You added a chatbot to your site to take load off your team. Instead, it quoted a customer a price you don’t offer, promised next-day delivery on a route that takes three, and answered — with full confidence — a question that has no answer at all. Every owner who has tried to automate conversations eventually asks why a chatbot gives wrong answers. And almost always, the root cause isn’t where they look for it.
People assume the bot is “bad” or “dumb.” In reality, the bot does exactly what it was set up to do. It doesn’t lie on purpose — it completes an answer from whatever it has. And if you never gave it a reliable source of truth about your business, it fills the gap from its training data, meaning the entire internet at once. That is the “hallucination”: an answer that looks right in form but contains none of your actual facts.
The Bot Doesn’t “Know” — It Completes
The core misunderstanding: the owner pictures the bot as a staff member who studied the price list and the rules. In fact, a modern language model is a machine for probable text continuation. It sees a question and generates the most plausible continuation. “Plausible” and “true” are not the same thing.
When a customer asks “how much is installation,” the bot doesn’t check your spreadsheet. It recalls how thousands of similar answers sound online and produces something in that style — with a number pulled from nowhere. Perfect form, total confidence, invented content. That is exactly why wrong answers are so dangerous: they don’t look like errors. The customer believes it, treats the number as yours, and you clean up the mess.
Hence the first conclusion. The question isn’t “how smart is the bot.” The question is where it gets facts about your business at the moment it answers.
The Root Is in the Data Source, Not the “Brains”
When someone comes to us with “the bot talks nonsense,” we almost always find the same thing: the bot has no connected, current source of truth. Either it’s missing entirely, or it’s outdated, or it exists but the bot isn’t required to stick to it.
Picture the difference between two setups. First: the bot answers “off the top of its head,” leaning on general knowledge. Second: before every answer, the bot looks into your current material — service descriptions, delivery terms, FAQs, catalog — and builds the reply only from what it found there. To the customer it’s the same chat window. Inside, it’s a chasm. In the first case the bot invents; in the second it quotes you.
This is why a chatbot gives wrong answers even when it was “supposedly trained on everything.” Texts loaded into it once go stale the moment you change a price or a term. If the source lives separately and the bot pulls facts from it on the fly, answers stay current without retraining. But if the facts were “baked in” long ago, you’re doomed to a gap between what the site says and what the bot says.
Why Invented Prices Are the Costliest Mistake
Of all hallucinations, the ones about money and commitments hurt most. An invented price, an invented deadline, an invented guarantee. Because that’s no longer just an inaccuracy — it’s a promise the customer remembered.
From there, two scenarios, both bad. Either your manager walks back what the bot said, and the customer feels cheated because “your own chat told me.” Or you’re forced to honor an invented commitment at a loss, to protect your reputation. Both are the direct result of letting the bot speak about things it doesn’t actually know.
Doing nothing costs more than it seems. One such conversation, screenshotted and posted in reviews or a community chat, damages trust more than a dozen correct answers build it. Automation that occasionally lies about prices isn’t a saving on staff — it’s a drain on reputation.
An Honest “I Don’t Know” Is a Sign of a Good Bot
The owner’s healthy instinct is “I want the bot to answer everything.” That’s the trap. A bot that answers everything is a bot that answers some questions with fabrications.
A properly configured assistant has clear boundaries. For questions within your source — a precise answer drawn from your data. For questions outside it — an honest “I can’t advise on that, let me pass you to a person,” or an offer to leave contact details. That isn’t weakness; it’s strength. A customer told “a human will confirm shortly” is far less annoyed than one who was confidently lied to.
Technically this rests on a few things: the bot works only with your source, is instructed not to invent beyond it, and deliberately hands off risky topics — money, deadlines, legal questions — to a human. That’s how wrong answers stop being a lottery and become a narrow, managed, predictable flow.
What to Actually Do About Your Bot
The sequence we walk clients through is simple in logic.
First, define the source of truth. One living, current set of facts about your business that someone owns. Not ten scattered documents, but something that updates when a price or a term changes.
Next, connect the bot to that source so it builds answers on it, not on “general knowledge.” And separately, set the boundaries: what it speaks to, and what it must hand to a person.
Then, test it on real customer questions — including the awkward and provocative ones — and watch where the bot starts inventing. Those spots reveal holes in the source, not a “bad temper” in the bot.
Finally, upkeep. Business changes; new services and questions appear. The source has to stay alive, or in six months the bot drifts from reality again. We’ve worked with systems like this since 2018 on our own infrastructure precisely because a good assistant isn’t a one-time setup — it’s a system that’s maintained.
If your bot already invents prices or answers off-topic, that’s a good place to start a conversation. Show us a few real dialogues where it slipped, and together we’ll see exactly where the hole in the data source is. Often the first brief already makes clear what needs fixing.