Make AI useful, understandable, and reviewable

A model capability is not a product outcome. Start by naming the user problem and comparing an AI approach with simpler ways to solve it.

When AI is useful, make its role clear. Show what the system can and cannot do, offer a way to review important outputs, and provide a safe path when confidence is low.

Treat evaluation as ongoing product work. Test representative inputs, monitor failure patterns, and keep a human decision-maker involved where the consequences require it.