WhatsApp AI Chatbot: How to Build One on the WhatsApp Business Platform

What a WhatsApp AI chatbot is, and isn't

A WhatsApp AI chatbot is your business's assistant on WhatsApp. Customers message your business number; the bot understands free-text questions, answers from your own information, takes actions such as checking an order or booking a slot, and passes the conversation to your team when needed.

It runs on the WhatsApp Business Platform, the API version of WhatsApp for businesses, not the WhatsApp Business app on a phone. The platform gives programmatic access to messages, templates, interactive buttons and lists, and media, which is what makes automation possible at scale.

What it is not, since January 2026, is a general-purpose AI assistant offered to the public through WhatsApp. Meta's terms now restrict the Business Platform to AI that serves a business's own customers, which is exactly the use case most companies want anyway.

Most businesses use a WhatsApp AI chatbot for three jobs: answering customer questions at any hour, handling simple tasks such as order tracking, bookings and payment links without an agent, and sending timely notifications that customers have asked for. The best results come from starting with one of these, measuring it, and adding the next once it works well.

The WhatsApp platform rules that shape every bot

WhatsApp is strict about protecting users from spam, and the rules directly affect how a bot is designed.

The practical effect: a WhatsApp bot is mostly reactive and conversational inside the window, and uses a small set of well-designed templates to start conversations that customers actually want, such as delivery updates, appointment reminders and payment confirmations.

Meta's 2026 policy on AI chatbots

In October 2025 Meta changed the WhatsApp Business Solution terms to stop AI providers distributing general-purpose assistants through the Business Platform. The change applied to new users from mid-October 2025 and to all existing users from 15 January 2026, and services such as public ChatGPT-style bots on WhatsApp stopped working.

For most businesses this policy changes little in practice, but it is a good reason to document what your bot is for, which topics it covers and how it declines others, so the design is easy to explain if Meta or your own compliance team asks.

WhatsApp AI chatbot architecture

A production WhatsApp bot is a small service between the WhatsApp Cloud API and your systems, with an AI layer and an agent inbox.

• Webhook endpoint: receives messages and delivery statuses, verifies Meta's signature and responds quickly.
• Conversation store: history, the 24-hour window, opt-in status and whether a human has taken over.
• AI layer: retrieval from approved content plus tools for orders, bookings and CRM, with guardrails.
• Template sender: business-initiated messages through approved templates with variables.
• Agent inbox: handover with history and summary, and a way to hand back to the bot.

A minimal webhook: receive a message and reply

This is the core of a WhatsApp bot using the Cloud API: verify the webhook, check the request signature, read the incoming text and send a reply. The answer function is where your RAG chatbot plugs in.

# app.py — pip install flask requests
import hashlib, hmac, os
import requests
from flask import Flask, request, abort

app = Flask(__name__)
VERIFY_TOKEN = os.environ["WA_VERIFY_TOKEN"]      # you choose this in the Meta app settings
APP_SECRET = os.environ["META_APP_SECRET"].encode()
TOKEN = os.environ["WA_ACCESS_TOKEN"]
PHONE_NUMBER_ID = os.environ["WA_PHONE_NUMBER_ID"]
GRAPH = f"https://graph.facebook.com/v21.0/{PHONE_NUMBER_ID}/messages"

@app.get("/webhook")
def verify():
    if request.args.get("hub.verify_token") == VERIFY_TOKEN:
        return request.args.get("hub.challenge", "")
    abort(403)

def valid_signature(req) -> bool:
    sent = req.headers.get("X-Hub-Signature-256", "").removeprefix("sha256=")
    expected = hmac.new(APP_SECRET, req.get_data(), hashlib.sha256).hexdigest()
    return hmac.compare_digest(sent, expected)

def answer(user_id: str, text: str) -> str:
    # Plug in your RAG chatbot here: retrieve approved content, call tools,
    # keep answers about your business, and offer a human when unsure.
    return "Thanks! I can help with orders, bookings and account questions."

@app.post("/webhook")
def incoming():
    if not valid_signature(request):
        abort(401)
    for entry in request.json.get("entry", []):
        for change in entry.get("changes", []):
            for msg in change.get("value", {}).get("messages", []):
                if msg.get("type") != "text":
                    continue
                reply = answer(msg["from"], msg["text"]["body"])
                requests.post(GRAPH, timeout=10,
                    headers={"Authorization": f"Bearer {TOKEN}"},
                    json={"messaging_product": "whatsapp", "to": msg["from"],
                          "type": "text", "text": {"body": reply}})
    return "ok", 200

In production, acknowledge the webhook immediately and process messages in a queue, deduplicate by message ID because webhooks can be retried, and store each customer's last-message time to know when the 24-hour window closes.

Message types that make bots easier to use

Pure free text works, but WhatsApp's interactive features make bots faster and less error-prone, especially on small screens.

Let customers type freely at any point, though. A good bot treats buttons as shortcuts, not as the only way to talk, and the AI layer understands the free-text version of every choice.

Keep a consistent tone across buttons, lists and free-text replies, and make sure every interactive message still makes sense if read out of context in the chat history, because customers often scroll back days later to find an order number or booking time.

Making the AI answers reliable

WhatsApp conversations feel personal, so wrong answers do more damage than on a website widget. The same principles as any RAG chatbot apply, with a few WhatsApp-specific ones:

Handing over to people

The best WhatsApp bots are judged as much by how they hand over as by what they answer. Customers should never feel stuck with a bot.

• Offer a person whenever the customer asks, the bot isn't confident, or the topic is sensitive (complaints, payments, cancellations).
• Pass the full history and a short summary so the agent doesn't ask the customer to repeat themselves.
• Pause automatic replies while an agent is handling the conversation, and hand back to the bot when the agent closes it.
• Tell customers honestly when a person will reply, especially outside working hours.
• Track handover reasons; they are the best list of what to improve next.

Set expectations for response times by channel and hour. A bot that says a person will reply within minutes, when the team is offline until morning, damages trust more than an honest message about when someone will be available.

What a WhatsApp AI chatbot costs

There are two kinds of cost: building the bot, and running it. Running costs include Meta's charges for template messages, which vary by category and country, AI model usage, hosting, and any business solution provider's fees. Replies inside the 24-hour customer service window are generally not charged by Meta, which makes reactive support bots inexpensive to run; check Meta's current pricing before you plan volumes.

Ranges match our AI chatbot development page. You get a fixed quote after a discovery call.

Launch checklist

WhatsApp numbers carry a quality rating, and a poor launch can limit how many customers you can message. Work through this list before the bot meets real customers.

• Business verified, display name approved and the phone number registered on the Cloud API.
• Opt-in collected where customers expect it (checkout, account settings, forms) with clear wording, and stored with time and source.
• Templates submitted for the messages you will start: order updates, reminders, passcodes, with the right category for each.
• Webhook signature checks, retries, deduplication and a queue in place, and alerts if replies start failing.
• Answers tested on real customer questions in every language you support, including questions the bot should decline.
• Handover to the agent inbox tested end to end, including outside working hours.
• Privacy notice updated to explain the use of AI and how chat data is stored and for how long.

Common WhatsApp chatbot mistakes

These are the issues we most often see when businesses come to us to fix an existing WhatsApp bot.

Frequently asked questions