbot-multiidioma) — German, French, Dutch, Swedish — using the property data to build the response in the visitor's language. Russian, Arabic, Polish... all supported. ## What Plan Do You Need? The Resales Online integration requires the **Starter plan (€19/month) or higher**. Starter includes unlimited messages and supports automatic CSV syncs. Each filter you enable (Sale, Short term rental, etc.) counts as 1 synced CSV. If you want to test before committing, you get **7 days free on the full PRO plan** with the Resales Online integration included. You sync your properties, see how the chatbot handles your real catalog, and decide afterwards whether it's worth keeping. For larger agencies with thousands of properties or multiple chatbots, the Pro and Enterprise plans offer more capacity and priority support. ## Resales Online + AI: A Combination That Made Sense Resales Online has been the backbone of real estate agencies on the Costa del Sol for years. It has a solid API, a collaboration network between agencies that works, and an ecosystem of connected portals. Your properties are already on your website. What was missing was someone to handle them: answering inquiries outside business hours, in the client's language, with the exact data from your catalog. That's what this integration does. It doesn't replace anything you already have — it complements it. Your Resales catalog, always up to date, with a chatbot that knows it inside out and serves every visitor as if they were the only one. ### Got property listings in Resales Online? Connect your CRM, let the chatbot handle inquiries 24/7, and receive qualified leads by email. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Chatbot for Ad Campaigns: How to Stop Wasting Money on Google Ads and Meta Ads **Category:** Analysis | **Read time:** 8 min | **Date:** Feb 22, 2026 | **URL:** https://bravos-ai.com/en/blog/chatbot-ad-campaigns You pay for every click on Google Ads. You pay for every impression on Meta Ads. The ads work — people land on your website. But they arrive at 10 PM, on a Sunday, or during your lunch break. Nobody's there. They fill out a form — maybe — and leave. That click you paid for? Wasted. A chatbot fixes this. It responds instantly, around the clock, qualifies visitors, and notifies you with full context. In this article, we explain why it's critical if you run paid ads — and how to implement it. ## The Problem: You Pay for 24/7 Traffic, But You're Only Available 8 Hours Ad campaigns don't sleep. Google Ads and Meta Ads show your ads at any hour, any day. And people click when it's convenient for them — not when it's convenient for you. According to the [Zendesk CX Trends Report 2025](https://www.zendesk.com/newsroom/articles/2025-cx-trends-report/) , **74% of consumers expect customer service to be available 24/7**. And **88% expect faster responses than they did a year ago**. But reality tells a different story. Most businesses take **hours or days** to respond to a contact form submission. By then, the potential customer has found another option — or simply lost interest. ## What You Lose While You Wait The [MIT and InsideSales.com study](https://cdn2.hubspot.net/hub/25649/file-13535879-pdf/docs/mit_study.pdf) is the most cited research on this topic — and for good reason. They analyzed **15,000 leads and 100,000 call attempts over 3 years**. The results: - Contacting someone within **5 minutes vs 30 minutes** = **100 times** more likely to reach them - Qualifying them within **5 minutes vs 30 minutes** = **21 times** more likely to convert into a real opportunity [Harvard Business Review](https://hbr.org/2011/03/the-short-life-of-online-sales-leads) published similar findings in "The Short Life of Online Sales Leads": - Contacting within **1 hour** = **7 times** more likely to qualify - Waiting **24 hours or more** = **60 times less** likely You don't lose a little effectiveness. The probability of closing that sale collapses. And meanwhile, your competition — if they respond first — takes the sale. According to the same studies, **35% to 50% of sales go to whoever responds first**. ## The Form Is Not the Solution "But I have a contact form." Yes, and according to a study by [The Manifest](https://themanifest.com/web-design/6-steps-avoiding-online-form-abandonment) , **81% of people who start filling out a form abandon it before submitting**. Forms have several problems: - **They don't answer questions** — the visitor has doubts, the form doesn't resolve them - **Friction** — required fields, captchas, error messages - **Zero immediacy** — the visitor knows they won't get a response now A form is a mailbox. You leave your message and wait. What you need is a store clerk — someone who responds in the moment, answers questions, and never closes. That's exactly what a chatbot does. ## Chatbot on Your Landing Page: Instant Response, 24 Hours, 7 Days A chatbot on your landing page does exactly what a form cannot: **Responds instantly.** The visitor asks "Do you ship to Alaska?" and gets an answer in 2 seconds. Not tomorrow. Now. **Qualifies while conversing.** Instead of a cold form, the chatbot asks questions: what they need, their budget, their timeline. When you get the contact, you already know if it's a real opportunity. **Captures data with context.** You don't just receive a name and email. You receive the full conversation: what they asked, what interests them, what doubts they have. You can respond with relevant information, not with a generic "thanks for reaching out." **Works when you can't.** At 11 PM, on a Sunday, in August. The chatbot is there. According to [Gartner](https://www.gartner.com/en/newsroom/press-releases/2024-12-09-gartner-survey-reveals-85-percent-of-customer-service-leaders-will-explore-or-pilot-customer-facing-conversational-genai-in-2025) , **85% of customer service leaders are already exploring or implementing conversational AI**. And they [predict](https://www.gartner.com/en/newsroom/press-releases/2025-08-27-gartner-survey-finds-self-service-and-live-chat-will-surpass-traditional-channels-as-top-customer-service-technologies-by-2027) that live chat and self-service will surpass traditional channels (phone, email) as the primary customer service technologies by 2027. This isn't a future trend. It's happening now. ## The Impact on Your Ad Campaigns [Forrester](https://www.forrester.com/report/Market+Overview+Chat+Solutions+For+Customer+Service/-/E-RES92941) found that live chat increases **average order value by 10%**. And the cost per contact is radically different: **$8.01 for assisted channels vs $0.10 for self-service**. But beyond generic numbers, think about your specific case: ### Example with real numbers - You invest **$1,200/month** in Google Ads - Average CPC: **$2.40** → 500 clicks per month - Landing page conversion: **5%** → 25 contacts - But **40%** of those clicks arrive outside your business hours (200 clicks) - Of those 200, many don't even fill out the form because there's no immediate response - Those who do fill it out — by the time you respond the next day, they're cold or have bought from someone else **With a chatbot that responds 24/7:** - Those 200 after-hours clicks get instant responses - Visitors resolve their questions in the moment and leave their details - You receive the contact with full conversation context - When you call the next morning, the contact is still warm — and you know exactly what they need If the chatbot helps just half of those 200 clicks convert at the same 5% rate, that's **5 extra contacts per month**. Contacts that were previously lost — and that you already paid for. And this doesn't count how the chatbot can improve conversion during business hours too, because it responds faster than any human. ## What Your Chatbot Needs for Ad Campaigns Not just any chatbot will do. To work in the context of paid advertising, you need: ### Responses based on YOUR information The chatbot must answer about your business, your products, your prices. Not with generic AI responses that [make things up](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo) . Look for platforms that use **RAG** (Retrieval-Augmented Generation): the chatbot searches your documents and responds only with what you've given it. With [Bravos AI](https://bravos-ai.com/) , you upload your website or documents and the chatbot trains automatically. ### Data capture with instant notification It's useless if the chatbot captures a contact at 3 AM but you don't find out until Monday. You need **real-time email notifications** with the complete conversation context. With Bravos AI, every time someone leaves their details, you get an email with the full conversation. ### Catalog filtering (if applicable) If you sell products with filterable attributes — price, category, location — the chatbot should filter exactly what the customer asks for. "Apartments in Miami under $300,000 with 2 bedrooms" → exact results, not approximations. Most chatbots can't do this — they treat your catalog as text and give approximate answers. [With Bravos AI we do](https://bravos-ai.com/blog/chatbot-catalogo-productos) : upload a CSV and the chatbot filters with precision. ### Simple integration One line of code to install. If your landing page is on WordPress, Webflow, or any CMS, it has to be plug-and-play. With Bravos AI you copy a script and you're done — there's also a [WordPress tutorial](https://bravos-ai.com/blog/chatbot-wordpress). ## How to Get Started with Bravos AI With [Bravos AI](https://bravos-ai.com/) you can create a chatbot for free and have it running on your landing page in under 10 minutes: - **Create your account** at [app.bravos-ai.com](https://app.bravos-ai.com/register?lang=en) and start the 7-day PRO trial - **Upload your content** — paste your website URL and the system extracts information automatically - **Customize** — your brand colors, welcome message - **Install** — one line of code on your landing page - **Enable notifications** — receive contacts by email with full context You get 7 days free on the full PRO plan to test it in real conditions. After that, plans start at €19/month. ## Conclusion Every dollar you invest in Google Ads or Meta Ads buys attention. Attention from people who have a problem you can solve. But that attention lasts seconds. If there's nobody to help them — if they have to fill out a form and wait hours or days — they leave. A chatbot doesn't replace your sales team. What it does is ensure that when someone lands on your website — at any hour — there's someone who responds. Who qualifies. Who captures. And who notifies you so you can close the sale. MIT proved that responding in 5 minutes multiplies your chances of closing a sale by 21. A chatbot responds in 2 seconds. With [Bravos AI](https://bravos-ai.com/) you can create a chatbot for free, train it with your website and documents, and test it in under 10 minutes. If you have a product catalog, you can upload a CSV and the chatbot will automatically filter by price, category, or whatever you need — no code, no APIs. [Try it free for 7 days](https://app.bravos-ai.com/register?lang=en). ### Running Google Ads or Meta Ads? Stop losing clicks you already paid for. Create your chatbot for free in under 10 minutes. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Chatbot for Small Business: The Complete 2026 Guide to Automating Customer Service **Category:** Practical guide | **Read time:** 10 min | **Date:** Feb 19, 2026 | **URL:** https://bravos-ai.com/en/blog/chatbot-for-small-business You run a small business. You get inquiries by email, through your website form, on social media. The same questions over and over: business hours, pricing, availability, how your service works. You or someone on your team spends hours every week answering the same things. What if you could automate most of those inquiries without hiring anyone? According to a McKinsey case study, businesses that implement chatbots properly reduce [service interactions by 40-50%](https://www.mckinsey.com/capabilities/operations/our-insights/the-next-frontier-of-customer-engagement-ai-enabled-customer-service) . And 91% of small businesses already using AI say it increases their revenue. In this guide, we'll show you how to do it right — no jargon, no fluff, just the numbers. ## Why Small Businesses Are Adopting Chatbots (And Why You Should Consider It) Enterprise companies have been using chatbots for years. But in 2026, the technology has matured enough to make economic sense for small businesses. ### The Numbers - **40-50%** reduction in customer service interactions with well-implemented chatbots ( [McKinsey case study](https://www.mckinsey.com/capabilities/operations/our-insights/the-next-frontier-of-customer-engagement-ai-enabled-customer-service) ) - **20%+** reduction in cost-to-serve ( [McKinsey](https://www.mckinsey.com/capabilities/operations/our-insights/the-next-frontier-of-customer-engagement-ai-enabled-customer-service) ) - **75%** of SMBs are already experimenting with AI, and **91%** of those using it say it increases revenue ( [Salesforce](https://www.salesforce.com/news/stories/smbs-ai-trends-2025/) ) - **$80 billion** estimated savings in contact center costs from conversational AI by 2026 ( [Gartner](https://www.gartner.com/en/newsroom/press-releases/2022-08-31-gartner-predicts-conversational-ai-will-reduce-contac) ) ### What's Changed in 2026 A few years ago, setting up a chatbot required developers, complex integrations, and five-figure budgets. Today you can have one running on your website in under an hour, trained on your own information. Plans start at €19/month with a 7-day PRO trial if you want to test the performance before paying. The difference is **RAG** technology (Retrieval-Augmented Generation): the chatbot doesn't [make up answers](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo) — it searches your documents, your website, your FAQs, and responds only with what you've given it. If it doesn't know something, it says so. ## What a Chatbot Actually Does for a Small Business **Answers the same questions you're tired of answering** Business hours, pricing, availability, how your service works, return policy. The chatbot handles them in seconds, 24/7, without you or your team lifting a finger. **Captures potential customers while you sleep** Someone visits your website at 11 PM. The chatbot engages them, asks questions, understands what they need. The next morning you have an email with their contact info and the full conversation context. Without a chatbot, that potential customer would be gone. **Filters products from your catalog** If you have a catalog with prices, categories, features — the chatbot can filter exactly what the customer asks for. "Laptops under €600 with 16GB RAM" → exact results, not approximations. With [Bravos AI](https://bravos-ai.com/) you upload a CSV and it works automatically, no code required. **Stays updated with your data** Does your inventory change? Do your prices update? If you export your data to CSV or Google Sheets, the chatbot syncs automatically. No retraining or reconfiguring every time something changes. And if you use Claude or ChatGPT, with [our MCP connector](https://bravos-ai.com/conector-ia-mcp) the assistant you already use reviews the bot with our judgment inside and proposes what to improve, not just runs what you ask. ## What Does a Small Business Chatbot Actually Cost? It depends on the platform, how they charge (per message, per conversation, per resolution), and what features you need. With Bravos AI you can try the full PRO plan free for 7 days and plans start at €19/month. Other platforms have similar base prices, but then charge extra for every query the AI resolves — and that's where costs explode. We've done a [complete pricing comparison](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa) with Tidio, Intercom, Crisp, Zendesk, and Chatbase. Real numbers, verified, explaining the traps in each pricing model. ### The Calculation That Matters: Is It Worth It? Picture this typical situation: - Your team spends **20 hours/week** answering repetitive inquiries - Cost of that time: **~€900/month** (at €12/hour with overhead) - A chatbot resolves **50%** of those inquiries automatically (conservative based on McKinsey's documented 40-50% reduction) - Savings: **€450/month** in recovered time - Chatbot cost: **€49/month** - **Net savings: €401/month from day one** And that's not counting the potential customers you capture outside business hours — who today simply leave. ## How to Choose the Right Chatbot for Your Small Business Not all chatbots are equal. These are the questions you should ask before choosing: ### Can I train it with my own data without coding? The chatbot needs to answer about YOUR business, not give generic responses. Look for platforms that let you upload your website, PDFs, documents, without needing a developer. ### How does it handle my product catalog? If you have an Excel or CSV with products, prices, features — ask how it manages that. Most chatbots treat your catalog as text and give approximate answers. If a customer asks "Language courses in London under €200/month", you need the chatbot to filter exactly that, not give you "something similar". [We explain this problem in detail here](https://bravos-ai.com/blog/chatbot-catalogo-productos). ### What happens when it doesn't know the answer? A good chatbot says "I don't know" and offers to collect contact information. A bad chatbot [makes things up](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo) . Ask specifically what it does when it can't find relevant information. ### How does it notify me of new contacts? If the chatbot captures someone's details at 3 AM but you don't find out until three days later, you've lost the opportunity. Look for real-time email notifications. ### Is it GDPR compliant? If you operate in Europe or serve European customers, the chatbot must comply with data protection regulations. Ask where conversation data is stored and what privacy policies apply. ## Mistakes That Ruin Implementation ### Not giving it enough content A chatbot with three FAQs will answer "I don't know" constantly. [Prepare your content well](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo) : real frequently asked questions, product information, policies. The more you give it, the better it responds. ### Hiding it in a corner If the widget is a tiny icon nobody sees, it won't solve anything. Make it visible. Make it stand out. Make sure customers see it as soon as they land on your site. ### Not reviewing conversations The first few weeks, review the conversations. You'll see which questions it can't answer, what information you're missing. It's the fastest way to improve the chatbot. And if you'd rather not do it by hand, by [connecting your chatbot to Claude or ChatGPT](https://bravos-ai.com/blog/que-es-un-mcp) the AI itself reviews those conversations and points out what's failing. ## How to Get Started with Bravos AI ### 1. Create your account and start the PRO trial Go to [app.bravos-ai.com](https://app.bravos-ai.com/register?lang=en) and create your account. You start with 7 days free on the full PRO plan. ### 2. Upload your content - **Your website:** Paste the URL and the system extracts the content automatically - **Documents:** Upload PDFs with your business information - **Catalog:** If you have an Excel or CSV with products, upload it. The chatbot will be able to filter by price, category, whatever you need ### 3. Customize the widget Set your brand colors and welcome message. The widget appears on all pages of your website. ### 4. Install on your website One line of code. If you use WordPress, [there's a step-by-step tutorial](https://bravos-ai.com/blog/chatbot-wordpress); and if your site is built with Wix, there's a [Wix chatbot](https://bravos-ai.com/blog/chatbot-wix) guide too. ### 5. Enable the contact form Configure the chatbot's contact form so you receive the customer's details + the full conversation context by email. In 10 minutes you have the chatbot running. The first few weeks, review conversations and add content to fill the gaps. ## Conclusion 20 hours a week answering the same questions. Potential customers visiting your website at 10 PM and finding no one. Catalog inquiries that require digging through a spreadsheet. All of that can be automated today, with technology that didn't exist at this price three years ago. The key is choosing well: a platform you can train with your data, that doesn't make things up, and that has pricing that makes sense for your volume. With [Bravos AI](https://bravos-ai.com/) you can create a chatbot, train it with your website and documents, and test it in under 10 minutes. If you have a product catalog, you can upload a CSV and the chatbot will automatically filter by price, category, or whatever you need — no code, no APIs. [Start with 7 days free on the PRO plan](https://app.bravos-ai.com/register?lang=en). ### Ready to automate your customer service? Set up your AI chatbot in under 10 minutes. No code needed. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Why Your AI Chatbot Can\ **Category:** Analysis | **Read time:** 10 min | **Date:** Feb 11, 2026 | **URL:** https://bravos-ai.com/en/blog/ai-chatbot-product-catalog You've installed an AI chatbot on your website. You've uploaded your product catalog. A customer asks: "Do you have running shoes under $100 in size 10?" And the chatbot responds with something vaguely related — maybe shoes for $150, maybe size 8, maybe not even running shoes. Sound familiar? You're not doing anything wrong. The problem is the technology your chatbot uses under the hood. Most AI chatbots on the market — including those from well-known platforms like Intercom, Zendesk, Tidio, and Chatbase — use a technology called **RAG** (Retrieval-Augmented Generation) that works great for text documents but **isn't designed to filter structured data**. If you need an **ecommerce chatbot** with a large catalog or a **real estate chatbot** with price and location filters, RAG isn't enough. How that RAG works under the hood —and whether it even needs a vector database— is broken down in [vectorless RAG](https://bravos-ai.com/blog/rag-sin-vectores). In this article, we'll explain why this happens and what alternative exists. Spoiler: the solution is combining RAG with real SQL queries on your data — the best of both worlds. Almost no one does this natively — at [Bravos AI](https://bravos-ai.com/) we do. ## How an AI Chatbot Works Under the Hood (No Jargon) To understand the problem, you need to know — broadly speaking — how an AI chatbot searches for information when a customer asks a question. ### The RAG System: Searching by Meaning Most chatbots use **RAG** (Retrieval-Augmented Generation). Here's how it works: - **Chunk:** When you upload a document (a webpage, PDF, or text), the system splits it into small fragments. - **Convert to numbers:** Each fragment is transformed into a numerical representation (an "embedding") that captures its meaning. - **Search by similarity:** When a customer asks something, the chatbot converts the question into numbers and finds fragments with the most similar meaning. - **Generate response:** Using the most relevant fragments, the AI generates a natural language response. This system is brilliant for [text-based content](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). If a customer asks "what's your return policy?", RAG finds the fragment from your website about returns and responds perfectly. Same with business hours, services, FAQs, user guides. If you've seen the headlines about [PixelRAG](https://bravos-ai.com/blog/pixelrag-chatbot-empresarial) (a recent RAG variant that works with screenshots instead of text), it doesn't solve this either: it isn't designed to filter structured data, and for catalogs it ends up more expensive without gaining precision. ### Where RAG Fails: Data with Attributes Now imagine you have a shoe store with 500 products. Each has a name, brand, price, available sizes, color, type (running, casual, trail), and stock. A customer asks: "Running shoes under $100 in size 10" To answer correctly, the chatbot needs to: - Filter by type = "running" - Filter by price ≤ 100 - Filter by size that includes 10 - Return only products that meet **all three conditions at once** RAG can't do this. What it does is search for text fragments that "look similar" semantically to "running shoes $100 size 10". It might find a running shoe for $150 because the text is similar. Or one in size 10 but for hiking. Or it might mix information from multiple products into a response that sounds good but is incorrect. Chatbase's own chatbot admits this when you ask if it can filter products by price: "Approximate, not guaranteed perfect filtering." Approximate. Not guaranteed. ## The Real Problem: RAG Treats Your Catalog Like a Book When you upload a CSV with your catalog to a platform that only uses RAG, here's what happens: | What you see | What the chatbot sees | | --- | --- | | A table with columns: name, price, color, size | A long block of text | | Data filterable by attributes | Loose words with approximate meaning | | "Price: $79" as a comparable number | "Price: $79" as text that resembles other texts with numbers | It's like giving someone a spreadsheet and asking them to find data... but by reading it aloud instead of using filters. Technically they can do it, but the result is slow, imprecise, and frequently wrong. ## Does This Affect You? Not every business needs structured filtering. If your chatbot only answers FAQs ("what are your business hours?", "do you ship to Canada?"), RAG is perfect. But if you sell products with filterable attributes — price, size, color, location, category — or if your customers combine 3 or more criteria in a single question, a chatbot with only RAG will give you problems. An **ecommerce chatbot** needs to filter by price and category. A **real estate chatbot** needs to filter by location, bedrooms, and square footage. Training academies, car dealerships, consulting firms: any business with a structured **product catalog**. **How do you know if your current chatbot has this problem?** Ask it these questions: - "Show me products under [price]" — Do all results meet the filter? - "I want [product] but not [attribute]" — Does it exclude correctly? - Ask for something that does NOT exist — Does it say it doesn't have it, or does it [make something up](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo)? If it fails 2 out of 3, your chatbot is using RAG only. It's not a bug — it's a limitation of the technology. ## What the Platforms Themselves Say We've reviewed the official documentation from the leading chatbot platforms. Here's what they say about structured data and catalogs: **Chatbase:** Doesn't accept CSV. Their own chatbot recommends converting your CSV to a text document (PDF, DOCX) and uploading that. For real filtering, they say you need to build your own API. **Intercom (Fin):** Doesn't accept CSV as a knowledge source. To filter a catalog, you need to build your own REST API, your own database, and configure a Data Connector. Their documentation says it explicitly: filtering "should be handled at the API level". **Zendesk:** Accepts CSV, but as text articles (one row = one article). Not as filterable data. For real filtering you need their Integration Builder (only available on the most expensive plan) + your own API. **Tidio (Lyro):** Accepts CSV, but only as question/answer pairs (two columns). Not for catalogs. Product catalogs are only supported via Shopify integration, and only the standard data schema — if you use apps that modify product data, it won't work. **Crisp:** Accepts CSV and sections it by product for semantic search. It's the best on this list, but it's still RAG on text — no real SQL filtering. None of these platforms can execute a query like "products where price ≤ 100 AND category = running AND size includes 10" on your data. They all treat your catalog as text. ## The Solution: Combining RAG with Real Queries The problem isn't that RAG is bad. It's that RAG isn't enough for structured data. What you need is a hybrid system: RAG for text content (policies, FAQs, descriptions) and database queries for catalogs with filterable attributes. Almost no platform offers this natively. Most tell you "build your own API" or "convert your CSV to PDF". At [Bravos AI](https://bravos-ai.com/) we do it differently: - **Upload a CSV** — no converting to PDF, no building an API - **AI extracts filters automatically** from the customer's question: "apartments in Miami under $500,000 with 3 bedrooms" → City = Miami, Price ≤ 500,000, Bedrooms ≥ 3 - **Runs real queries** on your data — exact filtering, not approximate - **Supports exclusions** — "no garage", "not a penthouse" - **If there are no results, it says so** — doesn't make things up or offer "something similar" - **RAG still works** for what it's good at — policies, FAQs, service descriptions — but your catalog gets filtered with real queries The difference is that filtering is exact. If an apartment costs $501,000, it doesn't show up. If it has 2 bedrooms, it doesn't show up. No "approximate, not guaranteed". ## Conclusion AI chatbots have advanced enormously, but most still use a technology (RAG) that isn't designed for tabular data. For FAQs and documentation, they work wonderfully. For an **AI chatbot with a product catalog** that needs to filter by attributes, they don't. If your business depends on a catalog — whether products, properties, courses, or services — make sure the platform you choose can do real filtering on structured data. Not "approximate". Not "convert your CSV to PDF". Exact, automatic filtering, no coding required. With [Bravos AI](https://bravos-ai.com/) you can upload a CSV with your catalog — or connect your [Shopify](https://bravos-ai.com/blog/chatbot-shopify), [WooCommerce](https://bravos-ai.com/blog/chatbot-woocommerce) or [PrestaShop](https://bravos-ai.com/blog/chatbot-prestashop) store directly with a native integration — and the chatbot automatically filters by any attribute: price, location, category, whatever. No code, no manual setup. [Try it free for 7 days](https://app.bravos-ai.com/register?lang=en) and see how the filters perform on your own catalog. ### Chatbot can't filter your catalog? Try Bravos AI's exact filtering on your own catalog in under 5 minutes. