A confident wrong answer costs more than a hedged right one. Yet most assistants are built to sound certain, because certainty reads as competence in a demo.
In production the calculus inverts. The people relying on an assistant need to know when to double-check it, and they need to know it at a glance — not by reading a confidence score buried in a tooltip.
We treat doubt as a first-class design material: citations that sit beside the claim, language that softens when retrieval is thin, and a visible handover the moment the question leaves what the system actually knows.