How to Create a WhatsApp AI Chatbot for Your Business (2026 Guide)

Difficulty

Intermediate

Time required

3 hours

Tools needed

• A phone number not tied to personal WhatsApp
• Meta business portfolio
• A WhatsApp provider or Meta Cloud API access
• An LLM API or chatbot platform
• Your FAQs and policies

A WhatsApp AI chatbot is still one of the most useful things a small business can set up in 2026, and it's allowed, as long as the bot is there to serve your customers. The short version: sign up for the WhatsApp Business Platform (not just the free Business app) through a chatbot provider or Meta's Cloud API, feed the AI your own FAQs and policies, and make sure a human can take over when the bot gets stuck.

What you can't do anymore is run a general-purpose "ask me anything" AI assistant on WhatsApp. Meta's platform terms now bar AI providers whose AI is the main product being offered, while explicitly allowing AI that's "incidental or ancillary" to a business. A support bot for your bakery, clinic or online store falls on the allowed side. A ChatGPT clone doesn't.

Below, we'll explain the two WhatsApp products, the rules that trip people up (the 24-hour window, templates, opt-in, pricing), and then walk through setup on either a no-code platform or a small custom build. Plan on an afternoon for a basic bot, longer if you're writing code.

WhatsApp Business app vs WhatsApp Business Platform

These two get confused constantly, and the difference decides everything else.

The Business app's built-in features cover greeting messages, away messages outside business hours, quick replies, labels and a product catalog. That's enough for many shops. Meta also rolled out Meta Business Agent, an AI agent that answers questions, recommends products from your catalog and books appointments, announced for businesses globally on June 3, 2026. Meta said activation was free at launch and that access would move to "paid subscription offerings" in the following months, so check current terms in the app.

If you want your own AI model, your own knowledge base, a shared team inbox or links to your CRM, you need the Business Platform. The rest of this guide is about that.

The rules you have to design around

Read this section before you build anything. Most failed WhatsApp bots break one of these.

The 24-hour customer service window

When a customer messages you, a 24-hour window opens, and it resets with each new user message. Inside it, your bot can send free-form replies: AI answers, images, links, anything. Once it closes, you can only send pre-approved message templates. So an AI bot can chat freely with anyone who writes to you, but it can't start a conversation or follow up two days later with an improvised message.

Templates, and what they cost

Since July 1, 2025, Meta charges per delivered template message rather than per conversation. As of September 2026:

• Service messages (non-template replies inside the window) are free.
• Utility templates (order updates, appointment reminders) are free inside an open window and charged outside it.
• Authentication templates (one-time codes) are charged outside the window.
• Marketing templates are charged every time.

Rates vary by the recipient's country code, and Meta publishes rate cards on the same page. There's also a 72-hour free window when a customer reaches you through a click-to-WhatsApp ad or a Facebook Page button. For a support bot that mostly answers incoming questions, Meta's own fees can be close to zero. Your provider's fees are another matter (more on that below).

Opt-in

You can only message people who gave you their number and opted in to hear from you, per the WhatsApp Business Messaging Policy. How you collect opt-in is up to you, but it has to be clear and comply with local law. A checkbox at checkout saying "Send me order updates on WhatsApp" is the usual approach.

A human escape hatch

The same policy requires "prompt, clear, and direct escalation paths," such as a handoff to a human agent in the chat, or a phone number, email, web form or store visit. An AI bot with no way out breaks the rules as well as annoying customers.

The AI provider restriction

Meta's WhatsApp Business Platform terms define "AI Providers" as developers of LLMs, generative AI platforms and "general-purpose artificial intelligence assistants," and prohibit them from using the platform when that AI is "the primary (rather than incidental or ancillary) functionality being made available." According to Meta's AI provider pricing page, this took effect with the terms update of January 15, 2026. The only exception is where Meta is legally required to allow general-purpose assistants. In those markets Meta charged AI providers for each non-template message: in 29 European countries from February 16 to May 12, 2026, and in Brazil from March 11, 2026.