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # How Much Does an AI Chatbot Cost in 2026? Real Prices, Hidden Fees, and ROI **Category:** Practical guide | **Read time:** 12 min | **Date:** Feb 3, 2026 | **URL:** https://bravos-ai.com/en/blog/ai-chatbot-pricing If you search "AI chatbot pricing" online, you'll find two types of results. The ones that say "it depends" without giving you a single number. And the ones that throw out a range like "$0 to $500,000" that helps absolutely no one make a decision. This article is different. We're going to show you real prices from real platforms, explain the pricing models that exist (and the hidden traps in each), and give you a concrete method to calculate whether a chatbot is actually worth it for your business. With numbers. Because the real question isn't how much a chatbot costs. It's how much it costs you NOT to have one. ## The Three AI Chatbot Pricing Models in 2026 Before comparing platforms, you need to understand how they charge. A chatbot that looks cheap can end up costing three times more if you don't read the fine print. ### Model 1: Flat rate (fixed monthly price) You pay a fixed amount per month with a usage limit. This is the most predictable model and the one that works best for small businesses that need to control spending. **Advantage:** You know exactly what you'll pay each month. **Risk:** If you exceed the limit, you either pay extra or your chatbot stops working. ### Model 2: Pay per resolution or per conversation You pay each time the chatbot fully resolves a query without human intervention. A "resolution" is the entire conversation, not each individual message — a single resolution can involve 10-20 messages exchanged. It sounds fair ("you only pay when it works"), but the bill can skyrocket if your chatbot is good and resolves a lot. **Advantage:** If the chatbot doesn't resolve, you don't pay. **Risk:** Your bill grows unpredictably. ### Model 3: Pay per agent or per seat You pay for each person on your team who needs access to the dashboard. The chatbot may be an add-on or included, but the price scales with your team size. **Advantage:** Predictable if your team is small. **Risk:** Scales fast. A team of 5 people can mean hundreds of dollars in licenses alone, before you even count the AI chatbot. **Not all platforms measure the same thing.** Some charge per **message** (each individual message), while others charge per **conversation** or **resolution** (the entire interaction, regardless of how many messages it contains). A single conversation can involve 10-20 messages. If you don't know which unit each platform uses, you can't compare prices. In the table below, we specify what each one measures. ## AI Chatbot Pricing Comparison: Real Prices in 2026 Prices verified directly on each platform's official pricing page in February 2026. If you're reading this later, double-check that they're still current. The average chatbot pricing in 2026 sits between $20 and $60/month for small businesses, although enterprise tiers can easily exceed $2,000/month depending on volume. | Platform | Base price | AI chatbot | Billing unit | | --- | --- | --- | --- | | Bravos AI | Starter: $23/mo (€19) · Pro: $59/mo (€49) (7-day free PRO trial) | Included in all plans (RAG) | Unlimited | | Tidio | Starter: $29/mo | From $39/mo extra (50 conversations) | Conversations | | Intercom | From $29/seat/mo (annual) | $0.99 per AI-resolved query | Resolutions | | Crisp | Essentials: $114/mo (€95; AI not in the $54/€45 Mini plan) | Included (~450 automated conversations, $30/€25 in AI credits) | Credits | | Chatbase | Hobby: $19/mo | 2,000 credits/mo (1 credit ≠ 1 message) | Credits | | Zendesk | From $19/agent/mo | From $1.50 per AI-resolved query | Resolutions | **An example of how costs escalate:** If your chatbot resolves 1,000 queries per month on Intercom, you pay $990 in AI resolutions alone — on top of the per-seat base price. With Bravos AI you get unlimited messages from $23/mo (Starter) or $59/mo (Pro, with lead capture). No per-resolution fees. If you want to explore more options, directories like [Dang.ai](https://dang.ai) list thousands of AI tools with filters by category and pricing. ## How Much Does It Cost NOT to Have an AI Chatbot? Before deciding whether a chatbot is worth it, think about what you're already paying without realizing it. ### Your team answers the same questions over and over Think about the inquiries your business receives each week. How many are variations of the same thing? Business hours, pricing, availability, how the service works, return policies. In most small businesses, 50% to 80% of inquiries are questions that already have an answer on your website — but customers can't find it or prefer to ask directly. Each of those inquiries costs someone on your team time. Time they're not spending closing sales, handling complex cases, or growing the business. ### After hours, nobody's there If your business operates from 9 to 6, there are 15 hours every day — plus weekends and holidays — when a potential customer can land on your website and find no one. That lead either goes to a competitor or simply goes cold. This isn't just a hunch: a study by MIT and InsideSales.com analyzed over 15,000 leads and 100,000 contact attempts over 3 years. The finding: contacting a lead within the first 5 minutes makes you 21 times more likely to qualify them compared to waiting 30 minutes. A chatbot responds instantly, even at 3 AM. If you still rely solely on a contact form to capture leads, you may be losing more opportunities than you think. In our article on [chatbots vs forms](https://bravos-ai.com/blog/formulario-contacto-vs-chatbot) we analyze the data in detail. ### The real cost: an example with numbers Picture a typical small business: - **200 inquiries per month** (email, forms, phone, social media) - **1 person partially dedicated** to answering them (20 hours/week, company cost: $2,000/month) - **No after-hours support** That person spends 80 hours per month answering inquiries. If a chatbot resolves half of them — the repetitive ones, the ones with a clear answer — you free up 40 hours per month. That's $1,000 in recovered time that person can spend on tasks that actually generate revenue. | Concept | Value | | --- | --- | | Time freed per month | 40 hours | | Value of that time (50% of $2,000) | $1,000 | | Chatbot cost (Bravos AI Starter) | $23/month | | Net monthly savings | $977 | What if we're twice as conservative? At 25% automation, you free up 20 hours: $500 in savings versus $23 in cost. Still profitable from month one. And this doesn't even account for the leads you capture after hours — which today are simply lost. The person doesn't disappear. They stop answering "what are your business hours?" twenty times a day and focus on what actually grows your business. ## AI Chatbots in 2026: What the Data Says Gartner predicted in 2022 that conversational AI would save $80 billion in contact center labor costs by 2026. We're in that year now. But there's an interesting tension: 64% of customers would still prefer that companies didn't use AI for customer service, according to a Gartner survey of 5,728 consumers. The main concern: that it becomes harder to reach a human. The reason isn't that they hate AI. It's that they've had bad experiences with bad chatbots — the ones that gave generic answers or flat-out made things up. The difference in 2026 is that technologies like RAG allow the chatbot to respond exclusively with your company's information, not hallucinations. If you're worried about your chatbot making things up, our article on [why your chatbot fails and how to fix it](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo) explains exactly how to prevent it. ## Mistakes That Make Your AI Chatbot More Expensive ### Confusing a support platform with a chatbot Many of the platforms in this comparison (Intercom, Zendesk, Crisp, Tidio) **aren't chatbots — they're full customer support platforms** that happen to include a chatbot as one feature among many. What you're really paying for is a ticketing system, CRM, agent routing, integrations with Salesforce or Shopify, and analytics dashboards for support teams. If you have a team of 15 people managing thousands of tickets daily, that makes sense. But if you're a small business with 5-10 people, you're probably paying for infrastructure you'll never use. Here's something rarely mentioned: these platforms assume you have human agents available for when the chatbot escalates a query. But in a small business, that escalated ticket ends up answered the next day — which is exactly the same as a [contact form](https://bravos-ai.com/blog/formulario-contacto-vs-chatbot). If you don't have someone dedicated to handling escalations in real time, the routing and handoff system you're paying for adds no value. Before you sign up, ask yourself: **do I need a chatbot or a support platform?** They're different things, with very different price tags. ### Ignoring how the platform handles your data Not all chatbots process your data the same way. Most use RAG (Retrieval-Augmented Generation): they convert your documents into text and search for fragments similar to the user's question —how that retrieval works, and whether you even need a vector database for it, is covered in [vectorless RAG](https://bravos-ai.com/blog/rag-sin-vectores)—. For FAQs and documentation, this works well. But if you have a product catalog with filterable attributes (price, category, size, location), RAG alone gives approximate results, not exact ones. The alternative offered by platforms like Intercom or Zendesk is for you to build your own API — a cost that doesn't appear in any pricing table. With Bravos AI, you upload a CSV with your catalog and the chatbot automatically filters by any attribute (price, location, category) without writing a single line of code. And if you sell on [Shopify](https://bravos-ai.com/blog/chatbot-shopify) or [WooCommerce](https://bravos-ai.com/blog/chatbot-woocommerce), it connects directly to your store — you don't even need the CSV. ### Not training the chatbot properly An untrained chatbot is a chatbot that resolves nothing. And a chatbot that resolves nothing has negative ROI. Invest time in [preparing your content properly](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo): clear FAQs, well-structured product information, well-written policies. ### Paying for features you don't need Many platforms charge extra for things like removing their branding from the widget, adding channels like WhatsApp, or accessing analytics. Before signing up, make a list of what you actually need. If your business does $500,000 a year in revenue, you don't need a platform that costs $2,500/month. That's like renting a bus to go buy groceries. For the WhatsApp-specific case, our [complete WhatsApp AI chatbot guide](https://bravos-ai.com/blog/chatbot-whatsapp) covers the whole ecosystem (app vs API, free vs paid, setup) and our [comparison of the 13 best WhatsApp chatbot platforms](https://bravos-ai.com/blog/mejor-chatbot-whatsapp) breaks down the real prices and the hidden Meta surcharge almost nobody talks about. ## Conclusion The cost of an AI chatbot in 2026 ranges from $0 to thousands of dollars per month. For most small businesses, a SaaS solution between $19 and $60/month covers everything you need. But price is only part of the equation. What matters is how much time you free up for your team, how many leads you capture after hours, and how many repetitive inquiries you stop handling manually. As we've seen, even with conservative estimates, a chatbot pays for itself from the very first month. Our advice: don't overthink it. Start with the 7-day PRO trial, train it with your information, and measure it in real conditions with your own traffic. Real data from your business is worth more than any pricing table. And one advantage you won't see in any pricing table: managing and improving the bot doesn't force you to pay for a separate tool. If you already work with Claude or ChatGPT, with [our MCP connector](https://bravos-ai.com/conector-ia-mcp) you do it from that same assistant you already pay for —audit conversations, fix the content, adjust the tone—, without adding another subscription to administer it. With [Bravos AI](https://bravos-ai.com/) you can create your chatbot with 7 days free on the full PRO plan, train it with your company's data using RAG technology, and install it on your website in minutes. No commitment. And if you need more — choose from premium AI models like GPT-4o, Claude, or Gemini, fully customize how your chatbot behaves, or get dedicated support with guaranteed response time — our Enterprise plan adapts to whatever you need. ## Frequently asked questions ### What are the AI chatbot pricing models in 2026? There are three main pricing models. **Flat rate**: a fixed monthly price with a usage limit, the most predictable for small businesses. **Pay per resolution or per conversation**: you pay each time the chatbot fully resolves a query without human intervention; sounds fair but the bill can skyrocket if your chatbot is good. **Pay per agent or per seat**: you pay for each person on your team with dashboard access; predictable for small teams but scales fast with headcount. Important caveat: some platforms measure by message, others by conversation or resolution, so unit definitions matter when comparing prices. ### How much does an AI chatbot cost in 2026? Verified February 2026 prices across major platforms: Bravos AI Starter is $23/month and Pro $59/month with unlimited messages and AI included in all plans (7-day free trial of the full PRO plan). Tidio Starter is $29/month with AI chatbot from $39/month extra. Intercom is from $29/seat/month plus $0.99 per AI-resolved query. Crisp Essentials is $114/month ($95 EUR) with ~450 automated conversations included. Chatbase Hobby is $19/month for 2,000 credits. Zendesk is from $19/agent/month plus $1.50-$2 per AI resolution. The total depends heavily on the pricing model and your volume. ### Is an AI chatbot worth it for a small business? Yes, the math works out from month one in most realistic scenarios. Example: a small business with 200 inquiries/month and one person partially dedicated to answering them (80 hours/month, $2,000 cost). If a chatbot resolves half of those repetitive queries, you free up 40 hours/month — $1,000 in recovered time. Versus a Bravos AI Starter plan at $23/month, that is $977 net monthly savings. Even at 25% automation (conservative), you free up 20 hours for $500 in value versus $23 in cost. Still profitable. ### What mistakes make an AI chatbot more expensive than it should be? Four common mistakes. **Confusing a support platform with a chatbot**: Intercom, Zendesk, Crisp and Tidio are full customer support platforms (ticketing, CRM, agent routing) where the chatbot is one feature among many — overkill for a 5-10 person team. **Ignoring how the platform handles structured data**: RAG alone gives approximate results on product catalogs; if you sell products you need exact filtering, not similarity search. **Not training the chatbot properly**: an untrained chatbot resolves nothing, which means negative ROI. **Paying for features you do not need**: removing branding, adding WhatsApp, analytics dashboards. If you bill $500K/year, a $2,500/month platform is overkill. ### Want to calculate your ROI? Try the PRO plan free for 7 days and measure how much time it saves your team in real conditions. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # PixelRAG Explained: What It Is and Whether It Makes Sense for Your Business Chatbot **Category:** Analysis | **Read time:** 20 min | **Date:** June 30, 2026 | **URL:** https://bravos-ai.com/en/blog/pixelrag-business-chatbot When an AI paper comes out with eye-catching numbers, one of two things usually happens: either nobody reads it, or everyone makes up what it says. PixelRAG falls into the second group. The viral threads claim it "kills text-based RAG", that it improves accuracy by 18% with half the tokens, that there's no longer any point in training text *embeddings*. We read the full paper, downloaded the code and calculated real costs using current market pricing. What follows is the honest explanation for someone evaluating whether PixelRAG actually fits their business chatbot: what it is, how it works under the hood, what the real numbers say (not the headlines), what it costs in production and when it genuinely helps. With direct quotes from the paper itself. Index - [What PixelRAG actually is](#what-is) - [How it works under the hood](#how-it-works) - [PixelRAG vs traditional RAG: what changes](#vs-traditional) - [The real results (the +18% that isn't +18%)](#results) - [What it actually costs](#cost) - [The limitations the authors themselves acknowledge](#limitations) - [When PixelRAG actually makes sense](#when-yes) - [When it doesn't (most business chatbots)](#when-no) - [The alternatives that already exist](#alternatives) - [Quick test: is it for your chatbot?](#test) - [Viral headlines vs what the paper says](#headlines) - [How we do it at Bravos AI](#bravos) - [Summary and FAQ](#summary-faq) - [Sources](#sources) ## What PixelRAG actually is PixelRAG is a RAG (*Retrieval-Augmented Generation*, the technique most AI chatbots use to answer with your information instead of making things up) system published in June 2026 by researchers from UC Berkeley, Princeton, EPFL, Databricks and Renmin University. The difference with traditional RAG isn't in the language model. It's in how the information is prepared and retrieved. Traditional RAG — the kind 99% of chatbot platforms use, Bravos AI included — works like this: it takes your documents (a PDF, a website, a catalog), chunks them into text fragments, converts them into numerical vectors called *embeddings* and stores them in a vector database. When a customer asks something, the system finds the fragments most similar to their question and passes them to the language model. Works well for text: FAQs, descriptions, policies, manuals. Whether that vector database is even necessary is exactly the [vectorless RAG](https://bravos-ai.com/blog/rag-sin-vectores) debate. PixelRAG changes the first step. Instead of extracting text from your documents, it renders each page as an image (a screenshot, literally) and stores that image. When a customer asks, the system finds the most relevant images and passes them (not text) to a multimodal language model — one that can read images, like GPT-4o or Qwen3-VL — which reads them the way a human would and answers. Why do this? Because text, when extracted from a laid-out PDF, loses a ton of information: tables get broken, charts disappear, layouts stop making sense. A financial PDF with a ratios table becomes a list of disconnected numbers without context. An infographic stops existing entirely. PixelRAG preserves all of that because it treats each page as an image. PixelRAG is **open source** under Apache 2.0. The official repository is at [StarTrail-org/PixelRAG](https://github.com/StarTrail-org/PixelRAG) (5,700+ stars as of late June 2026). The technical paper is not on arXiv yet: it's a PDF uploaded to the repo itself, without peer review. That matters. ## How it works under the hood The end-to-end flow, per sections 3.1 and 3.2 of the paper: - **Render.** Every page of every document is rendered as an image with a headless browser (Chromium via Playwright). For all of Wikipedia (7 million articles) this takes about 2 days on infrastructure with 128 cores, 2 TB of RAM and 8 H100 GPUs. - **Tiling.** Each image is split into *tiles* of 875 pixels wide by 1024 tall, with no overlap. - **Embedding.** Each tile is turned into a 2048-dimensional vector using Qwen3-VL-Embedding-2B, a 2-billion-parameter vision-language model with a custom fine-tune on screenshot data. - **Indexing.** Vectors are stored in a FAISS IVF index. For the 30M tiles covering all of Wikipedia, the index takes ~120 GB and the images on disk take **5.6 TB**. - **Search.** When a question comes in, it's embedded with the same model and the closest tiles are retrieved (top-3 by default in the paper). - **Reading.** The selected tiles (images) are passed to a final multimodal model — the paper defaults to Qwen3-VL-4B — which "reads" the images and generates the answer. It's a coherent architecture. It solves a real problem: text-based RAG loses visual structure. And it solves it at scale (30M tiles for Wikipedia) without falling into ColPali-style multivector, which would be prohibitively expensive at that size. The question isn't whether the architecture is good. It's whether the problem it solves is *your* problem. ## PixelRAG vs traditional RAG: what changes The most intuitive difference: faced with a Wikipedia page with a table, this is what each system "sees" before passing it to the language model: Boston | State | Massachusetts | | --- | --- | | Population | 675K | | Metro | 4.9M | | Density | 13,938/mi² | , (the table is preserved as an image, the multimodal model reads it like a person would) , ]} /> That's the promise. When the content is plain text — your dental clinic's FAQ, your store's return policies, your service descriptions — there's no difference. Text-based RAG extracts the text perfectly. PixelRAG's promise is for content where layout carries information: tables, infographics, spec sheets, manuals with diagrams. This table summarizes the real technical differences (not the marketing ones): | Dimension | Text-based RAG | PixelRAG | | --- | --- | --- | | Storage | Plain text + vectors. A few GB for Wikipedia. | 5.6 TB of images + 120 GB of index for Wikipedia. | | Indexing | Minutes on CPU for mid-size corpora. | ~2 days on 8× H100 for Wikipedia. | | Tokens per query | ~1,700 (text). | ~2,625 visual (3 tiles × 875). | | Latency | Sub-second for retrieval. | Not reported in the paper. Processing images in the LLM is slower. | | Languages | Multilingual (100+ with OpenAI embeddings). | English only. No proven transfer to other languages. | | Shines at | Plain text, FAQs, descriptions, policies. | Tables, infoboxes, visual layouts. | | Fails at | Laid-out PDFs with complex tables/charts. | Lists, unstructured content, navigation across pages via links. | ## The real results (the +18% that isn't +18%) The viral headline is "PixelRAG improves accuracy 18% over text-based RAG". That number does appear in the paper's abstract. But it's the improvement **on the best benchmark**, not the average. When you break down the 6 actual benchmarks the paper reports, the picture changes: The peak of +15.5 percentage points is on EVQA (*Encyclopedic Visual Question Answering*), a benchmark of questions about Wikipedia infographics and visual elements. Exactly where you'd expect an image-based system to shine. On typical text questions (Natural Questions), the improvement is under 3 points. On tables (NQ-Tables) it's 6.3 points — decent, not spectacular. And there's something else the headline hides. The paper breaks down what type of evidence PixelRAG retrieves better in SimpleQA (section 5.2, table 2): - **Tables:** +9.1 points (where it really shines). - **Bordered infoboxes:** +4.6 points. - **Paragraphs:** +7.9 points (surprisingly high; the authors attribute it to infoboxes "displacing" relevant paragraphs from the top-3 in text-based RAG). - **Lists:** +0.5 points (essentially no gain, within error margin). And the most important detail of all for this discussion: **all 6 benchmarks are on Wikipedia and news articles (CNN, BBC, AP)**. None of them touch product catalogs, corporate FAQs, software manuals, service descriptions, legal policies or anything resembling what a real business chatbot handles. The paper's own section 5.1 says so. The second metric going around is "10× fewer tokens". That number is real, but also misleading out of context. It appears in section 5.4 and refers to use in **agents** of the ReAct kind: a model that does multiple searches and reasons in a loop, with up to 20 steps per question. In a typical business chatbot — question, single search, answer — that reduction doesn't apply. What applies is the comparison of tokens per single query, and there PixelRAG consumes *more* tokens (visual ones) than text-based RAG. One note about the paper's rigor before we make decisions: **it is not published on arXiv and has not gone through peer review** as of late June 2026. It's a preprint uploaded directly to the repo. That doesn't invalidate it, but it means the abstract's numbers have not been audited by the academic community or independently replicated yet. ## What it actually costs The paper does not publish end-to-end monetary costs. We calculated the real costs for a typical business chatbot (10,000 messages per month) using current official pricing: | Configuration | Cost per query | 10,000 messages/month | Notes | | --- | --- | --- | --- | | Standard text-based RAG (embedding-3-small + GPT-4.1-mini) | ~$0.00092 | ~$9.20 | What 99% of SaaS platforms use today. | | PixelRAG with self-hosted Qwen3-VL-4B | ~$0.00015 *API only* | ~$1.50 + GPU | Requires maintaining a dedicated GPU ($200-500/mo extra). | | PixelRAG with GPT-4o as reader (or an equivalent commercial multimodal model) | ~$0.0066 | ~$66 | **7× more expensive** than standard text-based RAG. | And that's just inference cost per query. Initial indexing also adds up. For a small 1,000-page corpus, PixelRAG needs ~$0.55 in GPU time to process the images. Text-based RAG with OpenAI embeddings processes the same for ~$0.20. Not the biggest gap, but it compounds if you reindex often. Then there's storage. The 5.6 TB Wikipedia images take vs the few GB the same content takes as text. In cloud, that's ~$110/month just in standard S3 (at $0.023/GB). For a mid-sized enterprise corpus (50,000 pages of internal docs) the numbers are more reasonable, but always 100× to 1,000× larger than the equivalent text-based RAG. The "cheap PixelRAG" number ($1.50/month) only applies if you self-host Qwen3-VL-4B. If your team doesn't have its own GPU infrastructure, that figure stops being representative. And if you use a commercial multimodal model (GPT-4o or Claude as the reader), PixelRAG ends up **more expensive** than text-based RAG, not cheaper. The "savings" story holds only in a very specific scenario. ## The limitations the authors themselves acknowledge Appendix E of the paper, titled "Limitations", is one of the more honest pages we've read recently in an AI paper. The authors list five serious limitations without sugar-coating them. Three directly impact any business chatbot: ### 1. English only all datastores in this work are English-only [...] introducing a language bias The embedding fine-tune was trained on English-only Wikipedia screenshots. There is no evidence the system works well in Spanish, French, German or any non-English language. For a business chatbot operating in non-English markets, this is a short-term deal-breaker. ### 2. Loses hyperlinks hyperlinks are visually rendered (e.g., as blue underlined text) but are not directly actionable; the system cannot follow a link to retrieve the target page The system sees links as blue underlined drawings, not as paths to other documents. If your chatbot needs to answer things like "per the return policy (link)…" by navigating between pages, it can't. For enterprise chatbots with interlinked knowledge bases (FAQs referencing policies, products referencing spec sheets), this breaks the flow. ### 3. Content moderation is harder screenshot-based retrieval faithfully preserves whatever appears on a rendered page, including potentially harmful, misleading, or private content. Unlike text pipelines, where filtering can operate on extracted strings, pixel content is harder to moderate automatically In an e-commerce or healthcare chatbot, where the knowledge base can contain sensitive data, filtering image content is operationally more expensive and less reliable than filtering text. For companies with GDPR or HIPAA requirements, this adds compliance friction. The other two limitations they acknowledge are the storage overhead (mentioned above) and the decision to use single-vector instead of multivector ColPali-style, which forces them to lose fine-grained granularity inside each tile. ## When PixelRAG actually makes sense So that this article doesn't end up as a hit-piece, let's look at the cases where PixelRAG *does* bring something text-based RAG can't. They're niche, but they exist. Three concrete profiles: ### Profile 1: Historical archives, museums, digital newspaper archives If your knowledge base is scanned old newspapers, historical maps, handwritten letters, catalog cards, photos with captions — content where plain text doesn't exist or isn't available — a system that processes images directly is clearly superior. Text-based RAG doesn't even compete here: you'd first have to OCR everything, losing layout information and non-text elements. PixelRAG (or ColPali, or similar systems) is the right path. ### Profile 2: Technical documentation with many diagrams and schematics Industrial manuals, safety data sheets, electrical schematics, exploded-view diagrams, engineering documentation where information lives in the drawings, not in the text. If your chatbot has to answer "where is the pressure regulator on this model?" and the answer is in a schematic, PixelRAG can help. **But**: the final multimodal model has to understand the domain (electrical schematics aren't the same as Wikipedia infographics). In most cases you'll need domain-specific fine-tuning, which multiplies the cost. ### Profile 3: Financial and legal documents with tables and complex layouts Annual reports with ratio tables, contracts with multi-level table clauses, quarterly balance sheets, fund fact sheets. Here PixelRAG competes with AWS Textract and Unstructured.io, which have been extracting this kind of table to structured text for years. PixelRAG can add precision, especially if layouts vary a lot. If your volume justifies it, it's worth evaluating. In all three profiles, "makes sense" means "worth running a 2-4 week proof-of-concept, comparing against alternatives", not "automatically replaces text-based RAG". And for all three, what the paper proves today is **potential**, not validated production in your specific domain. ## When it doesn't (most business chatbots) For typical business chatbot cases — the ones we see daily at Bravos AI — PixelRAG adds no advantages and adds costs. Concrete cases where text-based RAG is still clearly the right option: - **E-commerce with a CSV or JSON catalog.** "Waterproof jackets under $80 in size L" is a structured query better solved with SQL filtering over the structured data you have. Converting the catalog to images and running it through a multimodal model is overkill. (We expand on this in our guide on the [AI chatbot for product catalog](https://bravos-ai.com/blog/chatbot-catalogo-productos).) - **FAQs for clinics, restaurants, consultancies, agencies.** Plain text, service descriptions, hours, prices. Text-based RAG retrieves them without losing anything. - **Policies, legal terms, terms of use.** Sometimes in PDF, but usually plain text. Text-based RAG handles them well. - **SaaS support documentation.** Help center articles, usage guides, API docs. Text, code, occasional screenshots. Text-based RAG covers 95%. - **Real estate listings, hotels, restaurants, events.** Structured data (price, date, location, capacity). Again, SQL + text. - **Any multilingual use case.** If your chatbot operates in Spanish, German, French, Portuguese or any non-English language, PixelRAG isn't validated. - **Any case where latency matters.