Shipping AI features people trust
Loading states, mistakes and undo: the patterns that stop AI from feeling like a black box.
AI features fail in a particular way: they work most of the time, and nobody can tell when they will not. The design job is to make that visible and cheap.
Show the work, not a spinner
A spinner says nothing about what is happening. Showing the steps, even briefly, tells people the system is doing something sensible and gives them a moment to stop it.
Make mistakes cheap
AI will get things wrong. We design so a wrong answer costs a click, not an afternoon: suggestions instead of changes, drafts instead of sent messages.
Always offer undo
Anything the AI does on someone's behalf gets an undo, right where it happened. People try more when trying is safe.
Say what it cannot do
A short, honest line about the limits does more for trust than a perfect demo. People forgive a tool that told them the truth.
