Most AI support bots fail for one boring reason: they answer when they should shut up.

Everyone worries about the model being wrong. That's not usually what sinks these projects. What sinks them is a bot that gives a confident, fluent, completely made-up answer to a customer — once. After that, nobody on the support team trusts it, and it quietly gets switched off.

The fix isn't a better model. It's two things most builds skip:

The bot answers only from your documents. Not from what it "knows." If the answer isn't in your docs, there is no answer to give.

The bot is allowed to say "I don't know." Below a confidence threshold it stops and hands the conversation to a human, with the full transcript so the person isn't starting cold.

That second one feels like a downgrade. It isn't. A bot that resolves 60% of queries and escalates the rest cleanly is worth far more than one that attempts 100% and gets 15% of them wrong.

Here's the shape of it — deliberately simple, because the architecture isn't the hard part. Deciding when not to answer is.

If you've shipped one of these: where did you set your confidence threshold, and did you have to move it after launch?