How do you stop a support AI from making things up?

A language model writes the most likely next words. Ask it about your return policy when it has never seen your return policy, and it will write one. A plausible one. Usually a generous one.

So most hallucinations in support aren't mysterious. The bot was asked something it had no facts for, and nothing told it to stop. A generic bot with a friendly prompt and no access to your documents is set up to fail in exactly this way.

None of these is exotic. Together they turn a bot that improvises into one that looks things up and knows when to stop.

Collect a few hundred real questions from your inbox or chat history, including the awkward ones: angry customers, messages that are half English and half Arabic, questions about products you stopped selling. Write down what a good answer looks like for each.

Run the agent against that set before launch and again after every change to its content or rules. It's a boring step. It also catches most of the embarrassing answers before a customer ever sees them.

It will, now and then. Treat each wrong answer as a bug report about your content: find the document that was missing or unclear, fix it, and add the question to your test set so it stays fixed.

Then tell the customer the right answer. People forgive a mistake that gets corrected quickly far more easily than one that gets defended.