Stopping an AI assistant inventing answers

A confidently wrong refund policy is more expensive than no chat at all. Here is what actually reduces it.

A general model asked about your returns window will answer. It has read a million returns policies and will produce a plausible one. That is the failure mode worth designing against — not rudeness, not latency.

Ground every answer in something you wrote

The assistant should be answering from your policy text and your product rows, and should have nothing else to draw on. If it cannot find the fact, it does not have the fact.

Make "I do not know" a good outcome

Most systems are tuned to always produce an answer. The safer default is to hand off. A customer passed to a human in ten seconds is a good experience; a customer given an invented delivery date is a refund and a review.

Test the risky questions, not the easy ones

Anyone can check that it answers "what do you sell". Check what it says about damaged goods, late deliveries, and anything involving money coming back out of your account.

Read what it actually said

Transcripts are the only honest measure. A tool that does not let you read every conversation is asking you to take its word for it.

Why "add a disclaimer" does not work

A line under the chat saying answers may be inaccurate does not protect anybody. Customers do not read it, and it does not change what they were told about your returns window. The fix has to be in what the assistant is able to say, not in a note beside it.

The questions worth checking every month