** Processing images in the final model adds latency. For a chatbot where the customer expects a response in under 2 seconds, the latency cost may not be worth it. These cases are 90% (or more) of real business chatbots. For them, PixelRAG is an expensive solution to a problem you don't have. ## The alternatives that already exist (and have been in production for years) If your problem really is preserving the visual layout of complex documents, PixelRAG is neither the first nor the only option. Tools for this have existed for years, and some are already in production at thousands of companies: | Tool | Approach | Cost | Maturity | | --- | --- | --- | --- | | AWS Textract | Extracts tables to structured JSON, plug-and-play with standard text-based RAG. | $1.50 per 1,000 pages. | Production since 2019. | | Unstructured.io | Hybrid parser (rules + ML) that preserves tables as HTML/JSON. | Open source or $0.01-0.10 per page via API. | Mature, integrated in LlamaIndex and LangChain. | | pdfplumber / PyMuPDF | Local extraction of text and tables. | Free (open source). | Mature. | | Claude / GPT-4o with direct vision | Pass the PDF as image directly to the model. No separate pipeline. | ~$0.003 per page with Sonnet 4. | Production, supported in the APIs. | | ColPali | Visual RAG with multivector. Academic ancestor of PixelRAG (ICLR 2025). | Memory-intensive, ~256 KB/page. | Peer-reviewed. Vespa and Qdrant support it. | For most companies with laid-out PDFs, a combination of **Textract or Unstructured.io for preprocessing + text-based RAG** solves 90% of the problem at a reasonable cost. For very demanding cases, Claude with direct vision or ColPali are validated alternatives. PixelRAG enters as a sixth option, not the first. ## Quick test: is it for your chatbot? Five questions. Count how many you answer "yes": **1.** My main knowledge base is laid-out PDFs with many complex tables, diagrams or infographics. **2.** My chatbot operates *only* in English. **3.** I have budget for either maintaining my own GPU infrastructure ($200+/month extra) or paying $60-100/month in multimodal API costs. **4.** Latency is not critical (I can accept 5-10 seconds per response). **5.** I have a technical team to integrate research code (without commercial support) and maintain it. - **4-5 yeses:** running a proof-of-concept with PixelRAG is worth it. Compare against ColPali and against Claude with direct vision before deciding. - **2-3 yeses:** first look at AWS Textract or Unstructured.io combined with standard text-based RAG. They will very likely cover your case at a tenth of the cost. - **0-1 yes:** standard text-based RAG is your option. PixelRAG solves a problem you don't have. ## Viral headlines vs what the paper says Final pass, with citations. Five claims circulating in headlines and what the paper actually says: | Viral headline | What the paper says | | --- | --- | | "+18% better accuracy than text RAG" | Peak of +15.5 points on EVQA (visual Wikipedia). On text Wikipedia QA, +2.8 to +7.2 points. *(Section 5.2, table 1.)* | | "10× fewer tokens than text RAG" | In ReAct multi-step agents with up to 20 searches per question. Not in single-turn chatbots. *(Section 5.4.)* | | "The end of text-based RAG" | The authors explicitly propose a hybrid text + vision system in their Future Work section. *(Appendix E, p. 32.)* | | "Works out-of-the-box" | Requires domain-specific fine-tuning. Wikipedia fine-tune does not transfer well to news, per the paper itself. *(Section 5.2.)* | | "Production-ready for enterprise" | No arXiv, no peer review, 6 months old, no real adversarial technical debate on HN/Reddit yet. Solid research code, not a commercial platform. | ## How we do it at Bravos AI At [Bravos AI](https://bravos-ai.com/) we cover the typical business chatbot cases — FAQs, product catalogs, service descriptions, policies — in 13+ languages including English, French, German, Spanish and Arabic, with latency under 2 seconds. Plans from $23/month with unlimited messages. Will we evaluate PixelRAG or some descendant of it? Yes, when three things happen at once: peer-reviewed publication, validated multilingual support, and per-query cost below SaaS market price with a quality VLM. Today, none of the three is true. When they are, we'll evaluate. In the meantime, migrating would be bad engineering. ## In summary - PixelRAG is a genuine technical advance for a specific problem: preserving the visual layout of documents when plain text destroys it. - The "+18%" from the headlines is the peak on a visual Wikipedia benchmark (EVQA). On typical text QA, the improvement is 2.8 to 7.2 percentage points. - The "10× fewer tokens" applies in ReAct multi-step agents, not in single-turn chatbots. - Only validated in English. No proven transfer to other languages. - With a commercial multimodal model (GPT-4o as reader), it's 7× more expensive than text-based RAG. The "cheap PixelRAG" story only holds if you self-host Qwen3-VL. - The authors themselves acknowledge three serious limitations: English only, loses hyperlinks, harder content moderation. - For typical business chatbot cases (FAQs, catalogs, policies, descriptions), text-based RAG + SQL filtering is still the answer. - For niche cases with complex tables, diagram-heavy manuals or visual archives, also evaluate AWS Textract, Unstructured.io, Claude with direct vision or ColPali before deciding on PixelRAG. ### What is PixelRAG in a nutshell? A RAG system that, instead of extracting text from documents, renders them as images (screenshots), indexes those, and feeds them to a multimodal language model that reads them like a person would. Published by researchers from Berkeley, Princeton, EPFL, Databricks and Renmin University in June 2026, under Apache 2.0. ### Is PixelRAG better than text-based RAG? It depends on the type of content. For questions about tables, infographics and visual elements on Wikipedia-style pages, yes: the paper reports 6 to 15 percentage points of accuracy gain. For plain text, FAQs, product descriptions, policies and most typical business content, it doesn't help and ends up more expensive. ### What does PixelRAG cost in production? Depends on what model you use as reader. With self-hosted Qwen3-VL-4B, ~$1.50/month in API for 10,000 messages, but you need to maintain a dedicated GPU ($200-500/month extra). With GPT-4o as reader (or an equivalent commercial multimodal model), ~$66/month, which is 7× more expensive than a standard text-based RAG. ### Does PixelRAG work in non-English languages? There is no evidence it works well in Spanish, French, German or any non-English language. The embedding fine-tune was trained only on English data and the authors themselves acknowledge the bias in the limitations appendix. For a chatbot operating in non-English markets, this is a veto. ### Do I need to migrate my chatbot to PixelRAG? Almost certainly not. If your chatbot handles FAQs, service descriptions, policies, product catalogs in CSV or JSON, or any typical business content, text-based RAG + SQL filtering is still better and cheaper. PixelRAG solves a problem (preserving visual layout) that most chatbots don't have. ### Is PixelRAG better than ColPali? They're different approaches to the same problem. ColPali (Faysse et al, ICLR 2025) uses multivector retrieval with a smaller model; PixelRAG uses single-vector with a larger model to scale to Wikipedia-sized collections. ColPali is peer-reviewed, more mature, and has production integrations (Vespa, Qdrant). PixelRAG is newer and scales to more documents, but it's 6 months younger and without peer review. ### When will PixelRAG be worth it? When three things happen at once: the cost of multimodal models (visual token) drops at least 5×, validated multilingual support appears, and the academic community audits the results with peer review. Reasonably, within 12-18 months a mature variant of this paradigm will be relevant for some enterprise use cases. Not today. ## Sources - **PixelRAG original paper (PDF):** [github.com/StarTrail-org/PixelRAG/assets/pixelrag-paper.pdf ](https://github.com/StarTrail-org/PixelRAG/raw/main/assets/pixelrag-paper.pdf) — Wang et al, June 2026. UC Berkeley, Princeton, EPFL, Databricks, Renmin University. Apache 2.0. - **Official repository:** [github.com/StarTrail-org/PixelRAG ](https://github.com/StarTrail-org/PixelRAG) - **ColPali (academic ancestor):** Faysse et al, "ColPali: Efficient Document Retrieval with Vision Language Models", ICLR 2025. - **VisRAG:** Yu et al, "VisRAG: Vision-based Retrieval-augmented Generation on Multi-modality Documents", ICLR 2025. - **AWS Textract docs:** [docs.aws.amazon.com/textract ](https://docs.aws.amazon.com/textract/) - **Unstructured.io:** [unstructured.io ](https://unstructured.io/) - **OpenAI official pricing (text and vision):** [openai.com/pricing ](https://openai.com/pricing) ### Build your business chatbot without the hype At Bravos AI we build business chatbots that answer well on FAQs, catalogs, policies and service descriptions. In 13+ languages, with latency under 2 seconds. 7-day PRO trial, no commitment: we notify you before charging and if you cancel before day 7 you pay nothing. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Proactive vs Reactive Chatbot: Why Asking Questions Converts More **Category:** Analysis | **Read time:** 10 min | **Date:** Jan 27, 2026 | **URL:** https://bravos-ai.com/en/blog/proactive-vs-reactive-chatbot Most chatbots work the same way: the user asks a question, the bot answers. The user asks another question, the bot answers again. And so on until the visitor leaves. If you're using a chatbot for sales or lead generation, this reactive model leaves a lot of opportunity on the table. There's another approach. A proactive chatbot that doesn't just answer, but knows when to ask questions. And decades of sales data, buyer psychology research, and conversational AI insights suggest that this difference can be the decisive factor in your conversion rate. ## What Is a Reactive Chatbot? A reactive chatbot does exactly what the name implies: it reacts. The user types a question and the bot searches its knowledge base for the best answer. If it finds one, it delivers it. If not, it says it doesn't have that information (or, in the worst case, [makes something up](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo)). This is the most common chatbot model, and it makes sense for many use cases: FAQs, business hours, return policies, pricing. The user knows what they're looking for, the bot provides it. Done. The problem shows up when the user **doesn't know exactly what they need**. When they're exploring options, comparing products, or have a need they can't quite articulate. In those moments, a chatbot that only answers falls short — and you lose a potential conversion. ## What Is a Proactive Chatbot (and What It Is Not) When we talk about a proactive chatbot, we don't mean the typical popup that appears after 5 seconds saying "Hi! Can I help you?" That's an interruption, not proactivity. A proactive chatbot is one that, during the conversation, detects opportunities to ask questions that add value. If a user says "I'm looking at this for my company," instead of dumping a generic paragraph, the bot asks: "What problem are you trying to solve?" or "How many people would be using it?" The distinction is subtle but critical for AI chatbot conversion rates: - **Reactive:** Answers everything directly. Never asks discovery questions. - **Proactive:** Answers direct questions, but also asks when the user shows interest without providing enough context. And there's a key nuance: a proactive chatbot also knows when **not** to ask. If the user wants a specific piece of information, it delivers it immediately. If it detects frustration, it backs off. The goal isn't to turn every conversation into an interrogation — it's strategic chatbot lead qualification. ## Sales Data on Asking Questions: What the Research Shows There aren't many studies specifically comparing chatbots that ask questions versus chatbots that only answer. But there are decades of research on sales performance and buyer behavior. And since 41% of chatbots are used for sales (the #1 use case according to Intercom, ahead of customer support), the connection between conversational AI sales tactics and chatbot strategy is direct. ### Gong: What Top-Performing Sales Reps Do Differently Gong is a platform that analyzes sales calls using AI. They've studied hundreds of thousands of sales conversations to identify patterns. Some of their most relevant findings for sales chatbot strategy: - **Top closers don't talk the most.** The optimal ratio is 43% talking and 57% listening. Average reps talk up to 64% of the time in deals they lose. - **The sweet spot is 11-14 questions per call.** More isn't better: reps who asked 20 questions actually closed fewer deals. It's not about how many questions you ask — it's about which ones. - **Top sellers distribute questions throughout the entire conversation.** Average reps front-load them, as if they have a checklist to complete before launching into their pitch. - **When facing objections, top performers ask questions.** The best salespeople respond to objections with questions 54% of the time, compared to 31% for average performers. The takeaway is clear: asking the right questions is a competitive advantage in sales. And a chatbot is essentially a sales rep available 24/7. This is why a proactive chatbot approach directly improves AI chatbot conversion rates. ### Harvard Business Review: Speed-to-Lead and Buyer Confidence A study from MIT published in Harvard Business Review analyzed 2.24 million leads. The key finding: responding within the first 5 minutes makes you 21 times more likely to qualify a lead compared to waiting 30 minutes. After 5 minutes, the odds drop by 80%. But there's another less-cited data point: an analysis of 2.5 million sales conversations found that between 40% and 60% of B2B deals end with no decision at all. The buyer simply doesn't decide. It's not that they choose a competitor — they don't have enough confidence to choose anyone. Sales reps (and chatbots) who guide the decision with strategic questions reduce this decision paralysis and improve lead qualification outcomes. ### The Challenger Sale: Sales Experience Matters More Than the Product CEB (now part of Gartner) studied 6,000 B2B sales reps and discovered something that reshaped the industry: 53% of customer loyalty comes from the sales experience itself. Not the product, not the price, not the brand. It's how the seller guided the customer toward a decision. "Challenger" sellers — those who teach the buyer something new and reframe their priorities — win 40% more deals than the rest. And according to LinkedIn, 71% of B2B buyers value a deep diagnosis of their needs from the seller. They want to be asked. This data reinforces why conversational AI sales strategies built around discovery questions consistently outperform reactive-only approaches. ## The Psychology Behind Asking Questions in Sales Why does asking questions work? It's not just a sales technique. There are deep psychological reasons that apply equally to human conversations and AI chatbot interactions: **It demonstrates genuine interest.** When someone asks what you need, you feel they want to help you, not sell you something. Research published in the Journal of the Academy of Marketing Science found that when customers perceive the seller is actively listening, their trust and willingness to keep engaging increases significantly. **It helps the customer articulate their problem.** Many times users don't know exactly what they need until someone asks. "What are you trying to solve?" forces them to think. And when someone verbalizes their problem, they're already closer to seeking a solution. This is the core mechanism behind effective chatbot lead qualification. **The customer reaches the conclusion themselves.** This is the most powerful effect. If you tell a user "you need our product," there's natural resistance. But if you ask questions that lead them to realize they have a problem and that a solution exists, the decision becomes theirs. And self-made decisions generate far more commitment and higher conversion rates. **It generates longer, more detailed responses.** Gong found a direct correlation between the length of buyer responses and the probability of closing the deal. When the buyer talks more, they think more, they engage more. And questions are what trigger those longer, more revealing responses — exactly the kind of engagement that separates a high-converting chatbot from one that just deflects. ## Conversational AI Sales: What the Chatbot Data Shows While we lack definitive A/B studies on "chatbot that asks vs chatbot that doesn't," the available data from the conversational AI industry points strongly in the same direction: **Drift + Forrester: 670% ROI.** Forrester conducted an independent study on Drift's impact (a conversational marketing platform) and documented a 670% return on investment. What does Drift actually do? It deploys chatbots that ask lead qualification questions instead of using [static forms](https://bravos-ai.com/blog/formulario-contacto-vs-chatbot). Additionally, 55% of companies using chatbots generate higher-quality leads, and conversions to qualified leads increase by 75% to 100%. This makes proactive chatbots one of the most effective tools for chatbot lead generation. **Typeform: asking one question at a time works.** Typeform analyzed 2.6 million forms in 2023 and found that their conversational format (one question at a time) achieves a 47% completion rate, compared to the 21% industry average. More than double. These aren't chatbots, but the mechanics are identical: asking step by step instead of demanding everything at once. This validates the proactive chatbot approach of guided discovery. **Intercom: sales is the #1 chatbot use case.** 41% of chatbots are deployed for sales, ahead of customer support (37%). And AI-powered chatbots increase lead engagement by 25%. This makes optimizing your sales chatbot strategy not just nice-to-have, but essential. **Salesforce: 62% of customers prefer chatbots over waiting.** According to the State of Service Report 2025, based on surveys of 6,500 professionals, customers prefer interacting with a chatbot rather than waiting for a human agent. But there's a critical nuance: 74% only prefer them for simple questions. For everything else, they want to feel understood. And that requires questions, not just answers — the exact gap a proactive chatbot fills. McKinsey, in their B2B Pulse Survey with ~30,000 respondents, identified the most-cited lesson from high-growth companies: **"ask the customer what they want."** It sounds obvious, but most chatbots still don't do it. ## Proactive vs Reactive Chatbot: Key Differences and When to Use Each Asking questions isn't always better. There are situations where a reactive chatbot is exactly what you need: **A reactive chatbot works well when:** - The user is looking for specific information (hours, pricing, location) - The chatbot is primarily used for technical support or FAQs - Conversation volume is high and the goal is fast resolution - The user already knows what they want and any extra question adds friction **A proactive chatbot delivers more value when:** - The user is exploring options or comparing products - The chatbot has a commercial role (chatbot lead qualification, product recommendations) - There's a large catalog and the user needs guidance - The average deal size is high and investing in a deeper conversation pays off - The user shows interest but doesn't provide enough context for a helpful response The ideal scenario is having both modes available. Not every business needs the same approach, and within the same business, different moments in the conversation call for different behaviors. The best sales chatbot strategy combines both reactive efficiency and proactive engagement. ## Real-World Example: Same Conversation, Two Approaches Imagine someone lands on the website of a team management software company and types into the chatbot: "Hi, I'm looking for something to manage my team better" **Reactive chatbot response:** "Hi! Our team management software lets you assign tasks, track projects, manage time, and generate reports. You can see all the features on our product page. Want me to send you the link?" Technically correct. But it's generic. We know nothing about the user, so we dump everything and hope something sticks. This is where most AI chatbot conversion rates plateau. **Proactive chatbot response:** "Hi! Sure, I can help. What's the biggest challenge you're facing with managing your team right now?" The user responds: "We lose a lot of time coordinating tasks between departments." "Got it. How many people would need to use the tool?" Now the chatbot knows the specific problem and can recommend the exact feature that solves it. The user feels understood, not sold to. That's the difference between a reactive chatbot and conversational AI sales done right. ## Key Data Points: Proactive Chatbot Performance Summary | Metric | Source | | --- | --- | | 43:57 is the optimal talk-to-listen ratio in sales | Gong | | Challenger Sellers win 40% more deals | CEB/Gartner | | 53% of customer loyalty comes from the sales experience | CEB/Gartner | | Responding within 5 min = 21x more likely to qualify a lead | HBR / MIT | | 40-60% of B2B deals end with no decision | HBR | | 670% ROI with conversational chatbots | Forrester / Drift | | 47% completion rate when asking one question at a time | Typeform | | 41% of chatbots are used for sales (#1 use case) | Intercom | | 71% of B2B buyers value a deep needs diagnosis | LinkedIn | | 62% of customers prefer chatbots over waiting for an agent | Salesforce | ## Conclusion: The Future of AI Chatbot Conversion Is Proactive Most chatbots on the market today are reactive. They answer well, but they don't go further. They don't ask, they don't qualify, they don't guide. The sales data is consistent: asking the right questions improves outcomes across the board. From Gong (hundreds of thousands of calls analyzed) to Gartner (6,000 sales reps), to Harvard Business Review (millions of leads), the evidence points in the same direction. Those who ask convert more, build more trust, and lose fewer deals to buyer indecision. If your chatbot plays a commercial role — capturing leads, recommending products, qualifying visitors — it's worth considering whether it should do more than just answer questions. A proactive chatbot approach isn't just smarter sales strategy; it's what the data demands. It's not about bombarding the user with questions. It's about knowing when one good question can turn a casual visit into a real conversation — and a real conversion. At [Bravos AI](https://bravos-ai.com/) we're building exactly this: chatbots with two modes (reactive and proactive) so every business can choose the behavior that best fits their needs. ### Want a chatbot that knows how to ask? Two modes: reactive and proactive. Choose what fits your business. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # How to Write a System Prompt for Your Business Chatbot **Category:** Practical guide | **Read time:** 22 min | **Date:** June 25, 2026 | **URL:** https://bravos-ai.com/en/blog/system-prompt-business-chatbot The **system prompt** is the single piece of text that decides the most about how good a business chatbot really is, and paradoxically it is the piece most people pay the least attention to. They compare platforms by price, integrations and widget look-and-feel, then write a three-sentence system prompt along the lines of «you are an assistant for my company, reply nicely» and end up complaining that the bot makes things up, drifts off-topic or sounds robotic. This guide is the opposite: a practical, no-hype, technical guide for writing a system prompt that actually works in production. It covers the real anatomy of a good prompt, a copyable base template, the variations that change by sector, how to personalize it by behavior, how to iterate it when it starts failing, and the most typical mistakes with concrete before/after examples. The recommendations are based on Anthropic's official documentation, the work of Hamel Husain (one of the world's top voices in AI product evaluation, who trains teams at OpenAI, Anthropic, Google and Meta), and what we've learned shipping prompts in production at Bravos AI. Table of contents - [Why the system prompt is the most underrated piece](#importance) - [Quick summary](#summary) - [Anatomy: the 7 blocks that matter](#anatomy) - [The base template (copy-paste, commented)](#base-template) - [Sector variations](#by-sector) - [Behavior personalization](#personalization) - [Rules: what the bot should NOT do](#rules-no) - [Handling «I don't know» (anti-hallucination)](#dont-know) - [How to iterate when it starts failing](#iterate) - [Common mistakes (with before/after)](#mistakes) - [Why a longer prompt is not better](#length) - [How to tell if your prompt is working](#metrics) - [How we do it at Bravos AI](#bravos) - [Wrap-up and FAQ](#wrap-faq) ## Why the system prompt is the most underrated piece of a business chatbot The system prompt is the block of instructions that the language model (GPT, Claude, Gemini or whichever you use) receives **before** reading any message from your customer. It is where you tell it who it is, what it does, what information it has access to, how it should behave, what it should not do, and how to respond when it does not know something. Everything your customer experiences in the chatbot is filtered through that block first. The consequence is direct: **a weak system prompt ruins any platform, no matter how good it is**. You can have the best model in the world (Claude Opus 4.8, GPT-5.5), the best knowledge base and the best semantic search, but if your prompt only says «you are an assistant for my company, help the customer», you're going to get inconsistent, robotic, off-tone answers and, worst of all, made-up ones when the bot doesn't know. The reverse is also true: a good system prompt lifts an average model a lot. We've seen chatbots built on top of GPT-4o-mini (a cheap model) outperform chatbots running GPT-5.5 when the prompt is well written. That happens because the model spends its capacity doing what you explicitly asked, rather than improvising. Think of Claude as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result. The same idea applies to every other frontier model. The system prompt is the welcome handbook you hand to the new employee on day one: the clearer and more precise it is, the less you'll have to fix later. ## Quick summary If you're in a hurry, here's what your system prompt needs to have. The rest of the article develops each point with templates and examples: - **A clear role**, not a generic one. Not «you are an assistant», but «you are the virtual assistant of [specific company], specialized in [specific area]». - **Business context** even if you have RAG. Document search does not replace giving the model a paragraph on what your company actually does. - **Explicit rules in the positive form**. Telling it what to do works better than listing prohibitions. - **Explicit tone and verbosity**. If you don't tell it, the model picks for you, and usually picks poorly (too formal or too long). - **A clear anti-hallucination instruction**. This is the single most important rule: what to do when it doesn't have the information. - **How to hand off to a human or capture contact**, with the exact format. - **Iteration based on real conversations**. The first prompt isn't the good one. The good one is the one you've been refining for 2-3 months while reading what people actually ask. ## Anatomy of a business system prompt: the 7 blocks that matter A well-built system prompt is not free-form writing, it's an **ordered sequence of blocks**. Each block has a specific function and benefits from being placed in a specific order. This is the structure that has worked best for us in production and that lines up with both Anthropic's and OpenAI's recommendations: - **Role and identity.** Who the bot is and which company it represents. One sentence, two at most. - **Business context.** What your company does, who it serves, geographic scope, languages. Three or four concrete sentences. - **Data and sources.** What information the bot has access to (catalog, FAQs, uploaded docs) and, just as importantly, what it does **not** have (stale prices, other customers' personal data, case-specific legal advice). - **Positive behavior rules.** How it should reply: tone, format, typical length, language, emoji usage, escalation. - **Explicit rules of what NOT to do.** What it should not make up, not reveal, not promise. Few but firm. - **Handling «I don't know».** What to say and what to do when it doesn't have the information (this block has the biggest impact on perceived quality). - **How to escalate to a human** or how to capture contact data when it applies. Order matters. Anthropic states this explicitly in its [prompting guide](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices): *«Put long documents and inputs near the top of your prompt, above your query, instructions, and examples. Queries at the end can improve response quality by up to 30% in tests with complex multi-document inputs»*. For our case (system prompt + user chat), this translates to: context and data first, behavior instructions next, and the rules of what not to do last. ## The base template (copy-paste, commented) Here's a base template that works for almost any business chatbot. It's built to be copied, with brackets you replace, and used as a starting point. In the next sections we'll add sector variations and behavior tweaks on top. For a concrete sector example: how a [law firm chatbot](https://bravos-ai.com/blog/chatbot-para-abogados) is configured so it doesn't give legal advice. Two practical notes on this template. One: **the ALL-CAPS section headers are not decorative**; modern models pick them up as separators and follow instructions better when the prompt is well-chunked. Anthropic recommends XML-style tags ( , ) for Claude; OpenAI usually does fine with all-caps or markdown headers. Both work; what matters is that there's a clear visual separation between sections. Two: **the «what to do if you don't know» section sits before «what you should not do» on purpose**. Production testing tells us the model pays more attention to positive instructions (what to do) than to negative ones (what not to do). If you only say «don't make things up», it will still make things up more often than if you first say «when you don't know, say this and do that». ## Sector variations: the lines that change by business type The base template gets you 80% of the way. The remaining 20% are 3 to 7 sector-specific lines that the model can't infer on its own. Below, the variations we add at Bravos AI to the «HOW YOU SHOULD REPLY» section depending on the client's business type. ### System prompt for an e-commerce chatbot ### System prompt for a restaurant chatbot ### System prompt for a professional services chatbot (clinic, law firm, tax advisory) ### System prompt for a SaaS or technical support chatbot The philosophy is the same in all four cases: the base template defines the general behavior; the sector variations cover the **typical cases where the model, if you say nothing, will improvise badly**. Adding 5 sector-specific lines to your prompt prevents about 80% of the weird situations you'd otherwise see in the first weeks. ## Behavior personalization: how you want your bot to act Beyond sector, there are **personality and behavior decisions** that depend on how you want your bot to be perceived. These are the levers with the biggest impact and how they're expressed in the prompt. Some of the functionalities that follow (structured contact capture, robust language detection, advanced metrics) live **outside** the system prompt on serious platforms, handled by dedicated tools that the bot operator configures without touching the prompt. We cover how to give the model these instructions via prompt because it's useful in three cases: if you're building the bot by hand without a platform, if you're evaluating platforms and want to know what a good one should solve for you before you commit, and because many platforms combine the two (explicit configuration in the panel + reinforcement in the prompt). Where a serious platform has a better solution than the pure prompt instruction, we flag it in the corresponding subsection. ### Tone: friendly vs formal The difference between «you are a professional assistant» and «you are a friendly assistant who speaks like a real person» is huge. If you don't specify, the model picks a neutral corporate tone that reads like an instruction manual. Be explicit: ### Verbosity: short and direct vs explanatory Modern models tend to ramble. If you don't set a ceiling, they write three paragraphs where three sentences would have done. For WhatsApp chatbots or small widgets, low verbosity is almost always better. Anthropic explicitly recommends *«telling the model what to do instead of what not to do»*. Instead of «don't use long lists», say «when you list, max 5 items». The model follows it much better. ### Language: automatic detection vs forced language If you have an international customer base, automatic detection is most useful. If your business is single-market and single-language, forcing the language can be worth it to avoid weird drifts. **Solving language through the system prompt is the workaround, not the right answer.