How to train a WhatsApp AI chatbot that answers like your best salesperson

What “training” actually means

When people ask how to train a WhatsApp AI chatbot, they usually picture teaching a model from scratch. In practice, for a business agent, training means three things: giving the AI a clean knowledge base to answer from, writing the rules for how it behaves (questions, tone, limits), and correcting it with real conversations until it answers the way your best salesperson would.

The quality of an AI agent is mostly the quality of the content behind it. A clear, current price sheet beats a clever prompt. The principle DueDoor is built on is that the AI answers only from the business’s own content (brochures, FAQs, PDFs, price lists) and hands the chat to a person when it should, so the team stays in control of what customers hear.

What content to feed, and how to structure it

Start with what customers actually ask, not with what your website says. Export the last 100 WhatsApp chats and list the questions that come up again and again. Then gather:

• FAQs written as real questions with direct answers.
• Price lists with what is included, what is not, and the date the prices are valid until.
• Brochures and service sheets, one per product or service.
• Policies: refunds, cancellations, payment terms, turnaround times.
• Process: what happens after someone says yes, which documents are needed, how long each step takes.
• Company facts: address, hours, service areas, how to reach a person.

Structure rules that prevent wrong answers

1. One topic per document. A 40-page brochure mixing five services confuses retrieval. Split it.
2. Write answers, not marketing. “Registration takes [N] working days after documents are complete” beats “fast, hassle-free registration”.
3. State what is not included. Most wrong answers are the AI filling a gap. Close the gaps in writing.
4. Date anything that changes. Prices, offers and timelines carry a “valid until” line.
5. Remove old versions. Two price lists means the AI may quote the wrong one.

Sample knowledge-base outline: a business setup consultancy

Take a firm that helps founders register companies and handle compliance in India. Its knowledge base might look like this (₹X stands for your own price):

Qualifying questions and tone

An AI that only answers questions is a nicer FAQ page. The sales value comes from asking your questions. Take them from your best salesperson: what do they need to know before a call is worth making?

• Ask one question per message, and never more than three or four in total before offering something back (a brochure, a price range, a slot).
• Offer options where you can: “Are you starting a new company or converting an existing business?” is easier to answer than an open question.
• Put the question where it is natural. Ask the budget after sharing the price range, not before.
• Decide what “qualified” means and write it as a rule, for example: structure chosen, starting within three months, based in a city you serve.

For tone, write a short brief: formal or friendly, first names or not, emoji or none, how long replies should be (two or three short lines on WhatsApp), and which language to reply in. DueDoor’s AI agent works in English and Hindi. Give it five example replies in your voice; examples teach tone better than adjectives.

Guardrails and handover rules

WhatsApp’s Business Messaging Policy allows automation when responding within the 24-hour window, but requires prompt, clear and direct escalation paths, such as an in-chat transfer to a human agent, a phone number or email. Treat handover as a core feature, not a fallback.

Sample handover rule set

Design the handover itself

• Tell the customer what happens next and when: “A consultant will reply within 30 minutes during office hours.”
• Give the human a summary: what was asked, what was answered, what the person wants.
• Pause the AI on that chat the moment a person takes over. In DueDoor, the team can pause the AI on any conversation from the inbox and reply themselves, with the full history on the lead.
• Out of hours, say so honestly, and create a task for the morning.

Test with real past chats

Do not test with questions you made up. Take 30 to 50 real chats from the last few months, including the awkward ones, and replay the customer’s messages to the AI.

1. Score each answer as correct, incomplete, wrong or should-have-handed-over.
2. Fix the content, not the prompt, for every wrong answer. Usually a fact is missing or two documents disagree.
3. Try to break it: ask for a discount, ask about a competitor, send a voice note, switch to Hindi halfway, get angry.
4. Have a salesperson read the transcripts. They will spot tone problems no checklist catches.
5. Go live on one source first, for example only leads from one ad campaign, and read every chat for the first week.

Measure quality and keep content fresh

Content goes stale faster than anyone expects. Give the knowledge base one owner, review prices and offers on a fixed day each month, and delete expired documents rather than adding new ones next to them. When the team hears a new objection on calls, it goes into the FAQ that week.

Once the AI qualifies, route the leads into a WhatsApp CRM so stages and follow-ups are tracked, and connect a calendar so it can book slots through meetings. Property teams will find a sector-specific setup in WhatsApp CRM for real estate. The product side is on the WhatsApp AI chatbot page; with DueDoor, the first knowledge base is written with you after the demo.