AI chatbot on WhatsApp: how it works, how to train it on your business and when to hand off to a human

What an AI chatbot on WhatsApp actually does

An AI chatbot on WhatsApp is a bot that answers free-text messages with a large language model (LLM) instead of fixed replies. When Ritu writes "is there any weekend batch for class 11 near Vaishali Nagar and what's the fee", a keyword bot sees no keyword it knows. An AI step reads the whole sentence, looks at the facts you gave it about batches and fees, and writes a natural answer.

Under the hood, every reply is one round trip: the customer's message, the recent chat and your instructions go to the model; the model returns text; the platform sends it on WhatsApp. The model does not learn from each chat or browse your website on its own. It knows what you put in front of it, and nothing more.

That makes AI good at three jobs on WhatsApp:

• Answering questions phrased in many ways, in Hinglish, with typos, in several languages.
• Understanding intent: is this person asking about fees, complaining, or asking for a human?
• Pulling details out of a sentence, for example a name, a date and a city from "I'm Rahul, need a 2BHK in Wakad by March".

It is a poor fit for anything that needs live account data it cannot see (today's order status, a fee balance), legal or medical judgement, and complaints, where people want a person. For a comparison of AI with keyword and button bots, see keyword vs button vs AI chatbots.

Training it on your own content

"Training" here does not mean building your own model. It means writing the context the model receives with every question. In Broadchat you do this in AI → AI Agents: click Create New AI Agent, give it an Agent Name and write its Instructions. The instructions are the agent's whole briefing, so write them like an induction note for a new front-desk employee.

Three habits make the difference between a useful agent and an embarrassing one:

1. State facts as facts. "Weekend batch only at Vaishali Nagar" beats a pasted brochure full of marketing lines. Models repeat what you give them.
2. Write the refusals. Tell the agent what it must not do: promise results, quote discounts, give medical or legal advice. Without rules, a model tries to be helpful and invents.
3. Keep one master copy. Keep the instructions in a shared document, update fees and dates there first, and paste the full text into the agent when anything changes, so no field is left with an old or empty version.

Test in AI → Playground: pick the agent, then ask the twenty questions your front desk hears most, plus five it should refuse. If an answer is wrong, the fix is almost always in the instructions. Facts that change daily (seat availability, order status) should not live in the instructions at all; look them up from a sheet or your system with an automation step, as described in the Google Sheets integration.

Putting the agent into a WhatsApp flow

An agent does nothing until a chatbot uses it. In Chatbots → Chatbots, the Quick Menu's AI Agent group has the AI steps. A design that keeps humans in charge looks like this:

The loop of steps 2 to 6 lets the customer ask several questions in a row, and each new message is checked for "I want a person" before the AI answers it. The No Answer / Error output covers technical failures (no agent selected, the AI service unavailable), not "the model did not know"; that case is handled by your instruction to offer a counsellor plus the intent check. The AI Data Collector step is the better tool when you need a set of fields (name, class, preferred batch) from a chatty customer; it asks follow-up questions until it has them.

When an agent takes over, they can click Pause Automation in Live Chat so the bot stays quiet for that contact. The no-code chatbot guide shows the hand-off steps in detail.

Human hand-off is a requirement, not a nicety

WhatsApp's Business Messaging Policy expects automated experiences to give people a way to reach a human. Beyond policy, a bot that traps people gets blocked and reported, which lowers your quality rating. Hand off when:

• the customer asks for a person, in any wording (that is what the intent check is for);
• the topic is money already paid, a complaint, health or anything legal;
• the AI has answered twice and the customer repeats the question;
• it is a high-value lead you do not want handled by a model at all.

Outside office hours, say honestly when a person will reply, and still assign the chat so it waits in the right queue.

Limits you should plan for

• Confident mistakes. LLMs can state wrong facts fluently. Keep instructions short and factual, and review the AI's answers in the chat history every week.
• No proactive messages. AI replies are normal messages inside the 24-hour customer service window. To follow up after it closes you need an approved template, written and approved in advance, not generated by AI on the spot.
• Text in, text out. Voice notes are not transcribed for the bot automatically, so offer buttons or ask people to type.
• Short conversations work best. Long chats dilute the context. If a conversation goes past a handful of turns, hand it to a person.
• Shared capacity. The platform AI is rate-limited per workspace; during a big campaign that triggers many replies at once, some answers can be delayed. Broadchat's platform AI runs on Google Gemini, with Groq and OpenAI as fallbacks where configured. If you need a specific OpenAI model on your own API key, use the Open AI or ChatGPT app step (Popular Apps), where each run uses a paid automation task.
• Cost of messages. Each business phone number gets 1,000 free service messages a month; after that Meta charges per service message (₹0.115 per message to India, Meta rate card effective 1 Oct 2026, excl. GST). A chatty AI uses that allowance faster than a button bot. See WhatsApp API pricing.

Meta's policy on AI and general-purpose chatbots

In October 2025 WhatsApp updated its Business Solution terms with a section on "AI providers". As reported by TechCrunch, from 15 Jan 2026 providers of large language models and general-purpose AI assistants may not use the WhatsApp Business Platform when that assistant is the primary functionality offered; think of a ChatGPT-style "ask me anything" number. Meta's pricing page also lists a separate pricing policy for AI Providers, in place since 16 Feb 2026.

What stays allowed is AI used inside a business's own customer experience: answering questions about your courses, products, orders and bookings, routing chats and collecting details. The test is simple. If your number exists to serve your customers about your business and the AI is one tool in that, you are on the right side. If the number's main purpose is to be a general chatbot for anyone, it is not a fit for the platform. Read the WhatsApp Business Messaging Policy yourself before launching anything borderline, and keep the usual rules: opt-in for business-initiated messages, honour opt-outs, and a path to a human. For the platform basics, start with what the WhatsApp Business API is.

Agencies that set up AI agents and hand-off flows for clients can offer them under their own brand through the white-label partner program.