** Serious platforms don't leave it to the model to decide which language to reply in; they handle it outside the prompt, with explicit detection (by domain, path, browser, widget configuration or similar signals) and predictable rules the bot operator can review and modify. Leaving it solely to the model works 90% of the time but fails in edge cases: customers who mix languages in a single sentence, very short messages with no clear signal, or conversations where a topic switch triggers an unwanted language switch. If your chatbot is going to serve a serious multilingual audience, ask your platform how it handles this before signing up, and don't rely on the prompt instruction as the only defense. ### Contact capture and human handoff This is where each company has its own logic. Some want to capture contact whenever they can; others only when strictly necessary; others never. A well-placed line in the prompt changes the behavior completely: **Capturing contacts with loose instructions in the prompt is the workaround, not the right answer.** Serious platforms handle it with a structured capture tool that surfaces naturally inside the conversation when certain conditions you defined are met (the customer shows interest, asks for a quote, requests a call, etc.). That tool validates the data (well-formed email, real phone), prevents loss if the conversation gets cut off mid-way, and logs the contact in the operator's dashboard as a structured, filterable lead, not as a loose sentence from the bot you later have to dig out of the transcript by hand. If you're going to depend on chat to capture contacts, ask your platform how it does this before signing up; the prompt-only version will get you through the first days but you'll lose contacts along the way. ### Contact info and external routes If you want the bot to send the customer to a specific destination (booking page, appointment form, WhatsApp number, web form), tell it explicitly and give the exact URL: ### Bot persona: does it have a name? does it introduce itself? A minor decision that changes how the product feels. If you give the bot a name, say so in the prompt and tell it when to introduce itself: That last instruction is not optional. Starting August 2, 2026, the **EU AI Act** requires chatbots to identify themselves as AI systems when the customer asks. We cover the legal angle in detail in [this guide on the legal liability of your chatbot](https://bravos-ai.com/blog/chatbot-responsabilidad-legal). ## Explicit rules: what the bot should NOT do Prohibitions are the shortest section of the prompt and the one most people over-engineer. People try to shield the bot with 20 negative rules and the model ends up ignoring half. Few, firm, and well-placed rules work better: Rule 6 is what's known as **prompt injection defense**. For a typical business chatbot (reactive support, no autonomous agents, no web browsing), modern models (GPT-5.x, Claude Opus 4.x, Gemini) are reasonably resistant. For reference, [Repello AI](https://repello.ai/) reported in 2026 that Claude Opus 4.8 fails about 5% under sustained adversarial pressure, and GPT-5.x around 14%. For typical business chatbot scenarios, rule 6 plus a few careful response templates are enough. The situation changes if your chatbot **indexes user-generated content** (forums, reviews, comments). That's where *indirect prompt injection* shows up: someone posts a review with hidden instructions, your RAG indexes it and then serves it to the model. If that's your case, a good reference is the work of [Simon Willison](https://simonwillison.net/), who coined the term and maintains an up-to-date bibliography of defenses. For business chatbots whose knowledge base is the company's own controlled documents, this risk is marginal. ## Handling «I don't know»: the single most important rule Of all the prompt's instructions, the one with the highest impact on perceived quality is **how the bot behaves when it doesn't know the answer**. It's the difference between a bot that earns trust and a bot that loses customers by making things up. The minimum instruction that works, tested in production: Why this version works and «don't make things up» alone doesn't: **the model needs an explicit alternative to «making things up»**. If you only forbid invention, in practice it will still invent frequently because it lacks instructions for what to do instead. Giving it a script («acknowledge, offer alternative, escalate») makes it follow. If your chatbot uses RAG (retrieval + LLM), also add this critical instruction: This is called *grounding* and it's what separates a serious chatbot from one that makes up hours, prices or policies because «they sounded right». Recent research on RAG in production (a review of 12 enterprise implementations published in 2025) shows that **in 8 out of 12 cases the system cited a misleading document because the embedding overlapped with the query**. Explicit grounding mitigates this and forces the model to lean only on what it's actually given. ## How to iterate your prompt when it starts failing The first system prompt is never the good one. The good one is the one you've been iterating for weeks with real conversations. [Hamel Husain](https://hamel.dev/blog/posts/evals-faq/), one of the world's top voices in AI product evaluation (he trains teams at OpenAI, Anthropic, Google and Meta), proposes a process we'll call «errors first»: **before building automated evaluators, look at the real data**. The simplified process, adapted to a business chatbot, goes like this: - **Review 50-100 real conversations** from your bot. Not 5, not 10. Minimum 50. If you've been running it for a short time, wait until you have them. - **Categorize the failures.** Note each case where the bot didn't reply well: hallucination, tone drift, didn't answer when it should have, answered when it shouldn't have, escalated poorly, didn't escalate when it should have. Count cases per category. - **Identify the 2-3 most frequent failures.** Usually 80% of the issues come from 2 or 3 categories. The rest can wait. - **Ask yourself: did I actually tell the bot not to do this?** In most cases the answer is no. Husain puts it bluntly: *«engineers often discover their LLM doesn't meet preferences they never actually specified — shorter responses, specific formatting, step-by-step reasoning. Start by fixing those obvious gaps»*. - **Add a concrete instruction to the prompt** that addresses each of the frequent failures. One instruction, one sentence, as explicit as possible. - **Re-test against the conversations that were failing.** If they no longer fail, keep the change. If something new fails, adjust. - **Repeat every 2-4 weeks.** Before building an evaluator, spend 30 minutes reviewing 20-50 real outputs whenever you make significant changes. Don't wait for a perfect evaluation system: start with manual inspection. And a counterintuitive data point from his work: if your prompt passes 100% of your tests, your tests are probably too soft. A 70% pass rate tends to indicate a more meaningful evaluation. ## Common mistakes (with real before/after) These are the most frequent mistakes we've seen reviewing customer prompts and our own. For each one, the before (bad version) and the after (the fix). ### Mistake 1: vague role Symptom: the bot sounds generic, doesn't convey the brand, feels like any other assistant. ### Mistake 2: prohibitions only, no alternatives Symptom: the bot keeps making up prices or information despite the prohibitions. ### Mistake 3: unspecified tone Symptom: the bot sounds like a corporate manual, too formal or impersonal. ### Mistake 4: handoff without format Symptom: the bot says «I'll connect you with the team» but doesn't capture the data, so the team has no way to reach the customer. ### Mistake 5: ignoring grounding when there's RAG Symptom: the bot sometimes gives correct info and sometimes makes it up, with no clear pattern. ## Why a longer prompt isn't better (the length paradox) A common wrong intuition: «the more instructions I give the bot, the better it will perform». In practice, past a certain point the opposite happens. **The longer the system prompt, the more likely the model will ignore some instructions**. This is well documented in LLM-in-production literature and known as instruction following degradation. The causes are several and they stack: - **Attention is unevenly distributed.** Models allocate attention non-uniformly across the prompt. Instructions in the middle get less attention than those at the beginning or end (the «lost in the middle» effect, described by Stanford in 2023 and still relevant). - **Contradictory instructions.** The more text, the easier it is for two instructions to overlap or contradict each other without you noticing. The model picks one and discards the other, usually the earlier one. - **Redundant rules that cancel out.** Three prohibitions that say almost the same thing are not three times stronger; the model often treats them as a single fuzzy rule and complies with it worse. - **Token cost.** Each conversation spends the entire system prompt. A 2,500-word prompt in a chatbot with thousands of conversations a month shows up in the bill. The practical rule we've validated in production: **if your prompt goes past 1,500 words, it's almost certainly duplicating rules or covering cases that should live outside the prompt** (in the knowledge base, in the platform logic, or as separately loaded examples). The healthy range for a typical business chatbot is between 400 and 1,200 words. Put another way, the system prompt is just the *instructions* layer within a larger set —retrieved data, tools, history— that the model receives when it answers: managing that whole set well is what's called [context engineering](https://bravos-ai.com/blog/context-engineering). Three concrete ways to keep the prompt short without losing quality: - **Move factual information to context, not to the prompt.** Prices, hours, conditions, product lists, FAQs, etc., go in uploaded documents (the context the RAG retrieves per query), not in the system prompt. The prompt defines *behavior*; the context delivers *data*. - **Rewrite to condense.** If a rule takes 4 lines, try to reformulate it in 1-2 with the same force. Short, firm instructions are followed better than long explanations. - **Delete rules that don't respond to an observed failure.** If you added an instruction «just in case» and don't remember the real case that motivated it, it's probably extra. Every rule in the prompt should exist because it resolves a concrete problem you saw in production. ## How to tell if your prompt is working: metrics and signals There's no single metric that tells you «your prompt works». What's useful is combining quantitative and qualitative signals. These are the ones that tell us the most in production: - **Self-service resolution rate** (sometimes called *deflection rate*): the percentage of conversations where the bot resolved without requesting human escalation. A well-tuned reactive support bot usually sits between 60% and 85% depending on the sector. - **«I don't know» conversations**: how many times a week the bot says it doesn't have the information. If the number grows, it's a sign your knowledge base is falling short (opportunity: add content), not that the prompt is failing. - **Escalation rate**: the percentage of conversations where the bot captures contact and escalates. If it's very low (<5%) when you expected more, review the escalation instruction. If it's very high (>40%), check if the bot is escalating cases it could resolve. - **Average response length**: if your replies average more than 80 words, the bot is probably rambling. Tune the verbosity instruction. - **Cases of «the bot made it up» reported by customers or by manual review**: absolute zero is impossible, but the goal is to drive it under 1% of conversations. - **Weekly qualitative reading of 10-20 random conversations**: no metrics dashboard replaces this. Block half an hour on Friday to read real conversations. Don't try to sell your team on evals. Instead, show them what you find when you look at the data. A real conversation where the bot screwed up is worth more than ten slides of metrics. ## How we do it at Bravos AI A practical note if you're using or evaluating Bravos AI. The base template in this article is very similar to the one we use internally as a starting point for each customer. What our panel adds on top: - **Every bot starts from an internal base template** that already covers identity, tone, response style and anti-hallucination rules. On top of that, on the Starter plan you add your *custom instructions* from the bot panel. And from the PRO plan up you rewrite the entire system prompt (and the tone), not just append instructions. - **Custom instructions get injected with high priority** inside the prompt, not as a loose note — the model treats them with more weight than the generic content of the base template. - **Every change can be tested with a built-in test widget** before applying it to production. You can launch test conversations as if you were a customer. - **The full conversation history** is available so you can do the qualitative review Husain recommends without having to export anything. - **Catalog search is decoupled from the LLM**: structured filters (size, color, price, stock) go through an exact path, not through semantic embedding. That keeps price or stock hallucinations to a minimum without you having to write that instruction in the prompt. We explain this in depth in [why your AI chatbot can't find products in your catalog](https://bravos-ai.com/blog/chatbot-catalogo-productos). - **Multilingual without extra instruction**: language detection and response handling is done at the platform layer. If you want to force a single language, you do it by toggling an option, not by writing it in the prompt. - **Structured contact capture**: when the bot detects interest based on the instructions you gave it (quote, callback, follow-up), it opens an integrated mini-form inside the conversation with the fields you defined. The customer fills it naturally, the data arrives validated to the panel as structured, filterable contacts, not loose sentences from the bot you have to dig out of the history. - **You can refine the prompt from your own Claude or ChatGPT.** With [our MCP connector](https://bravos-ai.com/conector-ia-mcp), the assistant you already use carries inside the same writing guide we use for these prompts: it helps you rewrite the bot's instructions with that very judgment, without leaving the chat. - **7-day PRO trial** to try all of the above. We notify you before charging and if you cancel before day 7 you pay nothing. ## In summary The system prompt is the highest impact-to-effort lever you have in a business chatbot. The difference between a three-sentence prompt and one that covers the 7 blocks in this guide is huge: it changes tone, consistency, reliability, perceived quality. And the best part: it doesn't require switching platform or model, just a couple of hours of writing it well and another couple of iterating it on real data afterwards. If you had to keep five principles: - **Explicit role and identity**, not generic. - **Instructions in the positive form** (what to do) over prohibitions (what not to do). - **How to handle «I don't know»** is the single most important rule: define the explicit script (acknowledge, offer alternative, escalate) and block the model's prior knowledge if you use RAG. - **Base template + 5 sector-specific lines + 3-5 behavior lines**. Don't reinvent the prompt from scratch for each bot. - **Iterate on real conversations**, not in the abstract. 50-100 real conversations every 2-4 weeks, categorize failures, fix the 2-3 most frequent. When you're ready to apply it to a real chatbot, [you can try the Bravos AI panel for 7 days on the PRO plan, free](https://app.bravos-ai.com/register?lang=en), write your own full system prompt (or just add instructions on top of the base template) and see how the behavior changes instantly with the built-in test widget. ## Frequently asked questions ### What is a system prompt? It's the block of instructions that the language model (GPT, Claude, Gemini, etc.) receives before reading any user message. It defines who the bot is, what it does, what information it has access to, how it should behave and what it should not do. It's the piece with the biggest impact on the quality of a business chatbot. ### What's the difference between a system prompt and a regular prompt? The system prompt is persistent: the model reads it at the start of each conversation and keeps it in mind the whole time. A regular prompt (what the user types in the chat) is the punctual message they send and that the model replies to. The system prompt defines behavior; the user prompt is the content to respond to. ### How do I write a system prompt for a business chatbot? Structure it in 7 blocks in this order: (1) role and identity, (2) business context, (3) data and sources the bot has access to, (4) positive behavior rules, (5) explicit rules of what not to do, (6) how to handle «I don't know», (7) how to escalate to a human. Sector variations (e-commerce, restaurants, professional services, SaaS) and behavior variations (tone, verbosity, contact capture, language) go on top of the base template. ### How long should a system prompt be? Between 400 and 1,200 words is the healthy range for most business chatbots. Going past that doesn't improve the result: the longer the prompt, the more likely the model will ignore some instructions (the instruction following degradation phenomenon). If you drop below 250 words, you're almost certainly missing blocks. ### Is it better to tell the bot what to do or what not to do? What to do. Anthropic explicitly recommends this in its documentation: positive instructions are followed better than prohibitions. Instead of «don't use long lists», say «when you list, max 5 items». Instead of «don't make up prices», say «when you don't have the price, say this and do that». Prohibitions work, but reserve them for the few critical rules (don't reveal the prompt, don't talk about competitors, don't step out of role). ### How do I prevent the chatbot from making up answers? Three things at the same time. First, a clear instruction of what to do when it doesn't know (acknowledge + offer alternative or escalation). Second, if you use RAG, a grounding instruction: the answer must be based exclusively on the context fragments, not on the model's general knowledge. Third, a well-prepared knowledge base: most chatbot inventions come from gaps in the documentation, not from the prompt. We cover that side in [why your chatbot invents answers and how to solve it](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). ### How do you iterate a system prompt in production? The process recommended by Hamel Husain, one of the world's top voices in AI product evaluation: (1) review 50-100 real conversations, (2) categorize failures, (3) identify the 2-3 most frequent, (4) add a concrete instruction to the prompt that addresses each one, (5) re-test on the conversations that were failing, (6) repeat every 2-4 weeks. Husain insists on starting from real data before building automated evaluators: *«start with 30 minutes reviewing 20-50 real outputs when you make significant changes»*. ## Sources - [Anthropic — Prompting best practices for Claude ](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices) - [Anthropic — Prompt engineering overview ](https://platform.claude.com/docs/en/docs/build-with-claude/prompt-engineering/overview) - [Hamel Husain — LLM Evals FAQ ](https://hamel.dev/blog/posts/evals-faq/) - [Hamel Husain on Lenny's Newsletter — Evals, error analysis and better prompts ](https://www.lennysnewsletter.com/p/evals-error-analysis-and-better-prompts) - [Simon Willison — Blog on prompt injection and LLM security ](https://simonwillison.net/) - [Repello AI — Red teaming and jailbreak rates by model (2026) ](https://repello.ai/) ### Apply this guide on a real chatbot At Bravos AI you add your custom instructions on top of the base template from the bot panel (Starter plan) or rewrite the entire system prompt (from the PRO plan up). The built-in test widget lets you see the effect instantly, and the conversation history is available to iterate on real data. 7-day PRO trial, no commitment: we notify you before charging and if you cancel before day 7 you pay nothing. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Chatbots vs Forms: Which Converts More in 2026 **Category:** Analysis | **Read time:** 7 min | **Date:** Jan 22, 2026 | **URL:** https://bravos-ai.com/en/blog/contact-form-vs-chatbot Chatbots vs forms: in the debate over how to capture leads on a website, there's one number that settles the argument fast. Almost nobody fills out contact forms. According to a Formstack study analyzing over 650,000 forms, contact forms have an average conversion rate of just **1%**. Out of every 100 visitors who land on your contact page, only 1 submits the form. That's the lowest form conversion rate across every type analyzed—contest forms hit 35%, and lead generation forms reach 17%. LinkedIn reported a similar finding: **81% of B2B tech users abandon gated forms** when they encounter content that requires registration. The form abandonment rate is staggering. In response, many companies are turning to AI chatbots for lead capture. But do chatbots actually convert better than web forms? Let's look at what the data says about the chatbots vs forms debate. ## Why Contact Form Conversion Rates Are So Low A contact form demands effort without offering anything immediate in return. The user has to: - Find the contact page (often buried in navigation) - Figure out what to write in the "message" field - Hand over personal data to a company they don't know yet - Wait for a response that might take hours—or days Each step adds friction. And friction kills conversions. HubSpot's analysis of 40,000 forms found that **reducing form fields from 4 to 3 increased conversions by 50%**. Adding a phone number field alone can reduce your form conversion rate by up to 5%, because users fear unsolicited sales calls. The fundamental problem is that a web form is a cold, asynchronous channel. The visitor doesn't know when—or if—they'll hear back. ## What Analysts Say About AI Chatbots for Lead Generation Here's the latest data from Gartner (2024–2025) on the chatbot vs web form landscape: **Real-world adoption:** In 2024, only 5% of companies had deployed generative AI chatbots. Another 11% were running pilots, and 44% were still exploring options. There's plenty of buzz around lead generation chatbots, but actual deployment remains low. **Documented success story:** Solo Brands (retail) implemented a generative AI chatbot and increased query resolution from 40% to 75% without human intervention. Gartner published this as a case study in 2025. **2029 prediction:** Gartner estimates that agentic AI will autonomously resolve 80% of common customer service issues, cutting operational costs by 30%. **Customer preferences (July 2024):** 64% of customers said they'd prefer companies NOT use AI for customer service. This sounds contradictory, but the explanation is straightforward: most chatbots people have experienced are terrible. They give generic answers, miss context, and create more frustration than value. This 2024 data reflects a reality that has already shifted. AI models have improved enormously—better contextual understanding, fewer hallucinations, more natural responses. A [well-implemented chatbot](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo) today is fundamentally different from what existed 18 months ago. ## Chatbots vs Forms: Why Chatbots Win on Lead Capture Even without perfect benchmarks, the logic behind chatbot conversion rates beating form conversion rates is compelling: **Proactive engagement.** A contact form sits passively on a page and waits. A chatbot can proactively appear and offer help the moment it detects visitor interest—dramatically improving your chances of capturing that lead. **Immediate response.** The visitor gets something instantly, even if it's an automated answer. With a form, there's total silence until someone manually reviews it. For lead generation, speed is everything. **Lower perceived friction.** Typing in a chat window feels like a conversation. Filling out form fields feels like paperwork. This difference in perception directly impacts your chatbot conversion rate. **24/7 availability.** An AI chatbot for lead capture works at 3 AM. A contact form does too, technically—but nobody will respond until the next business day. **Real-time lead qualification.** A chatbot can ask qualifying questions to understand what the visitor needs and route them to the right resource. A form captures data but resolves nothing in the moment. ## Chatbots vs Forms: When to Use Each **A contact form still makes sense when:** - You need specific structured data (technical quotes, formal requests, RFPs) - Your internal process requires information in a particular format - The user has a complex inquiry that requires detailed written context **A chatbot delivers better results when:** - You receive many repetitive questions that can be answered automatically - Your conversions depend on response speed (the visitor is comparing options) - You want to capture leads outside business hours - Your audience skews younger (Dashly data: 92% of users under 30 prefer chat) The combination usually works best: a chatbot for first contact and frequently asked questions, with the option to escalate to a human agent or form for complex cases. In sectors like legal this matters: a [law firm chatbot](https://bravos-ai.com/blog/chatbot-para-abogados) qualifies and captures the contact with the whole conversation, but leaves what touches the case to the attorney. This hybrid approach maximizes both your chatbot conversion rate and your ability to handle nuanced requests. ## What a Lead Generation Chatbot Needs to Actually Convert Not all chatbots are equal. A poorly designed chatbot can perform worse than a basic form. Here's what separates a high-converting AI chatbot for lead capture from a frustrating one: **Answers grounded in your actual data.** A chatbot that gives generic responses creates frustration. According to Gartner, 64% of users disengage when they feel the bot doesn't understand their needs. The chatbot must know your business—your products, pricing, hours, policies. Technologies like RAG (Retrieval Augmented Generation) let the bot query your real information before responding, which directly improves your chatbot conversion rate. **No hallucinations.** Generative AI chatbots can fabricate information. For lead generation, you need a bot that uses your data and admits when it doesn't know something—rather than inventing an answer that creates problems downstream. **Data capture woven into the conversation.** The goal is still collecting an email or phone number, but it should happen naturally within the chat flow. Not as a disguised form that blocks the conversation until every field is filled. **Catalog and database integration.** If you sell products, the bot should be able to search your catalog, filter by features, and provide current pricing. This requires integration with your data sources—CSV files, databases, or APIs. **Brand-appropriate tone.** A chatbot for a law firm shouldn't sound like one for a skate shop. The tone needs to match your brand voice for the experience to feel authentic. **Lightweight installation that won't slow your site.** If the chatbot takes forever to load or consumes heavy resources, it hurts user experience. It should be a lightweight script you can integrate in minutes. **Knowing when to hand off to a human.