What's still allowed, in the terms' own words: ancillary or incidental AI use, hiring an AI company as your solution provider, and fine-tuning a model purely for your own business. Meta's pricing page is explicit that the AI provider policy "does NOT change how Meta charges all other businesses," including those using AI for customer support.

Choose your path

Path A: a no-code chatbot platform. Best for most small businesses. A provider handles WhatsApp onboarding, gives your team a shared inbox and has an AI agent you train on your content. Some options, with prices as listed in September 2026:

• respond.io: Starter $79/month, Growth $159/month, Advanced $279/month (monthly billing). AI agents are included on Growth and above, not on Starter. Meta fees are charged separately. 7-day free trial.
• WATI: Growth from $29/month, Pro $99/month, Business $399/month. It includes a chatbot builder on all plans. Its Astra AI agents are a paid add-on on Pro and Business. Message charges are separate.
• Tidio: a free plan and paid plans from $24.17/month. Its Lyro AI agent is an add-on from $32.50/month, and its pricing page lists WhatsApp across plans. Tidio is website-chat first, so it suits you if WhatsApp is one channel among several.
• Twilio: a developer platform rather than a no-code tool. It adds $0.005 per inbound or outbound message on top of Meta's fees. Pick it if you have a developer and want Twilio's APIs, not a ready-made inbox.

If you're comparing AI support tools across channels, our roundup of the best AI chatbots for customer support goes deeper on features and pricing.

Path B: Cloud API plus n8n or a small webhook server. You connect directly to Meta's Cloud API, receive messages on a webhook, send them to an LLM with your knowledge base, and post the answer back. You pay Meta's fees and your LLM usage, with no platform subscription. You also build the inbox, handoff and monitoring yourself. It's a good fit if you already use n8n; our beginner guide to automating your business with n8n and AI covers the basics, and Zapier vs Make vs n8n helps if you haven't chosen a tool yet.

For most businesses we'd start with Path A. The subscription usually costs less than the hours you'd spend building a decent human handoff inbox.

Step-by-step setup

1. 1

   Step 1: Get a dedicated business phone number

   Meta's phone number rules say the number must be yours, have a country and area code (no short codes), and be able to receive a call or SMS for verification. A number active on regular WhatsApp can't be registered until that account is deleted. Once registered, you can still use it for ordinary calls and texts, just not in the consumer WhatsApp app.

   A cheap separate SIM or a VoIP number that can receive SMS or calls is the simplest choice. If you already run your business on the WhatsApp Business app, ask your provider about coexistence (see the callout above) rather than switching numbers.

2. 2

   Step 2: Set up and verify your Meta business portfolio

   You need a business portfolio in Meta Business Manager (Meta Business Suite). Path A providers walk you through creating or connecting one during signup. On Path B, you'll create a Meta app with the "Connect with customers through WhatsApp" use case and attach it to your portfolio.

   Then start business verification. You'll upload documents proving your legal business name and address. It isn't strictly required to send your first messages, but it matters. New portfolios can start conversations with only 250 unique people per 24 hours, and verification is one route to the 2,000 tier. Pick a display name that clearly matches your business, because Meta reviews it.

3. 3

   Step 3: Connect WhatsApp through your chosen path

   Path A: in your platform, choose "Connect WhatsApp," log in with Facebook, select your portfolio, and add and verify your number. Most providers use Meta's Embedded Signup, which takes a few minutes if your details are in order.

   Path B: in the Meta App Dashboard, send a test message from Meta's test number, then create a permanent access token through a system user with the whatsapp_business_messaging and whatsapp_business_management permissions. Next, register your webhook URL. Meta sends a verification GET request with hub.mode, hub.verify_token and hub.challenge, and your server must echo the challenge back (Meta webhooks docs). Here's a minimal Node.js sketch of the whole loop:

   // Sketch only: add error handling, retries and logging before production.
   import express from "express";
   import crypto from "crypto";
   
   const app = express();
   app.use(express.json({ verify: (req, _res, buf) => (req.rawBody = buf) }));
   
   const { VERIFY_TOKEN, APP_SECRET, WA_TOKEN, PHONE_NUMBER_ID } = process.env;
   
