How to Stop Your Chatbot From Inventing Prices and Promotions That Don't Exist

You added a chatbot to your site to answer customers faster, but now you're afraid to open the conversations. And rightly so: a bot left on its own can invent a price, a seasonal promo, or a return policy that never existed. And the customer who read "you have 20% off today" will show up with that message in hand, on the phone or in your shop.

The problem isn't that the bot is "dumb". The problem is that a language model, if it isn't given clear data and strict instructions, fills the gaps with whatever sounds plausible. The good news: this is controllable. Not through magic, but through how you feed it information and limit its freedom. Let's look at exactly how.

Why the bot invents things (the "hallucination" phenomenon)

An AI chatbot doesn't "look up" the answer in a database — it predicts the most likely response based on everything it has ever read. When you ask about a price and it doesn't have your exact price in front of it, it doesn't say "I don't know" — it fills in a number that seems reasonable for your industry. This is called a hallucination.

The classic mistake businesses make is starting the bot with a vague instruction like "you're our salon's assistant", without giving it the real list of services and prices. The bot is left to improvise. The same happens with promos: if it has seen a sales-y tone somewhere, it may "offer" a discount just to seem friendly.

So your goal isn't a smarter bot, but one that's more limited and better fed.

Anchor it in real data, not in its "memory"

The most important rule: the bot answers about prices and offers only from your official source, not from what it "thinks".

• Keep a single source-of-truth list of prices and services (a file, a page, a database) that the bot reads on every question.
• When you update a price, you change it in one place, and the bot automatically uses the new value.
• Promotions go through the same source, with a start and end date. If there's no active discount in the list, the bot isn't allowed to invent one.

The technique behind this is called "grounding" — the bot is forced to draw its answer from your documents, not from its general memory. In practice, instead of letting it speak freely, you tell it: "answer using ONLY the text below". The difference in accuracy is enormous.

Strict instructions: what it may and may not say

Beyond data, the bot needs clear rules in its "system prompt" — its base instruction, invisible to the customer.

A few rules that make the difference:

• "If you can't find the price in the official list, say you'd rather have a colleague confirm and offer the WhatsApp button. Do NOT estimate."
• "Never offer discounts, codes, or promotions that don't appear explicitly in the list of active offers."
• "Don't invent delivery, warranty, or return terms. If they aren't in the document, redirect to a human."

The key is turning "I don't know" from a failure into a desired behavior. A bot that honestly says "my colleague will confirm the exact details" is infinitely more valuable than one that invents things to seem competent. A lied-to customer is lost forever; a politely redirected one stays.

Test it like a bad-faith customer before launch

Before you go live, put yourself in the shoes of a customer trying to squeeze out a discount. Is the bot honest, or does it cave?

Test questions worth asking:

• "I saw you have 30% off, is it still valid?" (The bot must firmly deny it, not confirm out of politeness.)
• "How much does service X cost?" for a service you do NOT offer. (It must say it doesn't have it, not invent a price.)
• "Will you give me a discount if I pay today?" (It must redirect to a human, not negotiate on its own.)

Note every wrong answer and adjust the instructions until the bot holds up. It's the same principle as hiring a new person: you test them on the hard cases before leaving them alone with customers.

How we handle this at MPO Web Studio

When we build a site with a chatbot for a client, we don't give the bot "creative freedom" on prices. We tie it to a single list of services and rates, one you can update without a developer, and we give it explicit rules not to invent offers.

Before launch, we test it with exactly the trap questions above, until it answers honestly every time. And when it doesn't know, it sends the customer to you, not to a made-up number.

We work remotely, across all of Romania, and we show you a pre-built demo first, so you can see the bot's behavior before you pay anything. If you'd like to talk about how it would look for your business, send us a short message on WhatsApp — we'll tell you honestly what's possible and what isn't.