Your AI Chatbot Quoted the Wrong Price. How to Stop It Costing You Customers

Switch on an AI agent to answer customer messages and one fear shows up before any other. Not whether it sounds robotic, or whether people can tell it's a bot. The real fear is simpler and more expensive: that it tells a customer something that isn't true.

Picture a dental clinic. At 11pm a patient asks the clinic's WhatsApp how much scaling and polishing costs. The bot, working from a price list nobody updated after January, answers confidently with last year's figure. The patient books, turns up on Saturday, and is told at the counter that the price is $15 higher. Now the clinic has a choice between honouring a price it never set and arguing with a patient holding a screenshot. Either way, the fast reply that was supposed to win the booking has cost goodwill.

That's the problem. A confident, well-written wrong answer is worse than no answer, because the customer believes it. This article covers why AI agents give wrong answers and what you configure so yours doesn't.

Why AI chatbots make up answers

A general language model is trained to produce a reply that sounds right, not one that is right. When it doesn't know something, it doesn't stop and say so. It fills the gap with the most likely-looking words. The industry calls this a hallucination, and in a customer chat it looks like your bot inventing a delivery time, a refund policy or a price.

This matters because customers act on what the bot says. Pricing and policy are where wrong answers do the most damage, because those are the answers people screenshot and hold you to.

So the goal isn't a smarter model. Serious platforms mostly run on the same class of models. The goal is to limit what the model is allowed to answer from.

Free download

How to Write the Instructions for Your WhatsApp AI Agent So It Gives Correct Answers and Knows When to Ask a Human

A practical companion to this article. Free for your email address.

Get it free

What actually stops wrong answers

There's no single setting. It's a stack, and each layer closes a different gap.

• Grounding. The bot answers only from a knowledge base you control, not from its general training. This is the biggest lever by far, because the bot retrieves your facts instead of guessing.
• A sealed knowledge base. The bot can read the documents you gave it and nothing else. It can't wander into other data, and it can't answer on topics you never covered.
• Rules in the instructions. You tell it in plain language what it must never do: never invent a price, never confirm stock it can't see, never promise a discount, never say a treatment or course is suitable for a specific person.
• Handover triggers. When a question falls outside what the bot knows, or touches money, complaints or anything high-stakes, it passes the conversation to a person instead of improvising. "Let me get a colleague to confirm that for you" is a correct answer. A made-up figure isn't.
• Monitoring. Someone reads what the bot said, finds the gaps and fixes the source. Accuracy is maintained, not set once.

Miss one layer and the others leak. A perfect knowledge base with no handover rule still guesses at the question you forgot to cover.

The knowledge base is the product

Think of the knowledge base as the only book the bot is allowed to quote from. If a fact is in there, the bot can use it. If it isn't, a well-configured bot says it doesn't have that information and offers a person.

That means the quality of the answers is the quality of what you put in. Vague or missing source material is where wrong answers come from. The vendors describe it the same way: Wati's AI support agent answers "based on the information you upload", and respond.io trains its agent "only on verified files" (Wati and respond.io help documentation, as of September 2026). None of them can supply accuracy you didn't give them.

A few things to get right for any business:

• Prices in your currency, stated clearly. Say whether a price includes sales tax, whether it's a starting price, and which branch it applies to if prices differ between, say, your city-centre and suburban outlets.
• The languages your customers write in. If customers message in English, Spanish and Mandarin, test the bot in all three. A policy that's clear in English can come out vague in a translated reply. Where the wording matters, such as refunds, deposits or cancellation, give the bot an approved version in each language.
• One version of the truth. Delete old price lists and last year's promotion. When two documents disagree, the bot may pick the wrong one.
• Dates on anything temporary. A Merdeka promotion or a festive-season opening hours notice needs an end date, or the bot will keep offering it in November.

When your pricing, stock or policies change, the source has to change on the same day. Otherwise the bot keeps quoting yesterday's truth with total confidence. Review the knowledge base on a schedule as well, and treat every wrong answer as a gap to patch rather than a one-off.

When the bot should refuse and get a person

The teams who trust their bots are the ones who decided up front what the bot isn't allowed to handle. A sensible default set of handover triggers:

• Any custom price, discount, package or contract term the bot can't read from a live source.
• A complaint or an angry message, where tone matters more than facts.
• A question the knowledge base doesn't cover, where the honest answer is "let me check".
• High-stakes commitments: deadlines, medical or legal questions, anything that would be costly to walk back.

Writing these conditions well is a skill of its own. Phrase them too loosely and the bot hands over everything, which defeats the point. Too tightly and it answers things it shouldn't. Most teams underinvest here, and it's the part that decides whether the bot earns trust in week one or loses it.

Handover also needs somebody on the other end. If the bot passes the chat to "the team" at 11pm and nobody picks it up until Monday, you've swapped a wrong answer for no answer. Decide who receives handovers during office hours, and what the bot tells the customer after hours ("Our team will confirm this by 10am tomorrow").

Where a WhatsApp AI agent isn't the right answer

Some businesses shouldn't put an AI agent on the front line at all.

If your answers depend on judgement more than facts, such as a lawyer weighing a case or an accountant reading a specific tax position, a bot grounded in documents will be thin. Use it to triage and book time with the expert. If your business is built on a personal, consultative first impression, automating that first touch can cost you the thing that makes you different. And if your knowledge isn't written down anywhere, no platform can ground a bot in it. You have a documentation problem before you have an AI problem, and switching on a bot first just automates the guessing.

An AI agent is excellent at answering the same fifty questions correctly, at any hour, and knowing when to step aside. It can't replace expertise it was never given.

The quieter reason bots get things wrong

Sometimes the bot is wrong because it can't see enough. If your AI can read a saved PDF but can't see the current pipeline stage, the customer's last booking, or what a colleague promised them yesterday, it will answer from stale scraps and sound certain doing it.

So the honest question about any AI CRM isn't how clever the model is. It's how much of your business the AI can read from and act on. That decides whether its answers come from your real, current situation or from a document written three months ago.

Yalla's agent sits on the same system as your pipelines, contacts and conversation history, so its answers are grounded in what the business currently knows, and it hands over to the right person on your team when a question crosses a line you've set. Yalla Brain acts across pipelines, WhatsApp, Instagram, Facebook, email campaigns and reporting from one instruction. The accuracy comes from that connection to the rest of your records, far more than from the model.

Frequently asked questions

Can any AI chatbot be 100% accurate?

No, and any vendor promising that is overselling. What you can do is make wrong answers rare and safe: ground the bot in a maintained knowledge base, add rules against inventing prices or policies, and route anything uncertain to a person. A good setup never guesses on the things that would cost you.

Why does my bot give the wrong price when the right one is on my website?

Usually because the bot wasn't given that page as a source, the source is out of date, or two sources disagree and it picked the wrong one. Make sure the current price is in the knowledge base, remove old versions, and add a rule that any custom or non-standard price must be confirmed by a person.

How often should I update the knowledge base?

Every time your pricing, products, policies or processes change, plus a fixed review on top: monthly for most teams, weekly if you handle high volume or run frequent promotions. Treat every wrong answer a customer reports as a gap to close in the source.

Is a knowledge-base bot the same as ChatGPT?

They use similar underlying models. The difference is that a knowledge-base bot is restricted to answering from documents you control, while a general chatbot answers from its training and will fill gaps with confidence. For customer-facing work, only the restricted version is safe, because it can be told to say "I don't know" and hand over.