AI UX Patterns That Don't Suck
After hundreds of AI features, these are the interaction patterns that consistently work.
Most "AI features" in shipped products are bad UX dressed up as innovation. After designing and shipping AI into dozens of client products, these are the interaction patterns that consistently work — and the anti-patterns that consistently flop.
Five patterns that work
- Progressive disclosure. Show a draft, let the user accept, edit, or regenerate. Never block the user on a single LLM completion.
- Confidence indicators. "I'm pretty sure" vs "you might want to double-check." Users need to know when to trust and when to verify.
- Inline editability. The AI output goes straight into the editable field. Don't force a copy-paste step.
- Explicit "AI generated" badges. Counterintuitively, transparency builds trust. Users who know it's AI assume some imperfection and don't punish small errors.
- Undo always available. Even non-destructive AI suggestions benefit from a one-click revert.
Five patterns that fail
- Chat-only interfaces for non-conversational tasks. "Type your request" is bad UX for "schedule a meeting" — give a form.
- Infinite generation loops. "Regenerate" buttons that produce slight variations indefinitely. Cap at 3, offer "edit instead."
- AI that interrupts the user's flow. Modal popups suggesting things mid-task. Save it for a sidebar or dismissable toast.
- "Magic" buttons that do too much in one click. "Make this better" hides intent and produces unpredictable output.
- Hiding latency without explanation. Spinners with no progress indication. Even fake progress beats a blank wait.
The latency tax
Anything above 2 seconds needs streaming, skeleton states, or both. Anything above 8 seconds needs an explicit "this will take a moment" with a way to do something else. Above 30 seconds, queue it and notify when done.
The trust ladder
Users move through a predictable trust curve: skeptical → cautiously curious → habitual → trusting. Each stage needs different UX. New users need confidence indicators and undo. Power users need keyboard shortcuts and batch actions. Build for both, not one or the other.
Bottom line
Good AI UX is mostly normal UX with the right metaphors layered in: drafts not finals, confidence not certainty, editability not finality. If you wouldn't accept the pattern in a non-AI feature, the AI label doesn't fix it.