We tried to make our AI chatbot invent an answer. Here's what it did instead.

Stopping an AI chatbot from inventing answers means giving it a rule most chat tools skip: check what you actually know before you answer, and say so plainly when you don't. That's not a filter bolted on after the fact — it's the whole design of how Maya responds. Ask her something that isn't in the business's approved information, and she doesn't guess a plausible-sounding reply. She tells the visitor she isn't sure, offers the closest real help she can, and flags the question for the owner. This is a walkthrough of that loop, and why it's the reason a business can actually trust what Maya says.

Last updated 2026-07-16

Where most AI chat tools quietly go wrong

A language model is built to keep talking. Ask most AI chatbots something outside their knowledge and, rather than stopping, they'll pattern-match toward a confident-sounding answer — the wrong shipping cost, a return policy that doesn't exist, a price that was never set. Nobody programmed the bot to lie; it's just doing what it was trained to do, produce a plausible next sentence, with nothing in the way to stop it when the facts run out. The business finds out when a customer shows up expecting a discount that was never real.

We tried to make Maya do that — asked it questions with no real answer behind them, the kind that would tempt a normal chatbot into a confident guess. Maya didn't take the bait. It checked what it actually knew, came up short, and said so: offered the nearest real help it had, and quietly flagged the question for the owner instead of answering it. Calling that "a chatbot with guardrails" undersells it — a guardrail stops a car that's already swerving. This is closer to an employee who's been told, and actually believes, that "I'll check and get back to you" beats a wrong answer every time.

How the abstention loop actually works

It checks before it speaks — against the business's own approved information — every answer Maya gives is checked against the pages, catalogue, and business data it read from the actual site — not general knowledge about "businesses like this one." What that means for you: the answer a customer gets is grounded in your real business, not a statistically likely guess about businesses in general.

When it can't verify something, it says so — out loud, in plain language — there's no forced cheerful reply when Maya doesn't know. It tells the visitor plainly that it isn't sure, and offers the closest real help it can — a related answer, a way to reach a person, or an offer to find out. What that means for you: a visitor gets an honest "let me check" instead of a wrong answer they'll act on and later blame you for.

It flags the gap for you — instead of quietly failing — the unanswered question doesn't vanish; it lands in a list you can see, so you know exactly what Maya was asked and couldn't confirm. What that means for you: you're not guessing what your AI might be getting wrong — it tells you.

You teach it the answer in one sentence — not a prompt, not a rebuild — you read the flagged question and write the real answer the way you'd correct a new hire — "we don't do same-day delivery, but we do next-morning." Maya holds that going forward. What that means for you: fixing a gap takes a sentence, not a developer.

The same flags become a list of what customers actually wanted — when several visitors ask the same unanswered question, that's not just a gap to close — it's a pattern. "14 people this week asked about airport transfers" is market feedback your website is generating for free. What that means for you: you stop guessing what to add next; your own visitors are already telling you.

It holds the line by voice or text, the same way, every time — whether a visitor types the question or asks it out loud, the same check-then-abstain behaviour applies — abstention isn't a text-only safety net. What that means for you: the reliability doesn't depend on which channel a customer happened to use.

What happens when Maya doesn't know the answer?

This whole page IS the abstention answer, but to say it plainly: when a visitor asks Maya something that isn't in the business's approved information, Maya does not invent a plausible-sounding reply. It checks what it actually knows, and when it can't verify the answer, it says so — offers the visitor the nearest real help, and flags the question for the owner rather than guessing. We are not claiming this is a mathematical or physical impossibility; it's a designed behaviour, and we tested it by trying to get Maya to answer things it shouldn't.

What that means for you: a business gets an AI that gets more useful every week without ever being rewarded for confidently inventing things. It won't quote a price you don't charge or promise a service you don't offer — and every time it says "I'm not sure," you get a one-sentence chance to make it a little smarter, instead of finding out about the gap from an angry customer.