   // 1. Webhook verification handshake
   app.get("/webhook", (req, res) => {
     if (req.query["hub.mode"] === "subscribe" &&
         req.query["hub.verify_token"] === VERIFY_TOKEN) {
       return res.send(req.query["hub.challenge"]);
     }
     res.sendStatus(403);
   });
   
   // 2. Incoming messages
   app.post("/webhook", async (req, res) => {
     // Check the X-Hub-Signature-256 header so only Meta can call you
     const expected = "sha256=" + crypto
       .createHmac("sha256", APP_SECRET).update(req.rawBody).digest("hex");
     if (req.get("x-hub-signature-256") !== expected) return res.sendStatus(401);
   
     res.sendStatus(200); // acknowledge fast, then do the slow work
   
     const msg = req.body.entry?.[0]?.changes?.[0]?.value?.messages?.[0];
     if (!msg || msg.type !== "text") return;
   
     if (await isHandedOffToHuman(msg.from)) return; // bot stays quiet
   
     const answer = await askModel({
       system: SUPPORT_PROMPT,                          // see Step 4
       context: await searchKnowledgeBase(msg.text.body),
       history: await getRecentMessages(msg.from),
       question: msg.text.body,
     });
   
     if (answer.needsHuman) {
       await flagForHuman(msg.from); // notify your team (email, Slack, CRM)
     }
   
     // 3. Reply (free-form is fine: the customer just opened the 24h window)
     await fetch(`https://graph.facebook.com/v23.0/${PHONE_NUMBER_ID}/messages`, {
       method: "POST",
       headers: { Authorization: `Bearer ${WA_TOKEN}`, "Content-Type": "application/json" },
       body: JSON.stringify({
         messaging_product: "whatsapp",
         to: msg.from,
         type: "text",
         text: { body: answer.text },
       }),
     });
   });
   
   app.listen(3000);

   The helper functions (askModel, searchKnowledgeBase and so on) are placeholders for your LLM provider's SDK and your storage. In n8n, the same flow is a WhatsApp Trigger node, an AI Agent node with your knowledge base, and a WhatsApp Business Cloud node to reply. One quirk from n8n's docs: WhatsApp allows only one webhook per Meta app, so testing a workflow can overwrite your production webhook. Use a separate Meta app for testing if you can.

4. 4

   Step 4: Connect your knowledge base and write the bot's instructions

   Your bot is only as good as what you give it. Gather the material customers actually ask about: opening hours, pricing, shipping and returns, booking rules, product specs, and your 20 most common questions with your best answers. Paste your recent support emails into a chatbot and ask it to list the recurring questions if you don't know them offhand.

   On Path A, upload documents or add your website URL in the platform's AI agent settings. On Path B, the simplest approach is to keep a short FAQ document in the prompt. For larger catalogs, store it in a vector database and retrieve relevant chunks per question.

   Then write the system instructions. Here's an illustrative example for a fictional bike shop:

   You are the WhatsApp assistant for Riverside Bikes, a bike shop and
   repair service. Answer only questions about our products, repairs,
   bookings, orders, opening hours and store policies, using the
   KNOWLEDGE BASE below.
   
   Rules:
   - Keep replies short: 1-3 sentences, plain language, no markdown.
   - If the answer isn't in the knowledge base, say you're not sure and
     offer to connect them with a team member. Never guess prices,
     stock or repair times.
   - Politely decline unrelated requests (general trivia, homework,
     coding) and steer back to how you can help with their bike.
   - Hand off to a human immediately if the customer asks for a person,
     is upset, mentions a safety issue, or wants a refund or complaint
     handled. Set needsHuman = true.
   - Never ask for card numbers or passwords.
   
   KNOWLEDGE BASE:
   [paste FAQ, policies, hours, price list]

   That "decline unrelated requests" line matters for Meta's AI provider rules as well as for your LLM bill. For more on writing instructions like this, see our guide to writing better prompts.