** Gartner estimates 80% of routine queries can be automated. But the remaining 20% need human intervention, and the chatbot must recognize when to escalate. At [Bravos](https://bravos-ai.com/), we built our chatbot to solve exactly these problems: it uses RAG to answer with your real data, can query CSV-based catalogs, integrates a lead capture form naturally within the conversation without blocking the chat, and installs with a lightweight snippet that won't slow your website. And if you use Claude or ChatGPT, with [our MCP connector](https://bravos-ai.com/conector-ia-mcp) the assistant fine-tunes the bot applying the same experience we use to build it. If you want to try it, you have [7 days free on the full PRO plan](https://app.bravos-ai.com/register?lang=en). ## Chatbots vs Forms: The Verdict Contact forms have conversion rates between 1–7% according to the studies. Live chat and AI chatbots consistently show better numbers, though the most optimistic claims (the often-cited "3x more conversions") tend to come from vendors selling these solutions. What is proven is that **immediate response matters enormously**: according to HubSpot, 90% of customers consider instant response important or very important when they have a question. A chatbot solves that. A contact form does not. If your website depends on lead generation and you only have a contact form, you're almost certainly leaving opportunities on the table. Not because chatbots are magic, but because forms put the entire burden on the visitor and offer nothing in return until someone manually reviews the submission. **The recommendation:** Deploy a chatbot alongside your existing form, measure for 30–60 days, and compare your own data. Your results will be worth more than any third-party statistic on chatbot conversion rates. ## Frequently asked questions ### Chatbots vs forms: which converts more? According to verifiable data, contact forms have an average conversion rate of 1% (Formstack, 650,000 forms analyzed) and 81% of B2B users abandon forms that require registration. Chatbots consistently show higher conversion rates, though the most optimistic figures (the often-cited "3x more") tend to come from vendors selling these solutions. What is proven is that immediate response matters: according to HubSpot, 90% of customers consider it important or very important when they have a question. ### Why do contact forms convert so poorly? A contact form demands effort without offering anything immediate in return. The visitor has to find the contact page, decide what to write, hand over personal data to a company they do not know yet and wait for a response that may take hours or days. Each step adds friction. According to HubSpot, reducing form fields from 4 to 3 increases conversions by 50%. The underlying problem is that a web form is a cold, asynchronous channel: the visitor does not know when, or if, they will hear back. ### When does a contact form still make sense over a chatbot? A contact form still makes sense when you need specific structured data (technical quotes, formal requests, RFPs), when your internal process requires information in a particular format, or when the user has a complex inquiry that requires detailed written context. For those cases, the combination chatbot + form usually works better than either alone: chatbot for first contact and frequent questions, form for complex cases. ### What does a chatbot need to convert better than a contact form? Answers grounded in your real data through RAG (Retrieval Augmented Generation), no hallucinations, data capture woven naturally into the conversation without blocking it, integration with catalogs and databases (CSV files, APIs), brand-appropriate tone, lightweight installation that will not slow your site, and the ability to hand off to a human when the conversation calls for it. According to Gartner, 64% of users disengage when the bot does not understand their needs, so the bot must know your business. ### Want to test the difference? Set up your chatbot and compare it with your current form on your own site. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # WhatsApp Chatbot with AI: The Complete Guide for 2026 **Category:** Complete guide | **Read time:** 20 min | **Date:** June 16, 2026 | **URL:** https://bravos-ai.com/en/blog/whatsapp-chatbot WhatsApp has **over 3 billion monthly active users** and, in most of Europe, Latin America, India, and increasingly the US, it has become the default channel customers reach out to before email. Setting up a **WhatsApp chatbot** is no longer an experiment: it is the most direct way to stop losing sales at 11pm or on a Saturday afternoon, on the channel your customer already uses every day. But the topic is full of confusion. WhatsApp Business this, WhatsApp Business API that, templates, the «24-hour window», and the famous «free chatbot» that turns out not to be free. This guide is the no-hype version: what a WhatsApp chatbot really is, what AI can actually do today, how much it costs, and how to set one up without wasting two weeks. We have spent months building our own WhatsApp chatbot at Bravos AI, we are a verified Meta Tech Provider, and our customers can connect their WhatsApp number in minutes, not days. Table of contents - [60-second summary](#summary) - [What is a WhatsApp chatbot](#what-is) - [WhatsApp Business app vs Business API](#business-vs-api) - [What an AI chatbot can do (and what it can't)](#can-and-cannot) - [Are there free WhatsApp chatbots?](#free) - [How it connects: signup, templates, 24-hour window](#how-it-connects) - [How to build a WhatsApp chatbot step by step](#step-by-step) - [How much it really costs](#pricing) - [AI chatbot vs button flows](#ai-vs-buttons) - [Use cases by sector: two examples](#use-cases) - [Common mistakes when launching](#mistakes) - [When to build it yourself with n8n, Make or Zapier](#n8n-make-zapier) - [How we do it at Bravos AI](#bravos-ai) - [In summary](#in-summary) - [Frequently asked questions](#faq) ## 60-second summary If you're in a hurry, here's what you need to know. The rest of the article goes deep on each point with honesty: - **A WhatsApp chatbot is a program that replies to your customers on your WhatsApp Business number, 24/7, without anyone having to be there.** If it has AI behind it, it doesn't rely on rigid buttons: it understands the question in plain language and answers using your business information. - **WhatsApp Business (the mobile app) and WhatsApp Business API (the technical connection used by chatbots) are two different things.** The app is free and operated by a person; the API is for connecting software like a chatbot and has its own per-message cost. - **Since July 2025, Meta no longer charges for conversations started by the customer.** If your chatbot only replies (no outbound marketing), Meta's cost is basically zero. You only pay the chatbot platform. - **The only thing that's truly free is the fixed away message of the official WhatsApp Business app.** Anything with real conversational AI costs money. The question is how much. - **A WhatsApp chatbot with real AI costs between $23 and $200/month** depending on platform and volume, without counting outbound marketing messages. - **Setup takes from minutes to several days** depending on the platform. If the platform is a verified Meta Tech Provider (we are), it's minutes. If they work through a BSP intermediary, it can be days. If you build it yourself with n8n or Make, plan for 13–26 hours the first time. ## What is a WhatsApp chatbot A **WhatsApp chatbot** is a program that receives the messages customers send to your business WhatsApp number and replies on its own, with no person typing on the other side. If it's built with modern AI, it is not a button menu: it understands what the customer wants even with typos, in another language or in very informal wording, and replies with real information from your business. In 2026, a typical WhatsApp chatbot does three things: - **Answer frequently asked questions:** hours, prices, location, payment methods, return policy, allergens, service conditions. - **Find specific information about your business:** «what treatments do you offer for back pain?», «what does your full coverage insurance include?», «do you ship to Hawaii?». - **Capture leads and collect data so your team can take over:** when someone shows interest (or when the machine can't close the case), the bot asks for email or phone and saves that contact in your dashboard. What a serious WhatsApp chatbot **doesn't do**: - It doesn't start conversations on its own to sell you things (that's outbound marketing and it has strict rules — we'll cover it below). - It doesn't access your personal WhatsApp conversations: it only works with the WhatsApp Business number you connect. - It doesn't replace the human team on complex cases. What changes is that your team stops answering the trivial stuff. ## WhatsApp Business app vs WhatsApp Business API: the confusion everyone has This is the most confused topic in the WhatsApp chatbot world. You google «WhatsApp Business chatbot» and you find two things with almost identical names that have nothing to do with each other. Let's break it down. ### WhatsApp Business app: the mobile app It's the **app you download on your phone**, different from regular WhatsApp, designed so a sole trader or small business can have a separate number from their personal one. It includes a business profile, a basic catalogue, customer labels and quick replies. It's **free** and operated by a person at the business. What *looks like* a chatbot inside the app is actually two very limited things: - **Greeting message:** fixed text when someone messages you for the first time. - **Away message:** fixed text when someone writes outside your business hours. That's not AI, that's not a conversational chatbot, it doesn't understand what the customer says. It's an automated sign. Enough for many sole traders; not enough for a small business with two people in customer service. If you want to understand the different levels of automated reply, we cover them in [WhatsApp auto reply: 3 real options](https://bravos-ai.com/blog/responder-whatsapp-automaticamente). ### WhatsApp Business API: the chatbot connection It's the **technical channel Meta offers so external software** (an AI chatbot, a customer service dashboard, a CRM) **can send and receive messages on your WhatsApp number**. It has no app, you don't see it on a phone: it lives inside the dashboard of whatever platform you're using. The original «Business API» (the old «On-Premise API», hosted on your own server) is barely used anymore. What everyone today calls «Business API» is actually the **WhatsApp Cloud API**, the version hosted by Meta and launched in 2022. The difference between «Business API» and «Cloud API» is only where the infrastructure runs; for the end customer they are the same thing. ### At a glance | Feature | WhatsApp Business (app) | WhatsApp Business API | | --- | --- | --- | | Who operates it | A person at the business, on a phone | Software (chatbot, dashboard, CRM) | | Meta cost | Free | Free for messages started by the customer; cost per outbound template | | Real AI chatbot | No (only fixed greeting/away text) | Yes, by connecting a platform | | Business verification | Not required | Not required to reply to inbound; only if you send more than 250 marketing templates per day | | Multiple agents | Up to 5 linked devices | Unlimited from the dashboard | | Web/store integration | Manual | Automatic | **Bottom line:** if you just want an automated sign about your hours and you'll reply by hand, the app is enough. If you want a chatbot that understands questions and replies with real data from your business without anyone typing, you need to connect your number to the API through a platform. ## What an AI chatbot can do on WhatsApp (and what it can't) This is where we have to be honest. Ads promise the world; the reality of a chatbot with **real AI** on WhatsApp in 2026 is better than people think for some things and worse for others. ### What a decent AI chatbot does well - **Answer questions with your business information.** If you upload your website, PDFs or a spreadsheet, it finds the answer there and replies in plain language. Quality varies a lot by platform: the good ones don't make things up when they don't have the data; the bad ones do. - **Remember what was said within the same conversation.** If you ask the price of product X and then say «and in size L?», it knows which product you mean. This is the bare minimum a modern conversational AI should offer. - **Be available 24/7.** No hours, no holidays, no waiting. The difference between replying in 30 seconds and replying the next day is the difference between closing the sale and not closing it. - **Collect customer data when needed.** If the question needs human attention or the customer asks for a person, the bot asks for phone or email and drops the contact in your dashboard so your team can pick it up later. - **Honestly say when it doesn't know** (on serious platforms). If you haven't uploaded the return policy, it won't make one up: it says it doesn't have it. Platforms that invent answers are a real problem in the market — worth testing before signing up. ### What it doesn't do (or rarely does) - **It doesn't pick up the phone.** It doesn't replace a call for an angry customer, a complex negotiation or an emotionally delicate case. - **It doesn't book appointments in your calendar by itself** on most platforms. What it can do is share available slots, send a link to your booking system, or take the data so your team confirms the appointment. The «books your full calendar» integration exists on very few platforms and usually requires extra configuration. - **It doesn't close complex sales by itself.** For a $50 sneaker purchase, sure, it can walk the customer to checkout. For an $8,000 kitchen renovation, no. It filters interest and drops the data in your dashboard so your team can close it. - **It doesn't send messages to your customers on its own unless you configure it to.** And when outbound is configured, WhatsApp imposes the «24-hour window» (we cover it below), which limits that type of message. - **It doesn't process orders or payments by itself.** It can share your store's checkout link; it doesn't act as a payment gateway. If the platform you're looking at promises things their product doesn't do (like a fully autonomous agent that closes sales, books calendar appointments and processes payments without intervention), be skeptical and ask for a demo with your case. The technical reality of most WhatsApp chatbots in 2026 is much closer to a great receptionist than to an autonomous salesperson. ## Are there free WhatsApp chatbots? What's actually free and what isn't A lot of search volume around WhatsApp chatbots adds «free» somewhere in the query. The short, honest answer: **a conversational WhatsApp chatbot with real AI that's free forever doesn't exist**. What does exist is several things marketed as free that sometimes are and sometimes aren't. Let's separate them. ### What's actually 100% free - **The away message and greeting of the WhatsApp Business app.** Fixed text, no AI, doesn't understand the customer. The most basic thing and, for many sole traders, the only thing they need. - **Trying several platforms for a limited time.** Many offer free trials of 7 to 14 days, including ours. That lets you see if it fits before paying. It's not a «permanent free plan». ### What's marketed as free but has small print - **«Free» plans of web chat platforms (Tidio, Crisp, Freshchat, etc.).** The free plan covers their web widget chat, not WhatsApp: to connect WhatsApp Business API you need a paid plan. And even when the free plan promises AI, that AI is usually quite limited — we checked it by asking their own official chatbots in [this experiment](https://bravos-ai.com/blog/tidio-crisp-freshchat-plan-gratis). - **Twilio «sandbox» trials.** Yes, you can connect WhatsApp for free for testing, but only with numbers you have registered and with pre-defined templates. Not viable for a real business. - **Building it yourself with self-hosted n8n.** Free n8n (hosted on your own server) does save you the subscription, but it costs you the server ($5–20/month), the OpenAI or Claude model you use ($5–50/month depending on volume) and the hours you put in. We break it down in [this comparison with n8n, Make and Zapier](https://bravos-ai.com/blog/chatbot-whatsapp-alternativa-zapier-make-n8n). ### What Meta always charges for This is important and almost no one explains it: **Meta charges for some WhatsApp Business API conversations**, regardless of the platform you use. That's a Meta cost, not a platform cost. Since July 2025 the model changed: - **Conversations started by the customer (service conversations):** free and unlimited. If the customer messages you first, neither Meta nor anyone else charges for the message. - **Conversations started by your business (marketing, reminders, notifications):** these have a cost. In the US, marketing messages run around $0.025 per message; in other regions it varies. If you only use the chatbot to reply to what comes in, this cost doesn't affect you. **Honest bottom line:** if you want a real AI chatbot on your WhatsApp Business, you're going to pay between $23 and $200/month to the platform. We sit in that range: from $23/month on the Starter plan. That's what it costs to have a maintained system, officially connected to Meta, with a decent AI model behind it. «Free forever» with real AI doesn't exist. ## How it connects: signup, templates, and the 24-hour window The signup process for a WhatsApp chatbot has parts that almost no one bothers to explain, but knowing them saves you headaches. These are the concepts you'll hear sooner or later. ### Do I need to verify my business with Meta? **For a chatbot that only replies to inbound, no.** Up until 2024 you did have to pass business verification in Meta Business Manager before using the API. That's no longer the case: you can connect your number, receive messages and reply without passing any verification. Replies to customers who message first are **unlimited** in this mode. Verification is only required if you're going to use WhatsApp for **heavy outbound marketing**: without verification, Meta lets you send business-initiated templates to up to 250 unique customers in a rolling 24 hours. If you complete verification (or if your message quality is high enough that Meta promotes you automatically), that limit goes up to 100,000 per day. Because we are a **verified Meta Tech Provider**, connecting your number through the Bravos AI platform takes minutes thanks to Embedded Signup. On platforms that work through intermediaries (so-called BSPs, «Business Solution Providers»), signup can take longer. ### Message templates If your business wants to **start** a conversation with a customer (sending an appointment reminder, a shipping update, an offer), Meta requires that message to be sent with a pre-approved text, called a template. You create the template and submit it to Meta for review; they approve or reject it within hours. This is what prevents businesses from using WhatsApp as spam. **Replies to a customer who messaged you first do not need a template.** Your chatbot can reply with whatever it wants, in free language, with no approval. This is what makes a reactive chatbot (like ours) much simpler to put live than one that pushes outbound marketing messages. ### The 24-hour window When a customer messages you, Meta opens what it calls a **24-hour service window**. During those 24 hours your business (or your chatbot) can reply freely, with no template and no cost. Once 24 hours pass, if you want to message that customer again, you can only do it with an approved template. So for regular customer service, the 24-hour window is not an issue: the bot replies in seconds, well before it closes. ### Phone number quality with Meta Meta tracks the behavior of your number. If customers block you or mark your messages as spam, your «quality» rating drops and Meta can limit how many templates you can send per day. This only affects you if your platform allows outbound and you're using it: in that case, segment well before sending. If your chatbot only replies to what comes in (like ours), the quality problem doesn't apply to you, because replies to customers don't count toward that score. ## How to build a WhatsApp chatbot step by step If you have the basics down and you want to set it up, here's the short path. The long version, with the Meta panel walkthrough, is in our dedicated guide [how to build a WhatsApp AI chatbot in 2026 without code](https://bravos-ai.com/blog/como-hacer-chatbot-whatsapp). - **Decide what you want the bot to do.** Answer FAQs, search your catalogue, collect data for an appointment or for your team to take over. Keeping a tight list prevents you from drowning in features you don't need in month one. - **Gather your business information.** Website text, PDFs of your menu or catalogue, a spreadsheet with products or services, return policy. If you have an online store on Shopify or PrestaShop (or your own system with a webhook), some platforms (ours included) connect directly and the catalogue syncs automatically, without you having to maintain a separate file. A chatbot is only as good as the information you give it, as we detail in [this guide on keeping your chatbot up to date](https://bravos-ai.com/blog/mantener-chatbot-actualizado). - **Sign up for a platform.** There are dozens; we compare the 13 most relevant in [this comparison of the best WhatsApp chatbots](https://bravos-ai.com/blog/mejor-chatbot-whatsapp). The choice depends on your budget, your language needs and whether you'll do outbound marketing or just inbound support. - **Create the bot and train it on your information.** On modern platforms this means uploading files or pointing it to your website URL. The platform processes it and can start replying based on that. - **Connect your WhatsApp number.** The platform walks you through the Meta Business signup and number assignment. If the platform is a verified Meta Tech Provider (like us), the connection takes minutes. If not, it can take longer. Business verification is only required if your platform allows outbound and you later want to send templates to more than 250 contacts per day. - **Test with real conversations before exposing it to customers.** Throw weird questions at it, typos, words in other languages, scenarios where you know information is missing. Adjust as needed. - **Activate and observe the first week.** Watch what people actually ask, what the bot gets wrong, what edge cases slip through. Most platforms (ours included) let you see the full conversation history so you can refine the loaded information. ## How much a WhatsApp chatbot really costs Price depends on three things: the platform you choose, the messages you send, and, if you build it yourself, the AI model behind it. Let's break them apart. ### Platform cost Platforms with a real AI WhatsApp chatbot cover different cases. These are the typical ranges in June 2026, with examples of platforms that fit each one: | Price range | Who it fits | Platforms in that range | | --- | --- | --- | | **From $23/month** | Small business or sole trader that wants to serve WhatsApp (and, depending on plan, also their website) with real conversational AI, lead capture and multiple bots | **Bravos AI** (Starter $23, PRO $59). The rest of the market starts above $50/month for chatbots with a real LLM. | | **From $50/month** | Business that also wants to send outbound marketing templates to its base, or already lives in a specific multichannel ecosystem | Tidio with Lyro ($56.67), ManyChat ($58), Landbot WhatsApp Starter (~$90), Wati Pro ($119), AiSensy with chatbot add-on (~$125 total — the $21 base plan does not include the chatbot) | | **From $100/month** | Mid-sized company with several human agents and many messaging apps (Instagram, Messenger, Telegram, etc.) connected at once | Respond.io Team, Interakt Advanced | | **Custom (from ~$500/month)** | Enterprise: large teams, multiple seats, dedicated support, custom integrations, specific terms | **Bravos AI Enterprise** (custom pricing), Intercom Fin, HubSpot Service Hub Enterprise | The detail for each one, with prices verified directly on their official pages, is in [the comparison of the 13 most relevant platforms](https://bravos-ai.com/blog/mejor-chatbot-whatsapp). ### Meta message cost Since July 2025, conversations started by the customer are free forever. You only pay Meta when your business starts the conversation with a template. For reactive customer service, this cost is basically zero. If you run mass marketing campaigns on WhatsApp, budget around $0.025 per marketing message in the US (it varies by region and category). ### The hidden markup Some platforms **charge a markup on top of Meta's cost** for every message. Wati, Interakt and Twilio apply margins from 2.8% to 40% depending on plan and volume. Others don't apply a markup and pass Meta's price through. Worth asking before you sign; we break it down platform by platform in the comparison. For a broader pricing comparison, not just WhatsApp, see [our complete guide on how much an AI chatbot costs in 2026](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa). ## AI chatbot vs button flows: why the difference matters When you search for «WhatsApp AI chatbot» or «WhatsApp ChatGPT chatbot», what you're asking for is a chatbot that understands what the customer says in free language. That's not the same as a button menu that many platforms still sell as a chatbot. ### Button flows (the classic chatbot) It works like this: when the customer writes, the bot offers options («1) Hours, 2) Prices, 3) Book»). The customer picks one, the bot gives a fixed answer, offers more options, and so on until they leave or the flow ends. If the customer writes anything outside the menu, the bot doesn't understand and usually replies with a generic «I didn't catch that». This is what **ManyChat, AiSensy on its base plan or Landbot** do. It's not bad; for highly structured processes (surveys, polls, first-line filtering) it can work. But most of your customers don't want to navigate a menu: they want to ask what they need in their own words. ### Conversational AI chatbot (what people mean today by «AI chatbot») The customer writes whatever they want, the way they would write to a person, and the bot understands. If they ask «what time do you open on Saturday?», the bot finds that information in what you uploaded and replies. If they ask about your return policy, the bot pulls it from your policy document and explains it in its own words, not as a copy-paste of the PDF. Under the hood, what the bot does is **combine your information with a large language model** (popularly called «the engine of the chatbot»): GPT-4, Claude, Gemini or another. The platform's job is to make sure the model doesn't make things up about your business: it only passes your information as reference. This is what we do at **Bravos AI**, what Tidio does with Lyro, what Wati does with Astra, what Intercom does with Fin. The difference between them is which model they use, how well they manage your information, and what they charge. If you want to go deeper, our [rule-based vs AI chatbot comparison](https://bravos-ai.com/blog/chatbot-reglas-vs-chatbot-ia) covers it. **Honest recommendation:** in 2026, launching a brand-new chatbot based purely on buttons rarely makes sense. Customers are used to talking with ChatGPT and Google's AI. If your chatbot doesn't understand them when they ask normally, they'll abandon it by the second interaction. ## Use cases by sector: two examples To show you what this actually looks like in production, here are two example conversations of the kind that hit businesses with a connected bot. ### Retail: outdoor gear store The customer wants boots for a specific trip. Watch what happens when the bot knows the catalogue and the product details: What this bot does well: filters the catalogue by size, by use case (the trip drives the recommendation), compares two specific models by a subjective feature (wide vs narrow last), and closes by sending the direct product link. It doesn't push the sale, doesn't ask for personal data if it doesn't need to. ### Professional services: dental clinic The patient has a specific question about a treatment. Watch how the bot informs with clinic data and sends them to the booking system without booking on her behalf: What this bot does well: replies with clinic data (prices, sessions, warranties), doesn't get into recommending medical treatments, doesn't decide for the patient and, most importantly, sends her to the clinic's booking system instead of promising a reservation the machine can't confirm. If the patient would rather be called, the bot takes the phone number for reception to ring back. Other sectors that fit a WhatsApp AI chatbot well: restaurants (menu, allergens, reservations), physical retail (hours, stock, location), real estate agencies (property filters), businesses with wide catalogues (filtering and product detail), training academies (prices, dates, terms), online stores (returns, order tracking, recommendations). ## Common mistakes when launching a WhatsApp chatbot We've spent months watching small businesses, sole traders and agencies launch WhatsApp chatbots. These are the most common pitfalls and how to avoid them. ### 1. Uploading incomplete information and blaming the bot If your website doesn't list the return policy, the bot isn't going to make one up (that's a good thing). But if you launch the bot without that content, you'll have customers asking things the bot can't answer. Before going live, double-check you have everything a real customer might ask. If something's missing, write it, upload it and verify. ### 2. Trying to reuse the same number you already have in the WhatsApp Business app A WhatsApp number can only be in one place at a time: either in the WhatsApp Business app on your phone, or on the API connected to a chatbot. Not both. If you want to move the number you use in the app over to the chatbot, you have to **migrate it** (Meta walks you through it) and that wipes the history you had in the app. If you don't want to lose that history, the easiest path is to get a new dedicated number for the bot. ### 3. Treating WhatsApp like email (when your platform allows it) Only applies if your platform allows outbound template sending, not if your chatbot only replies to inbound. On those platforms, sending marketing templates to 5,000 contacts who didn't opt in ends badly: customers block, Meta drops your quality rating, and your sending limits start to shrink. If you're going to do WhatsApp marketing, segment well before sending and start with fewer contacts. ### 4. Not reviewing conversations in the first week The first 50 conversations of your bot are gold: that's where you see the real questions people are asking and the gaps in your information. If you don't look at them, you can't improve the bot. Most platforms, ours included, let you see the full conversation history and filter the ones the bot got wrong. What almost no other platform lets you do is have the AI itself do that review for you: by connecting the bot to Claude or ChatGPT with [an MCP](https://bravos-ai.com/blog/que-es-un-mcp), it's the assistant that goes through the conversations and points out the gaps. ### 5. Expecting the chatbot to fully replace the human team A well-trained chatbot handles most of the routine queries on its own. The rest still need a person. If you frame it as «I'm replacing my customer service team», you're in for a disappointment. If you frame it as «my team stops answering hours, prices and availability and spends their time on what matters», it fits perfectly. ### 6. If you're going to do outbound, not checking the per-message markup If you're only doing inbound, this one doesn't apply to you. But if your platform allows outbound and you're going to send templates, watch out: a platform at $23/month can end up more expensive than one at $60/month if the first one charges a 25% markup on every Meta template and the second doesn't. Always look at both numbers: monthly fee and per-message pricing model. ## When it's worth building it yourself with n8n, Make or Zapier If you're technical (or work with someone who is), there are several routes: automation tools like n8n, Make or Zapier connected to OpenAI or Claude's API, or a more developer-oriented bot builder like Google's Dialogflow. We did this ourselves as an exercise with n8n and we wrote it up honestly in [this comparison with n8n, Make and Zapier](https://bravos-ai.com/blog/chatbot-whatsapp-alternativa-zapier-make-n8n). The summary: **it makes sense if** you're already heavy users of that tool for other things, you have a technical team maintaining flows in production, and you need something very custom that doesn't fit any closed platform. **It doesn't make sense if** you're a sole trader or a small business without technical people, if you don't have anyone who can put in the 15-25 hours the first build takes (plus the maintenance after: images, audio, conversation memory, 60-day token expiry), or if you want it live this month and not next. ## How we do it at Bravos AI If you made it this far, we owe you a quick run-down of what we do. At Bravos AI we build AI chatbots for small businesses and larger companies on a custom plan that want to [automate customer support](https://bravos-ai.com/blog/ia-atencion-al-cliente) over WhatsApp and on their website without going through expensive agencies or building anything technical. Here's what sets us apart, told without the marketing fluff: - **Verified Meta Tech Provider.