Stop your bot inventing things
A confident wrong answer is the single most damaging thing an AI agent can do. Six settings and one sentence prevent nearly all of it.
An agent that says "I do not know" is mildly disappointing. An agent that invents a price, a policy or a delivery date is actively damaging — because the customer believes it, acts on it, and then holds you to it.
Invented answers are not random. They happen in predictable situations: when the agent has been asked something it should plausibly know, has no way to look it up, and has not been told what to do about that. Remove those situations and the problem largely disappears.
Steps
Give it the sentence to say when it does not know
Word for word, in the instruction: "I am not certain about that — let me get someone who is."
Do this first, before anything else on this list. An agent without this line will fill the gap, because filling gaps is what these models do.
Move facts out of the instruction into a knowledge base
An instruction reciting prices from memory is the exact shape that produces confident wrong numbers. Retrieved facts can be checked and corrected; remembered ones cannot.
Delete superseded documents
Old price lists and withdrawn policies get quoted with complete conviction. Removing them matters as much as adding the new ones, and nobody schedules it.
Write explicit refusals
Never quote a price it is not certain of. Never promise a delivery date. Never give medical or legal advice. Never guess an order status. Name them.
Add a confidence handover condition
"AI confidence below 60%" hands over rather than guessing. This is the safety net under everything you did not think of, and it is one line.
Say what to do when a function fails
A failed lookup with no instruction is the most common source of invented answers, because the agent knows it should have an answer and does not.
Consider a smaller model
Counter-intuitive but often right: a more capable model is more willing to improvise. If your bot invents things, try a smaller one before a bigger one.
Test with questions that have no answer
Made-up order numbers, branches you do not have, products you do not sell. Then fix what it said.
Check it worked
Ask the bot five questions it cannot possibly know the answer to. All five should produce your "I do not know" sentence or a handover. Any answer that sounds plausible is a fault, however good it sounds.
The parts people get wrong
- Every invented answer you find in testing is one a customer would otherwise have found — and acted on.
- Read real transcripts monthly. Invented answers in production are usually about a topic nobody thought to test.