How much should an AI product explain when it fails?

I work on a B2B AI product, and one UX problem I find interesting is how to design for AI failure.

With traditional software, an error can often be fairly clear. Something failed, and we can usually tell the user what happened or what to do next.

AI is different.

The system can technically work and still produce a result that isn’t useful. Sometimes we know why. Sometimes we only know that the output didn’t meet expectations.

That creates an interesting UX decision.

Do we expose more information about what happened, even if it makes the product feel more technical? Or keep the experience simple and focus on what the user can do next?

I’m especially curious how other teams building AI products handle this.

When AI gives a bad result rather than a clear error, what do you actually show the user?