** That means connecting your WhatsApp with us takes minutes via Embedded Signup, no BSP intermediary, no 60-day token expiry. It's the official Meta model. - **Reactive bot, not proactive.** We don't start conversations on your behalf, we don't send campaigns to your base, we don't push you toward outbound marketing. The bot replies to whoever messages you. Safest way to avoid burning your number with Meta. - **Native integrations: Shopify, PrestaShop, Resales Online (real estate) and custom webhook.** We hook into your catalogue or your system directly and the bot stays current without Zapier in the middle. If your business lives in another system, the webhook brings it in without custom development. - **Exact search on your catalogue, not approximate.** If your customer asks for size 8 under $200, the bot finds exactly that and returns products that match, not «similar» ones. This sounds easy but it's much harder to get right than the market promises: most chatbots fall short the moment the customer combines multiple specific filters (size, color, stock, price range). We explain it in depth in [why your AI chatbot can't find products in your catalog](https://bravos-ai.com/blog/chatbot-catalogo-productos). - **No per-message markup on top of Meta's cost.** Whatever Meta charges is what you pay; we don't add our own margin on outbound messages. - **Multilingual out of the box.** If someone writes to you in German or French, the bot detects the language and replies in it. More in our [multilingual chatbot guide](https://bravos-ai.com/blog/chatbot-multiidioma). - **You improve it from the Claude or ChatGPT chat itself.** With [our MCP connector](https://bravos-ai.com/conector-ia-mcp), the assistant reviews the conversations and applies our best practices to tell you what to reinforce, without you opening the panel. - **7-day PRO trial including WhatsApp.** We notify you before charging and if you cancel before day 7 you pay nothing. No commitment. ## In summary In 2026, building a WhatsApp chatbot is no longer an experiment. Meta's official API is accessible to any business, customer-initiated conversations are free and unlimited since July 2025, and there are platforms with real conversational AI that work well without spending a fortune. The question is no longer *«is it free?»* but *«what do I need this bot to do, and how much does what I need cost?»*. If you only want to serve inbound (FAQs, catalogue search, contact capture), a platform with an official Meta connection like ours gets you live in minutes for **$23-59/month**. If you're also going to push outbound marketing with templates, you're looking at **$50-150/month** and a platform with a sending module. If your volume is enterprise, the floor is around **$500/month** with custom seats. The market traps are the usual ones: vendors who call something an «autonomous AI agent» when it's actually button flows with an AI-generated response inside, platforms that charge a 25% markup on every Meta template without telling you, or «free plans» that don't actually cover WhatsApp. The way to avoid them is the same as always: ask for a demo with your case, read the small print on per-message pricing, and test before you pay. When you're ready to start, [you can try the Bravos AI WhatsApp chatbot for 7 days on the PRO plan](https://app.bravos-ai.com/register?lang=en). We notify you before charging and if you cancel before day 7 you pay nothing. No commitment. ## Frequently asked questions ### What is a WhatsApp chatbot? It's a program that automatically replies to the messages sent to your WhatsApp Business number, with no one typing on the other side. If it has AI behind it, it understands what the customer asks in plain language and replies with your business information (website, catalogue, PDFs, hours). It's different from the fixed greeting and away messages of the WhatsApp Business app. ### What's the difference between WhatsApp Business and WhatsApp Business API? WhatsApp Business is the free mobile app a business downloads to have a number separate from the personal one, with a basic catalogue and fixed messages. WhatsApp Business API is the technical channel Meta offers so external software (like an AI chatbot) can send and receive messages on your number. The API has no app and is operated from the dashboard of whatever platform you use. ### What's the best free WhatsApp chatbot? A conversational chatbot with real AI that's free forever doesn't exist. What's 100% free is the fixed greeting and away message of the official WhatsApp Business app, no AI. If you want a chatbot that understands the customer and replies with real business data, you're looking at $23-200/month depending on the platform. Many, ours included, offer a free 7-14 day trial so you can see if it fits before paying. ### How do I create a WhatsApp chatbot? In short: decide what tasks you want it to do, gather your business information (website, PDFs, catalogue, hours), sign up to a platform, upload that info to train it, connect your WhatsApp number through Meta Business signup, and test before going live. For a chatbot that only replies to inbound, you don't need to pass Meta business verification. The full guide is in [how to build a WhatsApp chatbot step by step](https://bravos-ai.com/blog/como-hacer-chatbot-whatsapp). ### How much does an AI WhatsApp chatbot cost? It depends on the case. For inbound only (real AI, multichannel web + WhatsApp, contact capture), from $23/month (Bravos AI Starter and PRO at $59). If you also want to send outbound marketing templates, $50-150/month (Tidio with Lyro, ManyChat, Wati, Landbot). For mid-sized companies with many messaging apps connected at once, $100-300/month (Respond.io, Interakt). Enterprise is custom-priced from around $500/month (Bravos AI Enterprise, Intercom Fin, HubSpot). On top of that is Meta's cost per outbound marketing message (~$0.025 in the US); replies to customers who message first are free since July 2025. ### Can ChatGPT be integrated with WhatsApp? Yes, but not «ChatGPT» the app, rather the underlying OpenAI model (GPT-4, GPT-4.1, etc.) via its API. Most modern platforms, Bravos AI included, use that model (or equivalents like Claude or Gemini) under the hood. What the customer sees is a chatbot that understands well, not «ChatGPT on WhatsApp». If you want to build it yourself with OpenAI's API plus a tool like n8n, it's possible but it'll take 13-26 hours the first time. We compared it [here](https://bravos-ai.com/blog/chatbot-whatsapp-alternativa-zapier-make-n8n). ### What is WhatsApp Business API? It's the set of features Meta offers businesses to connect their WhatsApp number with external software: chatbots, customer service dashboards, CRMs. It has no mobile app of its own; it lives inside whichever platform decides to use it. Its two key features are message templates (approved text for the business to start a conversation with a customer) and the 24-hour window (during which your chatbot replies freely with no cost after a customer message). For a chatbot that only serves inbound, you don't need to pass Meta business verification. ## Sources - [Meta — WhatsApp Business Platform ](https://business.whatsapp.com/products/business-platform) - [Meta — WhatsApp Cloud API overview ](https://developers.facebook.com/docs/whatsapp/cloud-api/overview) - [Meta — WhatsApp Business pricing (July 2025 update) ](https://developers.facebook.com/docs/whatsapp/pricing) - [Meta — WhatsApp Commerce Policy ](https://business.whatsapp.com/policy) - [Meta — WhatsApp Cloud API launch (May 2022) ](https://about.fb.com/news/2022/05/whatsapp-cloud-api-launch/) ### Try the Bravos AI WhatsApp chatbot Connect it to your WhatsApp number in minutes, train it on your information in 10, and it starts replying. 7-day PRO trial, no commitment: we notify you before charging and if you cancel before day 7 you pay nothing. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Chatbot & AI Statistics: Data and Market Size (2026) **Category:** Data | **Read time:** 24 min | **Date:** July 10, 2026 | **URL:** https://bravos-ai.com/en/blog/chatbot-statistics Most "chatbot statistics" pages online share one problem: they recycle numbers from five years ago, cite each other in circles, and when you try to reach the original study, you hit a dead link. This page is built the opposite way. We opened **every figure at its primary source** —the analyst firm's press release, the company's report, the academic study, the court ruling— noted the year, and dropped anything we couldn't trace back to its origin. What follows is more than 50 statistics on chatbots and artificial intelligence in customer service, e-commerce and business adoption, each with its link and its date, grouped so you can find the one you need in seconds. And a warning almost no roundup makes: **the arrival of ChatGPT (late 2022) split the chatbot world in two**. Many of the statistics still circulating are from the era of button-menu bots, and today they mislead more than they inform. When a figure is from that earlier era, we flag it. When it's a forward projection rather than a measurement, we flag that too. **How to read this page.** Each figure carries its primary source, the year and a link. We distinguish what has been *measured* from what is *projected*, and the era of rule-based bots from the era of large language models (the "LLMs" behind ChatGPT, Claude or Gemini). You'll see wide ranges in the market-size figures: that's not an error, it's that each firm defines the market differently, so we always name the firm. Last updated: July 2026. We revise this page periodically. Index - [Key takeaways](#highlights) - [Adoption and market size](#market) - [Customer service impact](#service) - [Cost and return on investment](#cost) - [What consumers think and want](#consumers) - [E-commerce and messaging](#ecommerce) - [The generative-AI shift](#llm-era) - [Limits, risks and the human factor](#limits) - [Language and multilingual support](#language) - [Europe, Spain and Latin America](#regional) - [How we verified these figures](#methodology) - [Frequently asked questions](#faq) ## Key takeaways If you only have a minute, these are the most citable figures on the page. Each is developed in full, with its source and caveats, in the relevant section. - **75%** of customer service leaders expect that, within a few years, **80% of interactions** will be resolved without human intervention (Zendesk, 2025). - **64%** of customers would prefer companies **did not** use AI in their customer service (Gartner, survey of 5,728 people, 2024). - Klarna announced its AI assistant was doing the work of **700 agents**… and in 2025 it **reversed course and reinvested in people**. - **65%** of organizations already use generative AI regularly in at least one function, nearly double the year before (McKinsey, 2024). - **70.22%** is the average documented online shopping cart abandonment rate (Baymard Institute). - A Canadian tribunal held Air Canada responsible for what **its chatbot said**, rejecting the argument that the bot was "a separate legal entity" (2024). - **76%** of consumers prefer to buy products with information in their own language (CSA Research). - WhatsApp has surpassed **3 billion** monthly active users (Meta, 2025). ## Adoption and market size First warning, and it matters: **there is no single "chatbot market size"** figure. Each firm defines the market differently —some count only the bot software, others the whole of conversational AI including voice and infrastructure— so estimates for the same year swing enormously. That's why we always name the firm. Be wary of any page that hands you one round number as settled fact. The global **conversational AI** market was estimated at between $17.05 billion (MarketsandMarkets) and $19.21 billion (Precedence Research) in 2025. The forward projections are just as scattered: $49.8 billion by 2031 (MarketsandMarkets), $41.39 billion by 2030 (Grand View Research), or as much as $155 billion by 2035 (Precedence). The gap between firms shows the problem: the "chatbot market" defined narrowly is valued at $1.42 billion in 2025 (Precedence), while Mordor Intelligence, using a broader definition, puts it at $11.45 billion in 2026 —about eight times more. Same name, different markets. ### Chatbot market size by region With that warning in mind, here is what research firms publish for each region —always with the firm's name next to the figure, because definitions are not interchangeable—. Base year 2025; the projection horizon is whatever each firm uses. Data accessed July 2026. | Region | 2025 market (USD) | Projection | CAGR | Source | | --- | --- | --- | --- | --- | | Global | $7.66–9.6 billion depending on the firm | $32.45–60.21 billion (2031–2034) | 19.6–24.5% | [Straits](https://straitsresearch.com/report/chatbot-market), [Mordor](https://www.mordorintelligence.com/industry-reports/global-chatbot-market/market-size), [IMARC](https://www.imarcgroup.com/chatbot-market), [Fortune BI](https://www.fortunebusinessinsights.com/chatbot-market-104673), [Grand View](https://www.grandviewresearch.com/industry-analysis/chatbot-market) | | North America | 38.7% of the global market | — | — | [Mordor](https://www.mordorintelligence.com/industry-reports/global-chatbot-market/market-size) | | United States | $1,960.2M | $8,819.9M (2033) | 20.3% | [Grand View](https://www.grandviewresearch.com/horizon/outlook/chatbot-market/united-states) | | Europe | $2,100M | $9,200M (2034) | 17.6% | [IMARC](https://www.imarcgroup.com/europe-chatbot-market) | | Latin America | $885.0M | $3,573.8M (2033) | 18.7% | [Grand View](https://www.grandviewresearch.com/horizon/outlook/chatbot-market/latin-america) | | Brazil | $660.7M | $2,983.6M (2033) | 20.4% | [Grand View](https://www.grandviewresearch.com/horizon/outlook/chatbot-market/brazil) | | Mexico | $469.7M | $1,728.9M (2033) | 16.7% | [Grand View](https://www.grandviewresearch.com/horizon/outlook/chatbot-market/mexico) | | Canada | $559.6M | $2,829.9M (2033) | 22.0% | [Grand View](https://www.grandviewresearch.com/horizon/outlook/chatbot-market/canada) | | France | $265.7M | $1,056.9M (2033) | 18.5% | [Grand View](https://www.grandviewresearch.com/horizon/outlook/chatbot-market/france) | | Gulf (GCC) | $160.0M | $722.6M (2034) | 18.2% | [IMARC](https://www.imarcgroup.com/gcc-chatbot-market) | | "Bot services" (separate category) | $3,890M | $19,820M (2031) | 31.2% | [Mordor](https://www.mordorintelligence.com/industry-reports/bot-services-market) | Three honest clarifications. First, about **the global range**: it covers the five firms cited in that row, all using 2025 as the base year. It leaves out the two figures you saw just above —Precedence's $1.42 billion, with its much narrower definition, and Mordor's $11.45 billion, which is its 2026 estimate rather than 2025—. That's exactly the problem we warned about: same name, different definitions and different years. Second: the last row, "bot services," is a market Mordor Intelligence measures as its own category, separate from its chatbot market —which is why the figures don't reconcile: they don't measure the same thing—. And third, about what's missing: no firm publishes a specific public figure for **Spain or Argentina**, which sit inside the Europe and Latin America numbers. If anyone offers you a firm figure for those two markets, ask where it comes from. Retail spend transacted over chatbots would grow from $12 billion in 2023 to $72 billion in 2028, a 470% rise, driven —according to Juniper itself— by the falling cost of large language models like ChatGPT. Of U.S. small businesses say they use generative AI, up from 40% in 2024 and around 23% in 2023 (survey of 3,870 small businesses). Adoption among small businesses has roughly doubled each year. Generative-AI chatbots climbed to the **second most-used technology tool** among U.S. small businesses (44% adoption, behind only search engines), up from fifth place a year earlier. Of the small businesses surveyed by Intuit use AI regularly, up from 48% in July 2024 (more than 2,200 businesses of up to 100 employees). ## Customer service impact Two very different kinds of data live here: what customer service leaders *expect* to happen, and what companies *measure* already happening. They're not the same, and we say which is which in each case. 75% of customer service leaders **expect** that, in the coming years, 80% of interactions will be resolved without human intervention. It's an expectation about the future, not a resolution rate measured today. (Survey of more than 10,000 people across 22 countries.) In February 2024, Klarna announced its AI assistant (powered by OpenAI) had handled 2.3 million conversations in its first month —two-thirds of all its support chats— doing the work equivalent to 700 full-time agents, resolving errands in under 2 minutes versus 11 previously. In 2025, Klarna walked it back. Its CEO admitted the company had "underestimated the trade-off" of prioritizing cost and efficiency, and reinvested in human support: "it's so critical that you are clear to your customer that there will be always a human if you want." It's the caveat almost nobody cites alongside the 700-agents figure. Customer service teams estimate AI handles around 30% of cases today, and project it will reach 50% by 2027 (survey of 6,500 service professionals). Agents who use AI spend 20% less time on routine cases, freeing up roughly four hours a week for more complex work. Fin, Intercom's AI agent, reports an average resolution rate of 76% across more than 12,000 customers. It's a vendor-reported aggregate skewed toward high-volume customers: real deployments vary widely, and out-of-the-box and published-case figures tend to land closer to 42%–53%. Useful as a reference, not as an industry norm. Gartner predicted that by 2026, conversational AI in contact centers would cut agent labor costs by $80 billion worldwide. For context: up to 95% of a contact center's cost is labor, and there are around 17 million agents worldwide. ## Cost and return on investment The dominant narrative says chatbots always save money. It's half true, and in the generative-AI era there's a wrinkle almost nobody mentions: processing each answer with a large language model has a cost the old button bots didn't have. Here are both sides. The classic savings statistic: in 2018, Juniper projected chatbots would deliver $11 billion in annual cost savings across retail, banking and healthcare by 2023, up from around $6 billion in 2018. It's a projection from the rule-based era; cite it with its date. The real origin of the "hours saved" figure: the same Juniper report estimated consumers and businesses combined would save more than 2.5 billion hours by 2023 through chatbot interactions. A Forrester Total Economic Impact study commissioned by boost.ai modeled a 293% three-year return for a composite company, with $19.9 million in net present value and payback in under 12 months. It's a composite-case study commissioned by the vendor itself: treat it as a reference to check, not an industry average. Similarly, a Forrester TEI study commissioned by PolyAI (voice AI) modeled a 391% three-year return and payback in under 6 months for a composite organization handling 4 million calls a year. Same caveat: composite case, commissioned by the vendor. The figure that breaks the narrative: Gartner projects that by 2030 the cost per resolution for generative AI in customer service will exceed $3, higher than what many offshore human agents cost. In the era of large models, cost per answer can go *up*, not down. ## What consumers think and want This is the section where an honest roundup separates itself from a brochure. Sentiment data isn't applause for chatbots: when you ask well, with large samples, a lot of reluctance shows up. We include the figures for and against. Nearly two-thirds of customers would prefer companies **did not** use AI in their customer service (Gartner survey of 5,728 people, December 2023). Would consider switching to a competitor if they discovered a company was going to use AI in its customer service. Consumers' top concern: that it becomes harder to reach a person (60%), followed by AI giving wrong answers (42%). Of consumers want to **know whether they're talking to an AI** rather than a person (survey of 15,015 consumers across 18 countries). Acceptance is conditional: it rises to 45% if there's a clear path to escalate to a human, and to 44% if the agent's reasoning is explained. Of consumers say they trust AI agents more when they show human traits like friendliness and empathy. And 67% say they're ready to delegate tasks such as order tracking or personalized recommendations to AI. Nearly 70% of users admit to having sworn at a chatbot out of frustration. In the same study, 30% would rather wait for a human than get an immediate answer from a bot, and 11% would pay extra to skip the chatbot and reach a person. Of consumers now expect customer service to be available 24 hours a day, 7 days a week —an expectation that AI's own availability has helped normalize. A background counterpoint, with its date up front: back in 2018, 82% of U.S. consumers (and 74% outside the U.S.) said they wanted **more** human interaction in the future, not less. 59% felt companies had lost touch with the human element of experience. It predates modern AI assistants; use it as a historical baseline, not a figure for today. ## E-commerce and messaging The average documented online shopping cart abandonment rate. It's the mean of 50 separate studies (an average of averages, not a single measured population), and it's the foundation of the case for automated help before the sale. 43% of shoppers who abandon a cart do so because they were "just browsing." Among those who did intend to buy, the top reason for abandoning is extra costs (shipping, taxes, fees) being too high: 39%. 53% of online shoppers abandon a purchase if they can't find a quick answer to their question. The classic figure for on-site help; note its year (2016) when citing it. During the 2025 holiday season, traffic to U.S. retail sites from generative-AI assistants grew 693.4% year over year. Generative AI is becoming a real source of shopping visits, not just answers. Shoppers who arrived from a generative-AI assistant converted 31% more than other traffic sources, spent 45% more time on site, and viewed 13% more pages per visit. WhatsApp has more than 3 billion monthly active users (over 100 million in the U.S.), per Meta itself. Messenger, in turn, has more than 1 billion. It's the stage on which messaging chatbots operate. Meta reported more than a million weekly conversations with its "Business AIs" in early markets (Mexico, the Philippines), on a base of more than 1 billion daily threads between people and business accounts across its messaging platforms. Separately, WhatsApp's paid messaging passed a $2 billion annual run rate. E-commerce spend over rich-media messaging channels (RCS and WhatsApp-style apps) would grow about 30%, from $14 billion in 2025 to almost $19 billion in 2027. ## The generative-AI shift Everything in this section is after ChatGPT (late 2022). It's the change that makes so many earlier statistics stop being useful: it's not that chatbots improved a little, it's that the technology behind them changed. In early 2024, 65% of organizations surveyed said they were using generative AI regularly in at least one business function, nearly double the year before. Enterprise spending on generative AI reached $13.8 billion in 2024, more than six times the $2.3 billion of 2023. Customer support accounted for 9% of that spend. 70% of customer service leaders were rethinking their customer journeys with generative-AI tools, and 83% of those already using it in service reported a positive return. On top of that, 68% of consumers believe chatbots should have the same level of knowledge and quality as a skilled human agent: the bar went up. Gartner predicted that by 2027, chatbots would become the primary customer service channel for roughly a quarter of organizations. (The prediction is from July 2022, a few months before ChatGPT's public launch.) Gartner projects that by 2029, "agentic" AI (the kind that acts on its own, chaining several steps rather than just replying) will autonomously resolve 80% of common customer service issues without human intervention, with a 30% cut in operational costs. It's a prediction; weigh it against the sentiment and cost figures in the earlier sections. ## Limits, risks and the human factor No serious roundup of chatbot statistics is complete without the fine print: when they fail, what legal risk they carry, and why the human is still needed. Air Canada's chatbot gave a customer false information about bereavement fares. The tribunal found the airline liable for negligent misrepresentation and rejected its argument that the chatbot was "a separate legal entity responsible for its own actions": "it should be obvious to Air Canada that it is responsible for all the information on its website, whether it comes from a static page or a chatbot." The reference precedent on chatbot legal liability. In the specific task of *summarizing a document they're handed*, large language models still fabricate details not in the text, at rates from under 2% (the best) to over 20% (the worst); many mainstream models sit between 5% and 12%. It's a summarization-faithfulness test, not a "chatbots are wrong X% of the time" figure: it's the closest, most neutral analog to a support bot properly fed with your data. A Stanford study tested general-purpose 2023 models (GPT-3.5, PaLM 2, Llama 2) with more than 200,000 legal queries each: they hallucinated on 69%–88% of specific questions, and got a case's core ruling wrong at least 75% of the time. It's the proof of why a bare model, with no connection to verified data, is unreliable on questions of fact. Current models, and especially those connected to a knowledge base, perform considerably better. Gartner predicts that half the companies that cut customer service staff because of AI will rehire for similar roles before 2027. The Klarna story raised to a trend: AI is not yet replacing human support the way early adopters expected. Hallucinations are the reason a serious business chatbot shouldn't work "from memory," but query your data on every answer and admit when it doesn't know something. We cover this in [why your chatbot makes things up](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). ## Language and multilingual support Language is one of the most underrated variables in automated customer service —and one of the best documented. A counter-intuitive point for English speakers: even if you operate in English, most of your potential market doesn't. 76% of online shoppers prefer to buy products with information in their own language, and 40% never buy from websites in other languages (survey of 8,709 consumers across 29 countries). Of consumers are more likely to buy from the same brand again if after-sales support is in their own language: the direct argument for multilingual automated support. English is the content language of just under half of the world's websites. In other words, most of the internet —and of the people using it— runs in other languages. (The figure updates daily; "just under half" is the stable read.) When they can choose a language, nearly 9 in 10 European internet users always visit websites in their own language, and only 53% would accept an English version if it's not available in their language. 44% said they had missed interesting information because it wasn't in a language they understood. (2011 survey; the pattern holds, but cite it with its date.) ## Europe, Spain and Latin America The big studies are global or U.S.-based. These are the regional figures we could verify in official sources. A note on Spain: there are two legitimate "companies using AI" figures that look contradictory and aren't, because they come from different surveys and methodologies. We always give the year and the source. Of EU enterprises (with 10 or more employees) used AI in 2025, up from 13.5% in 2024. The most common application is analyzing written language —which includes chatbots— at 11.8%. The gap by size is wide: around 55% of large enterprises versus 17% of small ones. The EU AI Act, in force since August 2024, provides for fines of up to €35 million or 7% of total worldwide annual turnover (whichever is higher) for breaching the prohibited-AI-practices rules. A reference framework for any company deploying customer-facing AI in Europe —wherever the company itself is based. Of Spanish enterprises with 10 or more employees used AI in the first quarter of 2025 (8.7 points more than a year earlier). In the services sector, the figure rises to 25.7%. This is Spain's statistics office (INE) own, most recent survey. Under Eurostat's harmonized methodology, 11.4% of Spanish enterprises used AI in 2024, two points below the EU average. Among large enterprises (more than 249 employees), adoption reaches 44%. This doesn't contradict the INE figure: it's a different wave and methodology, which is why we give both with their year. For Latin America, comparable primary data is scarce —many per-country figures live in paid reports— but one fact is well documented: WhatsApp is the dominant contact channel. Outside China and Russia (where it's barely used), 69.8% of internet users aged 16+ use WhatsApp, and its penetration in Spain, Mexico, Brazil or Argentina is among the highest in the world. It's the arena where automated customer service over messaging plays out. ## How we verified these figures The standard for this page, in a few lines, because it's what sets it apart: - **Primary source or out.** Every figure comes from the original study, press release, report or ruling, not from a blog quoting another blog. If a widely-repeated number couldn't be traced to its origin, we dropped it rather than reproduce it. - **Year always visible.** A statistic with no date is worthless in a field that changes every quarter. Each figure carries its year, and we flag what predates the ChatGPT era. - **Measurement vs projection.** We distinguish what has already been measured from what a firm *projects* for years ahead. They're not the same. - **The firm's name on market size.