How do you catch AI hallucinations before real users find them?

Internal testing always feels safe because our team knows how to prompt the app without breaking it. But the second it hits real users they throw wild multi part scenarios at it and the model would rather guess an incorrect pricing rule than just say it has no clue. Manual testing can't keep up with that kind of edge case volume.

What automated checks or validation layers are you using to stop the model from making up answers when it gets stuck?