To build a WhatsApp AI assistant, register your number on Meta’s WhatsApp Business Platform (Cloud API), connect a webhook to a small server or an n8n workflow, give the AI model your own business information to answer from (retrieval, or “RAG”), and add a handoff to a human. Since January 2026, Meta only allows assistants that support a real business service — not general-purpose chatbots.
This guide explains each step the way I plan these projects for clients: what Meta allows, how the pieces connect, what it costs, and how to test it before customers see it. If you want it built for you, see my AI chatbot and WhatsApp agent service.
What Meta allows in 2026
From 15 January 2026, Meta’s WhatsApp Business Solution terms prohibit general-purpose AI chatbots on the platform — assistants where the AI is the product, such as an open “ask me anything” bot. AI that supports your own business is still allowed. Typical permitted uses:
- Answering questions about your services, prices, timings and policies
- Booking appointments and sending reminders
- Order confirmations, delivery updates and order-status questions
- Qualifying leads and passing them to a salesperson
- Admissions and fee questions for schools and institutes
Keep the assistant focused on your business, and always give customers a way to reach a person.
WhatsApp Business App vs WhatsApp Business Platform
| WhatsApp Business App | WhatsApp Business Platform (Cloud API) | |
|---|---|---|
| Who it suits | One person or a small team replying by hand | Businesses that want automation, AI replies or several agents |
| Automation | Greeting, away messages and quick replies only | Full automation through webhooks and the API |
| AI assistant | Not possible | Possible (business-specific use only) |
| Cost | Free app | Per-message charges set by Meta, plus your hosting |
| Set-up | Install the app | Meta Business portfolio, verification, phone number registration |
An AI assistant needs the Platform. You can connect directly to Meta’s Cloud API, or through a Business Solution Provider (BSP) such as Twilio or 360dialog, which adds its own dashboard and fees.
How a WhatsApp AI assistant works
- A customer sends a message. Meta sends it to your webhook — a URL on your server or n8n instance.
- Your server finds the right information. It searches your knowledge base (FAQs, price list, policies) for the passages related to the question.
- The AI model writes a reply using only those passages and your instructions: tone, language, what it must never answer.
- Tools take actions when needed: check an order in your database, create a booking, save a lead to Google Sheets or your CRM.
- Your server sends the reply through the Cloud API.
- Handoff: if the question is outside the knowledge base, sensitive, or the customer asks for a person, the bot stops and notifies your team.
Step-by-step: building it
1. Prepare your Meta accounts
Create a Meta Business portfolio, complete business verification, add a WhatsApp Business account and register the phone number you will use. Use a number your business controls long-term — not a personal SIM that might change.
2. Choose where the logic runs
n8n is the fastest route for most small businesses: its WhatsApp trigger and AI Agent nodes handle the webhook, model calls and tools without much code, and it can be self-hosted to avoid per-task fees. A custom Node.js or Next.js server gives more control for high volume or complex integrations.
3. Build the knowledge base
Collect what customers actually ask: your top 30–50 questions from WhatsApp history, service and price lists, delivery or admission policies, timings and location. Write each answer once, clearly. Store it in a vector database such as Supabase pgvector so the assistant can retrieve the relevant parts. This retrieval step — RAG — is what keeps answers specific to your business instead of generic.
4. Write the instructions
Tell the model who it represents, which languages to use (English, Urdu, Roman Urdu), to answer only from the retrieved information, to say “I’ll connect you with our team” when unsure, and never to promise refunds, medical advice or prices that are not in the knowledge base.
5. Respect the 24-hour customer service window
When a customer messages you, a 24-hour customer service window opens and you can reply freely within it. Outside that window, you can only start a conversation with a pre-approved message template (for example, an appointment reminder). Design reminders and follow-ups as templates and submit them for approval early.
6. Add the human handoff
Route escalated chats to a shared inbox or a staff member’s WhatsApp, with the conversation history attached. Pause the bot on that chat until the person marks it resolved.
7. Test before launch
Run at least 50 real past questions through the assistant, including rude, off-topic and Roman Urdu messages. Check every answer against your knowledge base, fix gaps, then launch to a small group first.
What it costs to run
| Cost | How it is charged | Notes |
|---|---|---|
| Meta messaging | Per message, by category (marketing, utility, authentication, service) and country | Per-message pricing replaced per-conversation pricing on 1 July 2025. Meta has announced that from 1 October 2026 service replies are also charged; check the current rate card. |
| AI model | Per token (text processed) | For FAQ-style answers this is usually small per conversation; it grows with long documents and long chats. |
| Hosting | Monthly server or n8n plan | Self-hosted n8n on a VPS is often the cheapest option at volume. |
| Build | One-time | Depends on the knowledge base, integrations and languages. |
Rates change, so always check Meta’s WhatsApp pricing documentation for your recipients’ country before estimating. Conversations that start from a Click-to-WhatsApp ad or a Facebook Page button open a free entry-point window.
Launch checklist
- Business-specific purpose (no general-purpose chatbot)
- Verified Meta Business portfolio and registered number
- Knowledge base with your real top questions and answers
- Clear instructions: languages, tone, what never to answer
- Message templates approved for reminders and follow-ups
- Human handoff tested end-to-end
- Conversation logs reviewed weekly for wrong or missing answers
For broader automation — saving leads to sheets, sending order alerts, processing documents — see AI automation services, or read the wider guide to building an AI chatbot or agent.