** Estimates vary so much between firms that giving a single number would be misleading. - **One transparent exception.** A few figures (mostly Gartner, and one or two from Salesforce) come from press releases that block automated access. In those cases, the figure is taken from the firm's own official press release and cross-checked word for word against several reputable outlets that reproduce it; we say so here rather than pretend to a line-by-line verification we didn't do. If you spot an outdated figure or a broken link, let us know: keeping this page current is part of the job. ## Frequently asked questions ### How many businesses use chatbots or AI in customer service? It depends on the segment. Among U.S. small businesses, 58% say they use generative AI and chatbots are now the second most-used tool, at 44% adoption (U.S. Chamber of Commerce, 2025). Among EU enterprises, 20% used AI in 2025 (Eurostat), and in Spain 21.1% in the first quarter of 2025 (INE). Customer service teams estimate AI handles around 30% of cases today (Salesforce, 2025). ### Do customers prefer chatbots or humans? With caveats, they lean human. 64% of customers would prefer companies didn't use AI in their service, and 53% would consider switching companies over it (Gartner, 2024). Their biggest fear is that it becomes harder to reach a person. That said, acceptance rises sharply with transparency (75% want to know if they're talking to an AI) and a clear path to escalate to a human (Salesforce, 2024). ### How much do chatbots really save? The classic savings figures ($11 billion a year, 2.5 billion hours) are Juniper Research projections from 2018, from the rule-based era. In the generative-AI era the math is more nuanced: Gartner projects that by 2030 the cost per resolution for generative AI will exceed $3, higher than many offshore human agents. The savings are real in some cases, but they're not automatic. ### How often do AI chatbots get it wrong or "hallucinate"? There's no single reliable figure, and be wary of anyone who gives you one. In the narrow task of summarizing a document they're handed, large models fabricate details between 2% and 24% of the time depending on the model (Vectara). With general-purpose 2023 models and complex legal questions, error rates reached 69%–88% (Stanford). The practical takeaway: a reliable business chatbot shouldn't answer from memory, but query your data and admit when it doesn't know something. ### Is a company legally responsible for what its chatbot says? Yes, per the most-cited precedent. In 2024, a Canadian tribunal held Air Canada responsible for the false information its chatbot gave a customer, and flatly rejected the argument that the bot was "a separate legal entity." A company answers for what its chatbot says just as it does for any other page on its website. ### Does language matter in automated support? A lot. 76% of consumers prefer to buy with information in their language and 40% never buy in other languages (CSA Research). 75% are more likely to buy again if after-sales support is in their language. And since English is the language of less than half of the world's websites (W3Techs), serving customers in only one or two languages leaves out a huge part of the market. ## Sources Every figure links to its primary source in the figure itself. These are the organizations and studies the information on this page comes from: - Analyst and market-research firms: Gartner, McKinsey & Company, Forrester, Juniper Research, MarketsandMarkets, Precedence Research, Mordor Intelligence, Grand View Research, IMARC Group, Fortune Business Insights, Straits Research, Menlo Ventures. - Platforms and industry surveys: Zendesk (CX Trends), Salesforce (State of Service; State of the Connected Customer; Small & Medium Business Trends), Intercom, Klarna, Adobe Analytics, Tidio, boost.ai, PolyAI. - Official and academic sources: Eurostat, European Commission (AI Act; Eurobarometer), INE, ONTSI / Red.es, U.S. Chamber of Commerce, Intuit QuickBooks, Meta (quarterly results), Baymard Institute, W3Techs, CSA Research, Stanford HAI, Vectara, British Columbia Civil Resolution Tribunal (Moffatt v. Air Canada), DataReportal. Compilation maintained by the Bravos AI team. Last updated: July 2026. You're free to use these figures; if you link to this page as the source of the compilation, you help us keep it current. --- # Best WhatsApp AI Chatbot in 2026: 13 Platforms Compared by Price **Category:** Analysis | **Read time:** 16 min | **Date:** June 10, 2026 | **URL:** https://bravos-ai.com/en/blog/best-whatsapp-chatbot The question «what is the best WhatsApp chatbot?» has no single answer. There are **13 relevant platforms** in the market in 2026 and each one tells a different story: real prices that do not show up on the marketing page, AI models that range from a true [AI agent](https://bravos-ai.com/blog/que-es-un-agente-de-ia) running GPT-5 down to «flow-based bots with one AI-generated step inside», markups over Meta's native rates that almost nobody mentions, and very uneven international and language coverage. We opened the official pricing pages of all 13, cross-checked with independent reviews on G2, Capterra and tech blogs, and **refuted 7 claims** that appear in widely cited guides. The output: an honest comparison of how much each one actually costs, what AI engine sits underneath, how they connect to Meta and who each platform fits. If you are in a hurry, jump straight to the summary table. If you want to understand why a «$99/mo Wati» does not exist or why Tidio Lyro now covers WhatsApp (contradicting several recent guides), keep reading. ## The 3 real categories of WhatsApp chatbot platforms Before comparing prices, a distinction most comparisons skip and that changes the decision criteria a lot: - **WhatsApp specialists.** Platforms built around WhatsApp itself. Wati, Respond.io, AiSensy, Interakt, SleekFlow and 360dialog are the best known. They usually have deeper integration with the official API, ready-to-use templates and pricing tuned for businesses that live on WhatsApp. Weak spot: most are Asian (India, Hong Kong, Malaysia) without real presence in Europe or the US. - **Multichannel platforms that added WhatsApp.** Tools that started in other channels (website chat, Messenger, Instagram) and added WhatsApp later. Tidio, Intercom, ManyChat, Landbot, HubSpot and Crisp fit here. Strong spot: if your business also needs web or social channels, they cover it. Weak spot: WhatsApp typically sits on a higher-tier paid plan. - **The technical DIY route.** Twilio is the textbook case. They give you the raw Meta connection and you build the rest. No chatbot, no inbox, nothing pre-built. A good fit if you have a dev team; a torture chamber otherwise. ## Comparison table: price, real AI, Meta surcharge and ecommerce Prices verified in June 2026 directly on the official pricing pages. We show the price of the **cheapest plan that actually includes a WhatsApp chatbot with real conversational AI** — not the entry-level decoy without AI, which is the usual trap. | Platform | Monthly price | Real AI | Meta surcharge | Shopify native | | --- | --- | --- | --- | --- | | Wati | $119 annual / $149 monthly (Pro) | Yes (Astra) | 2.8–21% | Yes | | Respond.io | $159 (Growth) | Yes (own RAG) | None | Yes | | AiSensy | $125 (Basic + chatbot) | Flows by default; LLM only on Enterprise | Not disclosed | Yes | | Interakt | ~$160 (Advanced + Haptik) | Hybrid | ~25% | Yes | | SleekFlow | ~$154 team (3-user min.) | Yes (GPT-4o + AgentFlow) | +$15/number/mo | Yes | | 360dialog | €49 (no AI) | None (pure BSP) | None | Via Zapier | | Tidio | $56.67 (Starter + Lyro) | Yes (Claude) | Not disclosed | Yes | | Intercom Fin | $29/seat + $0.99/resolution | Yes (Fin Apex + Claude) | Not disclosed | Yes | | ManyChat | $58 (Pro + AI) | Hybrid (single AI step) | None known | Via Zapier | | Landbot | ~€80 (WhatsApp Starter) | Hybrid (GPT-4 inside flows) | +€0.05/extra chat | Via Zapier | | HubSpot Service Hub | ~$90/seat + $1,500 onboarding + $0.50/resolution | Yes (GPT-4.x / GPT-5) | None | Yes | | Crisp | $95 (Essentials) | Hybrid (pick Claude/ChatGPT/Llama) | Markup on templates | Yes (+ PrestaShop, Woo) | | Twilio (DIY) | $0 + $0.005/msg + Meta | None (you build it) | ~20–40% effective | No | | Bravos AI | $23 (Starter) / $59 (PRO) | Yes (GPT-4 + RAG) | None | Yes (+ PrestaShop) | A «BSP» (Business Solution Provider) is an official partner Meta authorizes to connect businesses to the WhatsApp API. Some are verified *Tech Providers* (accredited directly by Meta) and others resell through third parties like Twilio or 360dialog. Prices read from official pricing pages in **June 2026**; these numbers move, double-check before signing up. For a broader take on what an AI chatbot really costs (not just on WhatsApp), our piece on [how much an AI chatbot costs in 2026](https://bravos-ai.com/blog/cuanto-cuesta-chatbot-empresa) has more context. If price is the only thing you care about, the three honest finalists are **Bravos AI**, **Tidio** and **ManyChat** in that order. If you need real conversational AI with no markup over Meta, narrow to **Respond.io**, **HubSpot** and **Bravos AI**. Detailed fiches below. ## The 6 WhatsApp specialists ### Wati (Hong Kong, $119–149/mo) Wati is the best-known specialist internationally. Their Astra AI Agents product runs on an LLM (they have not published which one) and the Shopify integration is native and well documented. The plan that actually includes a customer-facing bot is [Pro at $119/mo annual or $149/mo monthly](https://www.wati.io/pricing/) — not the $99 that circulates in some guides (that's the price of the standalone Astra product, not Wati Pro). The Growth plan at $59/mo only includes the AI Co-pilot that suggests replies to human agents, not a customer-facing bot. **Downsides:** Wati is a BSP and applies [a markup over Meta's rates between 2.8% and 21%](https://www.ycloud.com/blog/wati-pricing) depending on message category, which in real invoices translates to 2-3x the expected bill. G2 and Capterra reviews mention support responses taking multiple days, recurring UI bugs and weak analytics. Website is English-only with no documented Spanish UI or support. ### Respond.io (Kuala Lumpur, $159/mo) HQ in Malaysia, not Singapore as several lists claim. [The Growth plan at $159/mo](https://respond.io/pricing) includes their AI Agents with multi-agent architecture and their own RAG layer (they have not published the underlying LLM). One of the few specialists that puts [in writing that they do not apply markup over Meta's rates](https://respond.io/whatsapp-business-api), which on mid-volume invoices matters. **Downsides:** expensive for solo founders and small teams (no free plan and no entry below $79/mo for WhatsApp). Reviews mention a steep learning curve, basic reporting and no A/B testing. They run a Spanish version at respond.io/es/ , but support team language is not explicitly confirmed. ### AiSensy (India, $125/mo total) High customer volume in India and Southeast Asia. Basic at $45/mo does not include a chatbot; you need the Chatbot Flows add-on at $80/mo, totaling $125/mo. The default AI engine is **flow and intent-based**, not conversational LLM. To use real LLMs (OpenAI, Claude, Meta LLaMA or DeepSeek) you need to move to [their Enterprise tier with your own API key](https://aisensy.com/features/ai-whatsapp-chatbot) (BYOA). **Downsides:** harsh reviews on the [Shopify App Store](https://apps.shopify.com/aisensy-whatsapp-marketing/reviews) (2.3/5 with 71% one-star). Slow support, «AI chatbot» that in practice is rigid templated responses, and the Basic plan does not even let you schedule flows. ### Interakt (India, ~$160/mo with real AI) Subsidiary of Jio Haptik (Reliance Group), one of India's largest conversational AI players. The Advanced plan costs around ₹3,799/mo (~$45) and is rule-based; the real LLM comes from the Haptik AI Agents add-on at ~$115/mo. Total for a serious bot: around $160/mo. **Downsides:** [~25% markup over Meta on Growth and Advanced plans](https://www.zoko.io/post/interakt-pricing-guide) (₹0.871 vs Meta's ₹0.695 for marketing messages). Enterprise tier does promise «no conversation markups». Email-only support with delays, mobile app log-outs, and basic analytics outside the top tier. ### SleekFlow (Hong Kong, ~$154/mo team) Probably the most technically sophisticated option on the list. Their [AgentFlow product](https://sleekflow.io/news/sleekflow-announces-agentflow), launched July 2025, uses a multi-agent architecture with graph database, hybrid Graph-Vector RAG, and dynamic routing between GPT-4o, GPT-4o-mini, Gemini and o3-mini on Azure OpenAI with customer data isolation. Pro AI is HK$399/user/mo (~$51) with a 3-user minimum, putting the team invoice at ~$154/mo. **Downsides:** adds **$15/mo per WhatsApp number hosted** on top of Meta's pass-through. No Spanish version (product languages: English, Simplified Chinese, Traditional Chinese, Indonesian, Brazilian Portuguese). Reviews flag limited analytics depth, the 10-rule cap on the Pro plan and difficult data migration. ### 360dialog (Berlin, €49/mo without AI) The reference European BSP. [Regular plan at €49/mo per number](https://www.360dialog.com/pricing), Premium €99/mo (with account pre-verification and a 30-min response SLA) and High Throughput €249/mo for very high volume. It is pure WhatsApp infrastructure: **no chatbot, no AI included**; you add the conversational layer separately (third-party SaaS or self-built). Zero markup over Meta's pricing is the main draw. **Downsides:** Capterra reviews mention billing disputes, difficulty cancelling and slow support. As a standalone product it's just a connector to Meta — useful for agencies that want to control how they bill their own clients, useless for a small business that just wants their bot running. ## The 6 multichannel platforms with WhatsApp ### Tidio (Poland, $56.67/mo with AI) Probably the best-known multichannel option for small businesses. Their **Lyro** agent runs on [Anthropic Claude](https://claude.com/customers/tidio). Important refutation: several guides claim Lyro is not available on WhatsApp. [Tidio's own docs confirm it is](https://help.tidio.com/hc/en-us/articles/9003475527196). The pricing: Starter at $24.17/mo (with WhatsApp) + Lyro add-on from ~$32.50/mo = around **$56.67/mo** for a WhatsApp bot with real AI. **Downsides:** per-conversation billing that creates Black Friday bill shocks. Product Recommendations are English-only. Website crawler caps at 60 pages. Advanced API features only on much pricier Plus/PRO tiers. Connecting an existing WABA forces you to delete it and recreate — you lose all prior customer history. ### Intercom Fin ($29/seat + $0.99/resolution) The enterprise option par excellence. Essential at [$29/seat/mo annual](https://www.intercom.com/pricing) includes their Fin agent. WhatsApp is pay-as-you-go on top. The engine: [Fin Apex 1.0](https://www.intercom.com/blog/announcing-fin-apex-the-age-of-vertical-models-is-here/) (launched March 2026, proprietary model post-trained by Intercom) plus [Anthropic Claude](https://www.intercom.com/blog/fin-2-powered-by-anthropic-claude-llm/) in other parts of the system. **Downsides:** the per-resolution billing ($0.99 each) scales unpredictably. Reviews mention Fin hallucinating more than expected, add-ons (Copilot at $29-35/seat) stacking up fast, and a 5-10 agent team easily crossing $500/mo. Product is in Spanish but the main marketing site stays English-only. ### ManyChat ($58/mo with AI) Started as a Messenger marketing tool and expanded to WhatsApp. Pro at $29/mo + AI Add-on at $29/mo = $58/mo at the 2,500-contact tier (can start cheaper at fewer contacts). [ManyChat's AI](https://help.manychat.com/hc/en-us/articles/14281187288860-Manychat-AI-Step) is a **single step inside a rule-based flow**: the bot is still button-driven, and at a specific node it generates a reply with OpenAI. Not a free-flowing conversational agent. **Downsides:** AI is single-step, not conversational. No native Shopify or WooCommerce integration (Zapier required). Contact-based pricing balloons as you grow. Reviews mention thin support during Meta API changes. Translating flows to multiple languages is tedious. ### Landbot (Barcelona, ~€80/mo) A Spanish company headquartered in Barcelona. The [pricing page](https://landbot.io/pricing) shows a WhatsApp Starter from ~€80/mo and a WhatsApp Pro at €160 (annual) / €200 (monthly). The AI is **hybrid**: visual flow builder with an [OpenAI Block](https://help.landbot.io/article/sz9n4ri87v-open-ai) that calls GPT-4 inside a specific node. Dialogflow integration is available on WhatsApp Professional. Verified Meta Tech Provider. **Downsides:** adds **€0.05 per extra service chat** over the included quota, and €0.10 per extra AI chat. You're locked into whatever GPT model Landbot ships (no swapping to Claude or Gemini). The flowchart approach limits truly free-form conversations — customers still navigate through buttons. Reviews mention handover latency during volume spikes. ### HubSpot Service Hub (~$90/seat + $1,500 onboarding) The most enterprise piece on the list. Service Hub Professional runs around [$90/seat/mo annual](https://www.featurebase.app/blog/hubspot-pricing), with **$1,500 one-time onboarding**, and Breeze AI charges **$0.50 per resolved conversation**. Breeze runs on GPT-4.x for the Customer Agent and GPT-5 for Studio agents (effective January 2026 per HubSpot's public model cards). **Downsides:** HubSpot **is not a BSP**. You have to complete Meta business verification yourself and the native integration caps at [1,000 outbound templates/month](https://www.uptail.ai/blog/hubspot-whatsapp-integration-how-to-connect-what-you-get-and-where-the-gaps-are). No native WhatsApp flow builder (no drag-and-drop specifically for WhatsApp). Capterra and G2 reviews call the AI «behind the curve». A fit if you already pay for HubSpot anyway; overkill if you just want WhatsApp. ### Crisp (France, $95/mo Essentials) A French platform popular in Europe. [Essentials at $95/mo](https://crisp.chat/en/pricing/) includes WhatsApp Business and $25 in AI credits. Their Hugo agent **does not use a proprietary LLM** — we refuted that during verification. Instead, [you pick the underlying model](https://crisp.chat/en/chatbot/custom-chatbot/): Claude, ChatGPT or Llama. One of the few with **native PrestaShop integration** alongside Shopify and WooCommerce. **Downsides:** shallow analytics; many advanced features (triggers, extra agents, AI credits) are paid extras; Capterra reports unstable notifications; their AI does not auto-train from history. Markup applies on paid Meta template messages. ## The DIY route: Twilio [Twilio charges $0.005 per message](https://www.twilio.com/en-us/whatsapp/pricing) on top of Meta's native rates. No dashboard, no chatbot, no inbox: you build the agent by wiring OpenAI, Anthropic or whichever LLM you prefer. Effective markup over Meta in real invoices runs 20-40% depending on volume. It makes sense **only if you have a dev team** and need full control. If that's your case, we cover it in depth in [how to build a WhatsApp chatbot with n8n, Make or Zapier](https://bravos-ai.com/blog/chatbot-whatsapp-alternativa-zapier-make-n8n), which also applies to a direct Twilio + Meta integration. If you do not have devs, Twilio is a punishment. ## The hidden Meta surcharge nobody talks about The official WhatsApp Cloud API pricing page is public and Meta's per-category rates are open. But most platforms **add their own margin on top**. And almost none state it clearly on their pricing page. What we verified: - **Wati**: ~2.8% to ~21% depending on message category. - **Interakt**: ~25% on Growth and Advanced plans; no markup on Enterprise. - **Twilio**: $0.005/message on top of Meta — effectively 20-40% depending on volume. - **SleekFlow**: no per-message markup, but $15/mo per WhatsApp number. - **Crisp**: markup on paid Meta template messages. - **Landbot**: €0.05 per extra service chat over the included quota; €0.10 per extra AI chat. - **Respond.io, 360dialog, HubSpot**: explicitly state no markup over Meta's pricing. - **Bravos AI**: no markup. What Meta charges, you pay Meta directly — no intermediary. [Since July 2025](https://developers.facebook.com/docs/whatsapp/pricing/), Meta stopped charging for customer-initiated conversations: replies within the 24-hour service window are free and uncapped. This changes the math for businesses that only handle inbound messages: if your bot just replies (the typical small-business case), your Meta invoice is **zero**, and you only pay the platform subscription. The markup matters far more for outbound marketing templates than for reactive customer support. ## Best WhatsApp chatbot by use case ### Solopreneur or micro-business on a tight budget The cheapest option with real AI and predictable monthly bill is **Bravos AI from $23/month** (Starter, WhatsApp included). If you also need lead capture forms, the PRO plan ($59/month) adds them. The next step up is **Tidio** with Lyro at ~$56.67/month for Shopify-heavy use. For absolute minimum, **ManyChat** at $58/month works if your contact volume is low and you accept that the «AI» is a single step inside a flow. ### Small business with a Shopify store (catalog + support + marketing) Three balanced options with native Shopify integration: **Bravos AI from $23/mo** (real-time catalog sync, SQL-style structured search on products so the bot finds by size, price or category without making things up), **Tidio at ~$56.67/mo** with Lyro (Claude) on Shopify but with the catch of per-conversation billing during peak season, and **Wati at $119/mo** if your priority is outbound marketing with templates. For stores with large catalogs, our piece on [why most chatbots fail with product catalogs](https://bravos-ai.com/blog/chatbot-catalogo-productos) is worth reading first — not every platform handles structured search well. ### Agency deploying to multiple clients **360dialog** at €49/mo per number, no Meta markup, is the cleanest base for controlling how you bill your clients — reminder: it does not include a chatbot, you wire the AI agent on top. The product-complete alternative is **Respond.io Growth at $159/mo**, also no markup, with continuous RAG. Avoid Wati and Twilio at scale: per-message markup compounds across clients and chews through margins by month-end. Bravos AI offers team seats from the PRO plan and a multi-bot dashboard to manage several clients from one account. ### Dev team that wants total control **Twilio** remains the standard — pure WhatsApp infrastructure, opinionated about nothing, build the bot with whatever LLM you prefer. The catch is GDPR and Schrems II friction for EU customers. **360dialog** is the European alternative with the same philosophy and EU-hosted data. If you want to build with n8n or Make, we break it down in [this cluster article](https://bravos-ai.com/blog/chatbot-whatsapp-alternativa-zapier-make-n8n), where we calculate the real 13-26 hours of build time. ### Restaurant, clinic or local business **Bravos AI** ($23/mo) and **Landbot** (~$85/mo) are the budget-friendly options with real conversational AI for the typical case: answering FAQs, sharing booking links, handling after-hours inbound. **Crisp Essentials** at $95/mo is a solid alternative if you also need native PrestaShop. **HubSpot** is overkill for the case ($90/seat + $1,500 onboarding). Our article on [how to auto-reply on WhatsApp](https://bravos-ai.com/blog/responder-whatsapp-automaticamente) gives more context for this reactive use case. ### Enterprise team with many seats and a full suite Here **Intercom Fin** and **HubSpot Service Hub** are the serious options. Intercom from $29/seat + $0.99/resolution is cheaper to start but costs scale with volume. HubSpot Service Hub Professional from $90/seat plus $1,500 onboarding fits if you already pay for HubSpot anyway. Both ship a full suite (tickets, knowledge base, advanced reporting) that WhatsApp-focused tools do not cover. ## Where Bravos AI fits in this landscape Honest framing of where we sit: we are not a marketing-automation suite like ManyChat or AiSensy, nor an enterprise customer-service platform like Intercom or HubSpot. We're closer to **Crisp and Landbot**: European, with real Spanish-language presence, but with the difference that our product is built **around the customer's catalog or knowledge base**, not around a flow builder. **What sets us apart:** - **Verified Meta Tech Provider**: connect your number in a 60-second popup — no Twilio in the middle, no tokens that expire every 60 days. - **No markup on Meta's pricing**: if your bot only replies within the 24-hour service window (the typical case), your monthly bill is the subscription and nothing else. - **Honest pricing**: $23/mo Starter (WhatsApp and analytics included) or $59/mo PRO (3 bots and built-in lead-capture forms). New accounts get **7 days of PRO trial** with everything included. - **Real AI with RAG**: GPT-4 over your content (CSV, web, PDF) plus structured SQL search on catalogs so the bot answers with the real data of your business (price, size, color, stock) instead of making things up. - **Native store integrations**: **Shopify** (real-time catalog sync via webhooks), **PrestaShop** (multilingual catalog sync via API), **Resales Online** for real estate, and **Custom Webhook** to wire your ERP, SAP or internal database. - **The assistant you already use improves the bot for you**: with [our MCP connector](https://bravos-ai.com/conector-ia-mcp), Claude or ChatGPT carry our team's experience inside and tell you what to adjust; no other platform on this list allows it. - **Spanish-first**: web, product and support in Spanish, team based in Madrid. **What we don't cover yet:** - Native WooCommerce or BigCommerce integration. Crisp and Tidio cover those ecosystems better. - Multi-agent architecture at the level of SleekFlow AgentFlow — our RAG is single-agent with hybrid search (semantic + structured SQL). - Marketing and CRM integration catalog as broad as Tidio's or Wati's. We cover Sheets, CSV/Excel sync, native lead-capture forms and webhooks; we don't have one-click connectors to Klaviyo, Mailchimp or HubSpot yet. The bot replies with the clinic's information (hours, accepted insurance, address, consultation length) and shares the link to the online calendar. It does not book the appointment itself: the booking stays in the business's own system. If you want to try us, the 7-day PRO trial lets you see the bot working with your real data at no cost. More context in [how to build a WhatsApp chatbot in 2026](https://bravos-ai.com/blog/como-hacer-chatbot-whatsapp), where we walk through the four real paths step by step. ## FAQs ### What is the best WhatsApp chatbot in 2026? There's no single answer. For a small business on a tight budget, **Bravos AI at $23/month** is the cleanest mix of price, real AI and predictable bill. For a mid-size Shopify store, **Tidio with Lyro** at $56.67/month is the best value. For enterprise teams, **Intercom Fin**. For agencies juggling several clients without per-message markup, **360dialog** as infrastructure or **Respond.io** as finished product. For devs wanting full control, **Twilio**. ### Which WhatsApp chatbot vendor should you choose? There's no single "best provider" — it depends on your case. For real conversational AI at a low price with native ecommerce sync (Shopify, WooCommerce, PrestaShop), **Bravos AI** starts at $23/month; for enterprise CRM, **HubSpot** or **Intercom Fin**; **Landbot** if you want visual flow-building. This guide ranks the 13 real WhatsApp chatbot platforms so you can pick the vendor whose pricing and AI depth fit your use case. ### What is the cheapest alternative to Wati? Wati Pro is $119/month annual or $149/month monthly (not the $99 some outdated guides claim). Cheaper alternatives with real AI on WhatsApp: **Bravos AI at $23/month**, **Tidio with Lyro at $56.67/month**, or **ManyChat at $58/month** (with the caveat that ManyChat AI is a single step inside rule-based flows, not a free-flowing agent). ### Do these platforms charge a markup over Meta's pricing? Many do. Wati 2.8–21% depending on message category, Interakt ~25% on mid-tier plans, Twilio effectively 20–40% on top via its $0.005/message fee. Only Respond.io, 360dialog, HubSpot and Bravos AI **explicitly state no markup**. Important: since July 2025 Meta no longer charges for customer-initiated conversations, so the markup mostly affects outbound marketing templates, not reactive customer support. ### Wati vs Tidio for a Shopify store: which one? Depends on priorities. **Tidio is cheaper** ($56.67/mo with Lyro vs Wati Pro's $119/mo annual) and its Lyro agent runs on Claude, which in tests performs better than Wati's undisclosed Astra model. **Wati is stronger on outbound marketing**: its templates, campaign automations and abandoned-cart sequences are more polished. If your use is reactive support + catalog questions, Tidio. If you need to send marketing templates to your WhatsApp base often, Wati. For Shopify stores on a tight budget that need real conversational AI, **Bravos AI from $23/month** is a third path: native Shopify integration and structured SQL search on the catalog. ### Which platforms have real LLM AI vs only flow-based bots? Real conversational LLM: **Bravos AI** (GPT-4 + RAG), **Tidio** (Lyro on Claude), **Wati** (Astra), **Respond.io** (own RAG), **SleekFlow** (GPT-4o + AgentFlow), **Intercom Fin** (Apex + Claude), **HubSpot Breeze** (GPT-4.x / GPT-5) and **Crisp Hugo** (you pick Claude, ChatGPT or Llama). Hybrid or flow-based with one AI step: **ManyChat**, **AiSensy** by default, **Landbot**, **Interakt** base plans. No AI included: **360dialog** and **Twilio** — you bring your own. ## Bottom line Honest market snapshot, June 2026: **13 relevant platforms**, prices ranging from $23/mo at Bravos AI to over $500/mo at Intercom with several seats, AI quality everywhere from button flows with one generated step to multi-LLM agents with dynamic RAG, and a lot of fine print on Meta-rate markups. For most small businesses and local operations, the reasonable options are **Bravos AI**, **Tidio**, **Landbot** and **Crisp** in that order of price/language/real-AI balance. For Shopify-heavy stores with a bigger budget, add **Wati**. For enterprise, **Intercom** or **HubSpot**. For agencies and devs, **360dialog**, **Respond.io** or **Twilio** depending on how much control you want. For more context, check our piece on [how to build a WhatsApp chatbot in 2026](https://bravos-ai.com/blog/como-hacer-chatbot-whatsapp), on [how to auto-reply on WhatsApp](https://bravos-ai.com/blog/responder-whatsapp-automaticamente) or the more technical [building it with n8n, Make or Zapier](https://bravos-ai.com/blog/chatbot-whatsapp-alternativa-zapier-make-n8n). ### Try your WhatsApp chatbot in 60 seconds? Build your bot on Bravos AI from $23/month, connect your WhatsApp Business number and let the AI answer with the real info from your business. 7-day PRO trial with everything included. