How to stop a WhatsApp chatbot from hallucinating: official knowledge base and human handoff

Putting AI on your company’s WhatsApp is no longer hard. The hard part is making sure it doesn’t make things up. A customer who gets the wrong delivery time, a price that doesn’t exist or a return promise nobody approved won’t blame “the AI”, they’ll blame your brand. This guide covers the practices that keep a WhatsApp chatbot true to the facts, without losing the natural conversation that makes AI worth it.

What a hallucination is, and why it happens

Language models generate the most plausible answer to a question. When the right information is available, the plausible answer is usually the correct one. When it isn’t, the model doesn’t stop to say “I don’t know”: it produces something that looks right. That convincing, wrong text is a hallucination.

In customer service, the usual suspects are predictable:

• prices, fees and discounts;
• delivery times and opening hours;
• return, cancellation and refund policies;
• availability of products, slots or locations.

Those are exactly the answers where a mistake turns into a complaint, rework or a loss.

1. An official knowledge base the business controls

The first rule is to take the responsibility for your business’s facts away from the model’s memory. Everything the bot is allowed to state lives in an official knowledge base: locations, products, published prices, policies, hours. The assistant looks things up there to answer, and that’s what your team updates when something changes.

A good base is short and to the point. It doesn’t need to be a manual; it needs to be the source of truth for the questions customers actually ask.

2. “I’ll check” instead of guessing

The second rule is about behavior: if the answer isn’t in the base, the bot doesn’t answer on its own. It tells the customer it will check and hands the conversation to the team. It sounds small, but it’s the difference between a trustworthy assistant and one that is “almost always right”.

This behavior has to be explicit in the assistant’s instructions and, above all, tested (see point 5).

3. Handoff to a person, with history

Some conversations shouldn’t stay with a bot even when it knows the answer: complaints, payments, sensitive cases, unusual requests, or a customer who simply asks for a human. In those cases the chatbot should:

• notify the team immediately, on the channel they already use;
• pause the bot for that customer, so it doesn’t talk over the agent;
• hand over the full history, so nobody has to ask everything again;
• allow handing the conversation back to the bot once the case is solved.

A good handoff is what lets you use AI without fear: the bot handles the routine, and people handle what needs judgment.

4. Actions only with the customer’s confirmation

When the chatbot does more than answer, like booking a visit, reserving a slot or placing an order, the risk changes: a mistake stops being a sentence and becomes a commitment. The rule is simple: summarize and ask for an explicit “yes” before acting. “So I’ll confirm: visit tomorrow at 7am, at the downtown location?” prevents most misunderstandings.

5. Testing with evaluation conversations

The last practice is what separates a serious project from an experiment: a set of evaluation conversations that represent what customers really ask, including questions whose answer is not in the base. Before every change, whether to the base, the instructions or the model, those conversations run against the real assistant and the results are checked.

Good evaluation cases include:

• frequent questions answered by the base;
• similar questions with no answer in the base (the bot should call the team);
• requests that require handoff, such as a complaint or a payment;
• action flows, like a booking with and without confirmation.

Monitor it live

Even with all of that, visibility matters. A page that shows conversations in real time, and lets the team step into any of them with one click, turns the chatbot into a tool for your team instead of a black box.

Summary

• Facts come from an official knowledge base the business controls.
• With no information in the base, the bot says it will check and calls the team.
• Sensitive cases go to a person, with history, and the bot pauses.
• Actions only happen after the customer’s explicit “yes”.
• Every change runs through evaluation conversations before reaching customers.

That’s how a WhatsApp chatbot can be both human in tone and true to the facts.