5. 5

   Step 5: Design the handoff to a human

   Handoff is where cheap bots fall apart. Decide three things up front:

   • Triggers: the customer types "human," "agent" or "person"; the bot can't answer twice in a row; the topic is refunds, complaints, medical, legal or safety; the customer seems frustrated.
   • What happens next: the bot says something honest ("I'm passing you to Maria on our team, who's usually back within an hour during opening times"), pauses itself for that chat, and alerts your team with the conversation summary.
   • Out of hours: tell customers when a person will reply, and give an alternative (phone, email or form) so you meet Meta's escalation requirement.

   Path A platforms handle the pause-and-assign part in their inbox. On Path B, the isHandedOffToHuman check in the sketch above is the minimum: store a flag per customer and have your team reply from wherever they work. Remember the 24-hour rule applies to humans too. If your team replies after the window closes, they'll need a template.

6. 6

   Step 6: Test with real questions before going live

   Use Meta's test number or a private test number and run through a script of at least 30 real customer questions, including awkward ones:

   • Questions your knowledge base doesn't cover (does it admit it or make something up?)
   • Off-topic requests (does it decline politely?)
   • "Can I talk to a person?" (does the handoff fire, and does the bot stay quiet afterward?)
   • Voice notes, photos and messages in another language
   • Someone trying to get it to ignore its instructions

   Fix the knowledge base first, prompts second. Most wrong answers come from missing or outdated information, not the model.

7. 7

   Step 7: Create templates and launch

   Submit the templates you'll need outside the 24-hour window, like order confirmations, appointment reminders and "we've replied to your question" follow-ups, and wait for Meta's approval. Categorize them honestly. Meta decides whether a template is utility or marketing, and marketing costs more.

   Then add WhatsApp entry points: a click-to-chat link or QR code on your website, receipts and packaging, and your Google Business Profile. Opt-in checkboxes go wherever you collect numbers. Launch quietly, read every bot conversation for the first week or two, and keep updating the knowledge base with anything it got wrong.

What it costs in practice

A realistic monthly budget has three parts:

1. Meta's fees: often little or nothing for an inbound support bot, since service replies are free. Marketing and out-of-window utility templates cost extra, priced by country.
2. Platform or infrastructure: a Path A subscription (roughly $29 to $279 a month from the plans above, plus any AI add-ons), or hosting for your webhook or n8n instance on Path B.
3. AI usage: included in some platform plans, like respond.io Growth and above. Otherwise it's billed per conversation by the platform or per token by your LLM provider.

If WhatsApp is part of a bigger AI push, our pillar on the best AI tools for small businesses covers where else automation tends to pay off. If phone calls are your bigger bottleneck, see how to build an AI voice agent for your business.

Frequently asked questions

Is it allowed to use ChatGPT or another AI on WhatsApp Business?

Yes, if the AI supports your business, for example by answering questions about your products, orders or bookings. Since January 15, 2026, Meta's terms bar AI providers from using the WhatsApp Business Platform when general-purpose AI is the main thing they're offering. Using an LLM behind the scenes for your own customer service counts as incidental use, which is allowed.

Can I build a WhatsApp chatbot for free?

Partly. The WhatsApp Business app is free and has automated greeting and away messages plus quick replies. On the Business Platform, Meta doesn't charge for service replies inside the 24-hour window. You'll still pay for templates outside it, and usually for a chatbot platform or LLM usage. Some tools, like Tidio, have free plans with tight limits.

Can my chatbot message customers first?

Only with an approved template, and only to people who have opted in. Free-form AI replies are limited to the 24 hours after the customer's last message.

Do I need business verification to use the WhatsApp API?

You can start testing without it, but new business portfolios are limited to starting conversations with 250 unique people per 24 hours. Completing Meta business verification is one of the ways to raise that to 2,000, and it adds credibility to your profile.

Can I use my existing WhatsApp number?

A number on personal WhatsApp must be deleted from that app before you can register it on the platform. A WhatsApp Business app number can often be connected through coexistence, so you keep using the app alongside the API. Ask your provider whether it supports this.