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # How to Add an AI Chatbot to WordPress in 5 Minutes **Category:** Tutorial | **Read time:** 8 min | **Date:** Jan 9, 2026 | **URL:** https://bravos-ai.com/en/blog/wordpress-chatbot You have a WordPress website. You're getting traffic, but you can't be online 24/7 to answer questions. What you need is a WordPress chatbot — one powered by AI that handles visitors automatically. No hiring, no contact forms that never get a reply. An AI chatbot solves this. It answers automatically, around the clock, using the information from your own website. It doesn't make things up. It doesn't sound robotic. And the best part: **you can install it in 5 minutes without touching any code**. ## Two Ways to Add a Chatbot to WordPress There are two ways to install a chatbot on WordPress. Each one makes sense depending on what you need. ### Option A: WordPress Plugin You install a plugin from the WordPress directory (Tidio, WPBot, Collect.chat...). Everything is managed from the WordPress dashboard. It's convenient if you already have everything centralized there. **The problem:** these plugins are heavy, they slow down your site, and they force you to configure dozens of options before you have anything working. Some don't even use real AI — just predefined flows like "if they say X, reply Y". ### Option B: External Script You use an external platform and paste a single line of code on your site. The chatbot loads from outside, doesn't consume your server resources, and is usually easier to set up. This is the option we recommend if you want something lightweight, quick to install, and powered by real AI — a bot that understands questions and responds with your information, not canned replies. In this guide we're going with Option B. It's faster and doesn't affect your site's performance. ## What to Look for in an AI Chatbot for WordPress Before choosing a platform, there are four things you should check. - It trains on your content A generic chatbot is useless. You need one that reads your website, your FAQs, your documents, and answers based on that information. - No hallucinations Many AI chatbots make things up when they don't know the answer. Look for one with RAG: it only responds if it finds the information in your data. - Simple installation If you need 45 minutes and a YouTube tutorial, something is wrong. Look for something that works with a single-line script. - Customizable Colors, response tone, welcome message. The chatbot widget should look like part of your website. ## How to Install an AI Chatbot on WordPress Step by Step Here's how to do it on your WordPress site with [Bravos AI](https://bravos-ai.com/) — it checks all four boxes above and we give you 7 days free of the full PRO plan to see it on your own site. ### Step 1: Create Your Account Go to bravos-ai.com and [sign up](https://app.bravos-ai.com/register?lang=en). We give you 7 days free of the full PRO plan. We ask for a card to start the trial, we notify you before the first charge, and you can cancel in one click. ### Step 2: Create a Chatbot and Train It Give your chatbot a name and add your data sources. This is where most people make mistakes, so pay attention. You have three ways to feed your bot: - **URLs from your website:** The quick option. The bot scans the page and extracts the content. The downside: it also grabs headers, footers, menus... noise that clutters the context. - **PDF documents:** The recommended option. Create a PDF with your FAQs, service descriptions, and policies. Clean content, no noise. - **CSV files:** If you have a product or service catalog with names, prices, and features... a CSV is perfect. The rule of thumb: avoid URLs if you want the best results, use PDFs for important information, and CSVs for catalogs. If you want to dive deeper into why chatbots fail and how to fix it, we have a [technical guide here](https://bravos-ai.com/blog/por-que-tu-chatbot-falla-y-como-solucionarlo). ### Step 3: Customize Adjust the colors to match your brand. Write a welcome message that invites visitors to ask questions. Set the response tone — more formal if you're a law firm, more casual if you're an online store. You can also add custom instructions: "Don't mention pricing unless asked", "Always suggest scheduling a call", etc. The bot follows them. ### Step 4: Test It Test your chatbot on the **Bots → Preview** page. Make sure that: - It responds with your information, not generic answers - It doesn't make up data you haven't provided - The design looks good on both mobile and desktop ### Step 5: Copy the Script When you're done, go to the **Bots → Embed Code** page and the platform gives you a script. It's a single line of code like this: Script format ### Step 6: Paste It in WordPress **Option A: With a plugin (recommended)** Install "WPCode" or "Insert Headers and Footers". Go to the plugin settings, paste the script in the "Footer Scripts" section, and save. Works with any theme. **Option B: With a page builder (Elementor, Divi, etc.)** Go to the theme builder and edit your footer. Once inside Elementor, for example, add an HTML widget/block and paste the script inside it. **Option C: Editing footer.php (advanced)** Go to Appearance → Theme Editor → footer.php. Paste the script right before . Note: if you update the theme, you'll lose this change. ### Step 7: Test It on Your Website Open your website (outside of WordPress admin) and check that the chatbot widget appears and responds correctly. If it doesn't load, make sure the script is properly pasted. ## Common Mistakes When Installing a Chatbot on WordPress After seeing many installations, these are the most common mistakes. ### Choosing a chatbot plugin that slows down WordPress Some WordPress chatbot plugins load heavy scripts that hurt your page speed. If your site takes more than 3 seconds to load, you lose visitors. Before installing anything, measure your speed with PageSpeed Insights. ### Not training the bot with your content An untrained chatbot is a useless chatbot. If it only replies with generic phrases like "Thanks for your message, we'll get back to you soon", it adds no value. Spend 10 minutes uploading your information. ### Set it and forget it A chatbot is not something you install and forget about. Review the conversations every week. You'll find questions it can't answer — and that tells you exactly what information you need to add. If you'd rather skip that manual review, you can [connect your chatbot to Claude or ChatGPT](https://bravos-ai.com/blog/que-es-un-mcp) and let the AI itself go through the conversations and tell you what's missing. ### No fallback plan What happens when the bot doesn't know something important? It should be able to redirect to email or a contact form. Always configure a fallback for cases it can't handle. ## Conclusion: Your WordPress AI Chatbot in 5 Minutes Adding an AI chatbot to WordPress is easier than you think. You don't need to code, you don't need heavy plugins, and you can have it running in minutes. What matters is choosing a platform that trains on your content, doesn't hallucinate, and installs without complications. And if your WordPress site is a WooCommerce store, we have a dedicated guide: [a WooCommerce chatbot that understands your catalog](https://bravos-ai.com/blog/chatbot-woocommerce) — with the catalog synced (variations, stock and sales up to date), not just your site's pages. And if your site isn't on WordPress but on Wix? There's a guide too: [Wix chatbot](https://bravos-ai.com/blog/chatbot-wix). ## WordPress Chatbot FAQ ### Ready to try it? Build your chatbot in 5 minutes and try the PRO plan free for 7 days. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # WooCommerce Chatbot: How to Get One That Actually Understands Your Catalog **Category:** Use case | **Read time:** 13 min | **Date:** July 3, 2026 | **URL:** https://bravos-ai.com/en/blog/woocommerce-chatbot If you run a WooCommerce store and search for "WooCommerce chatbot", you'll land on plugin listicles that get rewritten every January. They all compare the same things: pricing, install counts and dashboard screenshots. And they all skip the only question that matters: **will the bot actually know what you sell?** Because that's the real problem. A chat that greets people beautifully but replies "I don't have that information" when asked about a size, a price or whether something is in stock doesn't save you work — it adds to it. This guide covers what actually decides whether a chatbot works on a WooCommerce store: variations, stock, sales, taxes, and how all of that stays up to date. No plugin list and no year in the title, because none of this will change next January. Index - [The problem: a chat that doesn't know what you sell](#problem) - [What a chatbot has to understand about a WooCommerce store](#what-it-needs) - [Do you need a WooCommerce chatbot plugin?](#plugin-or-not) - [How it connects (step by step, no coding)](#how-it-works) - [What happens when you change a price or run out of stock](#updates) - [What it looks like in a real conversation](#conversations) - [When you don't need it](#when-not) - [Summary and FAQ](#summary-faq) ## The problem: a chat that doesn't know what you sell Most chats installed on WooCommerce stores come in two flavors. The first is **live chat**: a little window where the customer writes and you (or someone on your team) answer by hand. It works — but only when someone is on the other side, and store questions arrive at 11 pm. The second is the **canned-response bot**: customer clicks "shipping", the shipping text appears; clicks "returns", the returns text appears. Fine for general questions. But the moment someone types "do you have mountain bikes in size M under $900?", it's over: that answer isn't behind any button, because it lives in your catalog and changes every week. And here's the trap with many "AI" chatbots: they bolt a language model onto the bot, but feed it your web pages as if they were running text. The result is a bot that writes beautifully and gets the important part wrong, because a catalog isn't text: it's **data with attributes** (price, size, stock, category) that needs filtering, not "similarity". We cover this in depth in [why most chatbots can't find products in your catalog](https://bravos-ai.com/blog/chatbot-catalogo-productos). ## What a chatbot has to understand about a WooCommerce store WooCommerce has quirks of its own that no "best chatbots" article mentions — and they're exactly where generic bots fail. If you're evaluating any solution (ours or anyone else's), ask about these four things: ### 1. Variations In WooCommerce, a bike that comes in three sizes and two colors isn't one product: it's a "parent" product and six child variations, each with its own price, its own stock and sometimes its own photo. A bot that only reads the product page sees the parent and misses the rest. The most common question a store customer asks — "do you have it in size M?" — is answered at the variation level, not the parent. If the chatbot doesn't go down to that level, it can't sell. ### 2. Stock, which WooCommerce expresses three different ways Some stores manage exact quantities ("4 left"), some only mark in stock or out of stock, and some variations inherit stock from the parent product. There's also the "available on backorder" status — purchasable even with no units on hand. A serious chatbot has to understand all three forms and never invent quantities where the store only says "in stock". ### 3. Prices as the customer sees them If a helmet is on sale, your customer sees "was $89, now $69" on your site — and the bot should say exactly that, not just one of the two numbers. Then there's tax: some stores save prices tax-inclusive, others save them tax-exclusive and add it at display time. If the bot reads the raw number from the database, it may quote a price that doesn't match what the customer has right in front of them on the product page. The correct rule is simple to state and easy to get wrong: **the bot quotes the price exactly as the store displays it**. ### 4. The real-world mess of WordPress Product descriptions on a real WooCommerce store are full of leftovers from visual page builders (Elementor, Divi, WPBakery): bracketed codes the customer never sees but that live inside the text. If the bot is fed that without cleaning, the garbage ends up in its answers. And if your store uses a translation plugin like WPML or Polylang, every product exists duplicated in several languages — a bot that swallows everything indiscriminately will mix languages in its replies. This is the exam we'd give any chatbot before putting it on a WooCommerce store: does it answer per variation or only per parent product? Does it distinguish "4 left" from "in stock"? Does it quote sales with the before-and-after price? Does its price match the product page, tax included? If any answer is no, the bot will fail precisely on the questions that sell. ## Do you need a WooCommerce chatbot plugin? It's the most common search and the decision almost nobody explains to you. There are two ways to put a chatbot on a WooCommerce store, and the difference matters more than the brand you pick: **A plugin installed inside your WordPress.** Everything runs on your hosting: the bot, its logic and its AI calls. The upside is it installs from the plugin directory like anything else. The downsides are the usual WordPress ones: it's one more piece loading your server (and WooCommerce stores typically live on shared hosting, where every plugin counts), one more piece to update, and one more piece that can conflict with your theme or another plugin. On top of that, the AI quality is capped by whatever that plugin does internally — and processing a catalog properly (variations, filters, structured search) is infrastructure work that hardly fits inside a plugin. **An external service connected through the WooCommerce API.** The catalog syncs outside, on the chatbot's platform, and the only thing living on your site is the chat widget (a few lines of code, like Analytics). Your hosting processes nothing: not the AI, not the search, not the conversations. The connection uses the REST API WooCommerce ships with — the same one management apps and invoicing systems use — with keys you generate yourself and can revoke anytime. | | Plugin inside WordPress | Service connected via API | | --- | --- | --- | | Load on your hosting | Everything runs on your server. | Only the chat widget; the rest runs outside. | | Maintenance | One more plugin to update, one more that can break with your theme. | Nothing to update in your WordPress. | | AI quality | Capped by what fits inside the plugin. | Dedicated infrastructure (structured search, real filters). | | If you uninstall | Leftovers in your database, like almost every plugin. | Remove the widget and revoke the key. Your WordPress stays intact. | | Installation | From the plugin directory. | Generate an API key + paste the widget. | We built Bravos AI the second way, and not by accident: the heavy lifting of understanding a catalog (variations, stock, "under $X in size Y" filters) needs its own database and search infrastructure. That can't be done well inside the customer's shared hosting. ## How it connects (step by step, no coding) Here's the full process with Bravos AI. No code to touch, nothing to install in your WordPress: - **Generate API keys in your store.** In your WordPress: WooCommerce → Settings → Advanced → REST API → Add key. Set the description to "Bravos AI", choose *Read/Write* permissions (needed so your store can notify us of changes) and copy the two keys that appear — they're shown only once. - **Paste them in the Bravos dashboard.** Under Integrations → WooCommerce: your store address and the two keys. When you test the connection you'll see what we detected: how many products you have, your currency, and how your store handles tax (prices tax-inclusive, tax-exclusive, or taxes disabled). If you have a multilingual plugin like WPML or Polylang, we let you choose which language the bot will work with. - **Choose how much stock the bot should reveal.** Three options: don't mention stock at all, say only in stock / out of stock, or give exact quantities if your store manages them. - **Connect, and the catalog imports itself.** Every product comes in with its name, category, description (already cleaned of page-builder codes), price as your store displays it, variations with their stock, and the link and photo so the bot can show them in the conversation. - **Put the widget on your site.** A few lines of code, pasted once. If your site is WordPress, the full guide is in [how to add a chatbot to WordPress in 5 minutes](https://bravos-ai.com/blog/chatbot-wordpress). From there, the bot answers about your catalog, your pages and your documents, in the customer's language (13+ languages, without you translating anything). ## What happens when you change a price or run out of stock This is the question you should ask any vendor, because a bot with an outdated catalog is worse than no bot: it recommends sold-out products and quotes old prices. With Bravos it works in two layers. **Instantly:** when you edit a product in your WooCommerce (price, stock, description, anything), your store notifies us and the change shows up in the bot within seconds. **And every night:** the full catalog is checked against your store, in case any notification got lost — which can happen in WordPress, because deliveries depend on an internal mechanism that's unreliable on low-traffic stores. The honest version: changes arrive instantly when your store sends the notification, and the nightly check guarantees nothing stays behind for more than a day. If the API key you generate is read-only, the connection works the same but without the "instant" layer: the catalog updates every night. The dashboard shows you which mode you're on with a badge ("Real time" or "Daily sync") — no fine print. This also means there's no "retraining" and no re-sync button to remember when you change products. It maintains itself. If you want to see how other platforms compare on this, we analyzed it in [how to keep your chatbot data up to date without Zapier or n8n](https://bravos-ai.com/blog/mantener-chatbot-actualizado). ## What it looks like in a real conversation Picture a cycling store. This is what changes when the bot actually filters the catalog instead of matching similar text: Yes — we have 3 mountain bikes in size M under $900: the **Ridge Trail 29″** ($849, in stock), the **Sierra Sport** ($759, was $899 — currently on sale) and the **Summit Base** ($649, 2 left). Want to see any of them in detail? }, , , ]} /> Look at the details, because that's where the difference lives: the price filter is real (no $950 bike sneaking in "because the text looks similar"), the sale is quoted with the before-and-after, the size is a variation with its own stock, and "available on backorder" isn't confused with sold out. All of that comes straight from the WooCommerce data, not from the model's imagination. Same for questions that aren't catalog questions — shipping, returns, sizing, warranty — because the bot is also fed your pages and documents. The combination of both is what makes the experience seamless for the customer. ## When you don't need it Honesty first — that's the whole point of not being a plugin listicle: - **If your catalog is a handful of products with no variations** and your customers always ask the same four questions, a free canned-response bot may be enough. - **If you barely get any inquiries**, fix that with a good FAQ page first. A chatbot amplifies support demand you already have; it doesn't create it. - **If what you want is human chat during business hours** and you have someone to staff it, simple live chat does the job. The chatbot shines exactly where the human doesn't reach: nights, weekends, and the 40 repeated questions a day. For everyone else — stores with a real catalog, variations, changing sales and customers asking at all hours — a catalog-connected chatbot is where it pays for itself. For the broader online-store case (abandoned carts, purchase doubts), there's more in our [ecommerce chatbot guide](https://bravos-ai.com/blog/chatbot-tienda-online). And if your store isn't WooCommerce, we have a guide for the [Shopify chatbot](https://bravos-ai.com/blog/chatbot-shopify) and the [PrestaShop chatbot](https://bravos-ai.com/blog/chatbot-prestashop). ## In summary - The criterion for choosing a WooCommerce chatbot isn't the feature list: it's whether it understands your catalog — variations, stock, sales and taxes as the customer sees them. - You don't need a plugin: connecting through the WooCommerce REST API leaves your hosting alone and the AI runs on dedicated infrastructure. - Connecting with Bravos AI is an API key + pasting the widget. No coding, nothing installed in your WordPress. - Catalog changes arrive instantly when your store sends the notification, and a full nightly check guarantees nothing stays outdated. - Works with WPML/Polylang (you pick the catalog language) and answers your customers in 13+ languages. ### Is there a free WooCommerce chatbot? Free canned-response bots, yes. A real AI chatbot that understands your catalog and is free forever, no — AI has a cost per answer and somebody pays it. What does exist is the trial: with Bravos AI you get the full PRO plan for 7 days; we notify you before charging and if you cancel before day 7 you pay nothing. ### Does a WooCommerce chatbot work with variable products (sizes, colors)? With Bravos AI, yes — and it's one of the things we cared most about getting right: every variation syncs with its own price, stock and photo, and the bot answers at the variation level ("do you have size M left?"). Grouped and external/affiliate products aren't synced for now, and the dashboard shows how many were skipped — no silent truncation. ### Will a chatbot slow down my WooCommerce store? Not if it connects through the API: the only thing living on your site is the chat widget, which loads like any external script without touching page load time. The catalog is processed outside your hosting, and the sync is designed to treat it gently: few simultaneous requests at a calm pace, built for the shared hosting where most WooCommerce stores live. ### Do I need to know how to code to add a chatbot to WooCommerce? No. The only two technical steps are generating the API key in WooCommerce (four clicks in your WordPress dashboard) and pasting the chat widget on your site once. Everything else is configured from the Bravos dashboard. ### What happens when I change a price or a product sells out in my WooCommerce store? With Bravos AI, your store sends the notification and the bot reflects the change within seconds. On top of that, the full catalog is checked every night in case any notification got lost. No retraining, no buttons to remember. ### Does the chatbot work with multilingual WooCommerce stores (WPML or Polylang)? Yes. On connection the multilingual plugin is detected and you choose which catalog language the bot will use; translation duplicates are discarded so languages don't get mixed in the answers. And this is independent from the language the bot serves customers in: it replies in the customer's language even if your catalog is in just one. ### Does the chatbot work if my product descriptions are built with Elementor or Divi? Yes — descriptions are cleaned before they reach the bot. The internal page-builder codes (those bracketed snippets your customer never sees) are stripped during import, so the bot works with the actual product description. ### Connect your WooCommerce and try it on your real catalog Generate the API key in your store, connect it in the dashboard, and within minutes the bot answers about your products — variations, stock and sales, all up to date. 7-day PRO trial, no commitment: we notify you before charging and if you cancel before day 7 you pay nothing. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en) --- # Why Your AI Chatbot Fails and How to Fix It **Category:** Practical guide | **Read time:** 10 min | **Date:** Dec 29, 2025 | **URL:** https://bravos-ai.com/en/blog/por-que-tu-chatbot-falla If you're building a RAG chatbot for your business, you've probably already hit the chatbot hallucination problem: the bot confuses topics, gives wrong answers, or flat-out makes up information that doesn't exist. Just uploading documents **doesn't work**. The good news: you can use AI itself to prepare your content properly. In this article, we give you the **exact prompts** to fix chatbot hallucinations step by step. ## Why Your RAG Chatbot Gives Wrong Answers RAG (Retrieval-Augmented Generation) systems work by splitting your documents into small fragments called *chunks*. When a user asks a question, the system searches for the most relevant chunks and generates a response based on them. Some variants avoid splitting the text at all, like [PixelRAG](https://bravos-ai.com/blog/pixelrag-chatbot-empresarial), and others drop the vector database entirely, like [vectorless RAG](https://bravos-ai.com/blog/rag-sin-vectores), though most chatbots use this classic method. The problem is that: - **Chunks lose context:** A fragment might contain information without any indication of what topic it belongs to. The AI sees text but has no idea where it came from. - **Similar topics get confused:** "Return policy" and "Exchange policy" sound similar enough that the RAG system retrieves the wrong chunk, leading to chatbot wrong answers. - **Content is full of garbage:** Navigation menus, footers, HTML tags, cookie banners — all of this noise confuses the retrieval system and degrades AI chatbot accuracy. ## Step 1: Extract All Content from Your Website The first step to train your chatbot properly is extracting all your content. Models like Claude or ChatGPT can analyze entire websites and extract structured content for you. ### Step 1.1: If Your Website Blocks AI Access Some websites have anti-bot protections, and the AI can't access them directly. Don't worry — if you don't mind feeling like a hacker for five minutes, there's a solution. Ask the AI to generate a scraping script you can run in your terminal. This script will crawl all your website pages, save the content into text files, and compress them into a ZIP. Then you just upload that ZIP back to the AI and continue with Step 2. It sounds complicated but it really isn't. You literally copy, paste, and hit Enter. The AI walks you through it step by step. Prefer not to deal with it? At [Bravos AI](https://bravos-ai.com/) we can do this for you. We've already cleaned and organized content for several clients. Just reach out and we'll take care of it. ## Step 2: Clean Your Training Data for RAG Web content comes loaded with garbage: residual HTML code, repeated navigation menus, duplicate text blocks, encoding errors. All of this noise is one of the main causes of RAG chatbot hallucination — the system retrieves irrelevant fragments and the AI tries to make sense of nonsense. ## Step 3: Optimize Titles to Prevent Chatbot Hallucinations This is **the most important step** for RAG content preparation. Generic titles are the number one reason chatbots confuse similar topics. When two sections have vague headings, the embedding vectors end up too close together, and the retrieval system grabs the wrong one. ## Step 4: Add Context That Survives RAG Chunking When the RAG system splits your document into chunks, each fragment must be understandable on its own. If a chunk just says "The price is $49/month" without mentioning what product or service it refers to, the chatbot has no way to give an accurate answer. This is a core cause of chatbot wrong answers. Before applying this step, check what chunk size your application uses. It's usually in the documentation or configuration. For example, at Bravos AI we use 800-character chunks with 150-character overlap. Adapt the prompt below to your specific RAG chunking optimization needs. ## Step 5: Verify All Links in Your Content There's nothing worse than a chatbot that confidently shares broken links. If your content includes URLs, you need to verify every single one before uploading it to your RAG system. ## Step 6: Configure Your System Prompt to Stop Chatbot Hallucinations Even with perfectly prepared content, your chatbot can still make up information if you don't explicitly tell it not to. This is the chatbot hallucination fix that most people skip — and it's the easiest one to implement. Below are the minimum anti-fabrication rules. If you want to go deeper into how to write the full system prompt (role, context, rules, tone, common mistakes and sector templates), we cover it step by step in our [guide on how to write a system prompt for your business chatbot](https://bravos-ai.com/blog/system-prompt-chatbot-empresa). A case where this matters most is a law firm: see how a [law firm chatbot](https://bravos-ai.com/blog/chatbot-para-abogados) is configured to inform and refer without giving legal advice. This step is critical. Without these anti-hallucination rules, your chatbot can confidently present fabricated data — phone numbers, prices, addresses — that don't exist. This is the most common chatbot hallucination problem and the easiest to prevent. ## Step 7: Test with Critical Questions to Validate AI Chatbot Accuracy Before going live, test your chatbot with two types of questions designed to expose problems: - **Confusion questions:** Ask about topics that could easily be mixed up (e.g., "What's your return policy?" vs. "What's your exchange policy?"). If the bot gives the same answer for both, your titles aren't differentiated enough. - **Trick questions:** Ask about something that is NOT in your knowledge base (e.g., a product you don't sell, a service you don't offer). If the bot makes up an answer instead of saying it doesn't know, your system prompt needs stronger anti-hallucination rules. ## Final Checklist: RAG Content Preparation {[ "Content fully extracted", "Garbage and duplicates removed", "Titles differentiated with unique keywords", "Introductory context in each section", "Context markers every ~500 characters", "Links verified", "System prompt with anti-hallucination rules", "Tests passed (confusion + trick)" ].map((item, i) => ( - ))} ## Conclusion: Content Preparation Is the Key to Fix Chatbot Responses Proper RAG content preparation is the difference between a chatbot that frustrates your customers and one that genuinely helps them. Yes, it takes time to do it right. But with the prompts in this guide, you can use AI itself to do the heavy lifting. If after following all these steps your chatbot still gives wrong answers or makes up information, the problem is probably not your content — it's how the system searches and retrieves information. That's where a properly implemented RAG architecture matters, and it's exactly what we've built at Bravos AI. And if you don't know where to start looking for the fault, you can ask the AI itself: by connecting your chatbot to Claude or ChatGPT with [our MCP connector](https://bravos-ai.com/conector-ia-mcp), it's the assistant that goes through your conversations and tells you which questions are going unanswered. Because it carries our team's judgment inside, it doesn't stop at spotting them: it also suggests how to fill those gaps. ### Is your chatbot still failing? Try Bravos AI's PRO plan free for 7 days and compare the results. [Try PRO free for 7 days ](https://app.bravos-ai.com/register?lang=en)