A multi-step AI workflow can fail at several points: model response, tool call, or the handoff between them.
Make clear what completed, what didn’t, and whether retrying could repeat an action.
A modal can look polished and still be difficult to use with a keyboard.
When it opens, focus should move into it; when it closes, focus should return to the control that opened it.
A polished AI answer doesn’t prove the workflow worked.
Test the outcome: did the right action happen? What happened when information was missing or a tool failed?
Fluency is visible. Reliability needs testing.