Show the evidence that matters
A long explanation is not always useful. Surface the few signals that help someone judge an output: data recency, source context, confidence, important assumptions, and known gaps.
Match automation to consequence
Low-risk, reversible actions can be fast and automatic. High-impact actions need review, comparison, and a clear record. The interface should make that difference obvious before the user commits.
Design for correction
Every AI experience needs a path for fixing wrong context, changing the request, or reversing an action. Correction should improve the immediate outcome and help users build a reliable mental model of the system.