WhatsApp AI chatbots that answer from your own documents (and know when to hand off)

For most businesses in India, WhatsApp is the customer service channel. That makes it the obvious place for an AI assistant — and also the most embarrassing place for one to get things wrong. A chatbot that invents a price or a policy in your company’s name is worse than no chatbot at all.

Here is how I build WhatsApp assistants that are actually useful.

Start with what customers ask

Before any AI, export a month of WhatsApp chats (or ask the team) and list the questions that come up most. In almost every business, around 20 questions make up the bulk of messages: prices, timings, location, availability, documents needed, order status.

That list becomes the scope. Anything outside it is handed to a person.

Answer only from your own content

The assistant shouldn’t answer from the model’s general knowledge. It should answer from your content: FAQs, price lists, policies, product catalogues, service descriptions. This pattern is called retrieval-augmented generation (RAG):

1. Your documents are split into small passages and indexed.
2. When a question arrives, the most relevant passages are retrieved.
3. The model is instructed to answer only from those passages, and to say it doesn’t know otherwise.

This is what stops the bot from making things up. It also means updating the bot is as simple as updating the document.

Design the hand-off first

The most important feature of a business chatbot is knowing when to stop. I design hand-off rules before anything else:

• The customer asks for a person.
• The question is outside the known scope, or retrieval finds nothing relevant.
• The topic is sensitive: complaints, refunds, medical or legal advice, negotiation.
• The customer is a qualified lead ready to buy.

On hand-off, the assistant tells the customer someone will reply, and sends the team the conversation with a one-line summary — on WhatsApp, email or a dashboard.

Capture leads properly

Every conversation is a potential lead. The assistant can ask one or two qualifying questions naturally (“Is this for your business or personal use?”) and write the answers to a Google Sheet or CRM. No more scrolling back through chats to find who asked about pricing last week.

The stack I typically use

• WhatsApp Business Platform (Cloud API) through Meta or a provider — required for automated replies on a business number.
• n8n to orchestrate: receive messages, retrieve content, call the model, apply hand-off rules, log everything.
• An LLM chosen for the job — fast, inexpensive models for routine replies; stronger models only where needed.
• Google Sheets or a database for leads and conversation logs.

Log everything, improve weekly

Every conversation is logged with the question, the retrieved passages and the answer. Once a week, read the logs: which questions got handed off, which answers were weak, what new questions appeared. Then update the content. A chatbot that’s reviewed weekly for the first two months gets dramatically better; one that’s launched and forgotten slowly gets worse.

Respect the platform rules

WhatsApp has rules about business-initiated messages (templates), opt-in, and message windows. Build around them from the start — a ban on your business number is far more expensive than doing it properly.

Is it worth it?

If your team answers the same questions on WhatsApp all day, almost certainly. The goal isn’t replacing people; it’s giving them back the hours spent on repeat questions so they can handle the conversations that need a human.

See how this fits into a bigger workflow in the CA firm case study, or tell me what your customers ask most.