What’s inside
Every AI pilot looks good on clean demo input. Production brings typos, PDFs, edge cases, slow calls and a tired user at 11 pm — and users judge the product by its worst moment. Trust is decided in the interface, and the fixes take weeks, not quarters.
- 01AI sounds sure when it isn’t
Show confidence and the source next to every answer; make “not sure” a designed state.
- 02No way for a human to step in
Preview before apply, edit the draft, undo any action, confirm anything risky.
- 03A blank box and a “magic” button
Start with examples and suggestions built from the user’s own data.
- 04Waiting with no explanation
Stream the answer, show the agent’s steps, keep partial results.
- 05Invisible to other AIs
Publish machine-readable data so assistants can find and cite you.
Plus: a 15-point checklist in five groups (confidence & sources, human control, first session, waiting & failure, visibility & learning) and a score that tells you whether you’re ready to scale. Examples come from products I designed and shipped: Diagnocat (AI diagnostics in 1,000+ clinics), VIZI, EngiBoard and Vigilo.