Designing AI features users trust: UX patterns for copilots and agents
AI features earn trust when the AI suggests and the person decides. Show where an answer came from, make every draft editable, let users undo, state the limits clearly and keep a human approval step for anything that moves money, data or customers.
Key takeaways
- Suggest, don't act — keep approval with the user for consequential actions.
- Show sources and confidence so users can check the AI's work.
- Make outputs editable drafts, not final answers.
- Design the failure state: what happens when the AI is wrong.
Adding a chat box is not an AI strategy. The products that get value from AI place it inside the workflow — drafting the reply, matching the invoice, planning the trip — and let the person stay in control.
Five patterns that work
- Suggested reply / draft
- The AI writes, the person edits and sends. Our support-desk concept uses this so agents answer faster without losing the human tone.
- Swap, don't regenerate
- Let users change one part of an AI output without losing the rest — like swapping one stop in a generated trip itinerary.
- Show your working
- Cite the record, document or rule behind each suggestion so users can verify it in one click.
- Confidence and limits
- Say when the AI is unsure, and say what it cannot do, instead of guessing.
- Approval gates
- Anything that pays, deletes or contacts customers waits for a human click.
Before shipping an AI feature
- Define the job the AI does in one sentence.
- Design the wrong-answer state as carefully as the right one.
- Log what the AI did so actions can be audited.
- Test with real users who distrust AI, not only enthusiasts.
You can see these patterns in our AI trip planner, support desk and automation builder concepts in the UI/UX portfolio, and we build them as part of UI/UX design and AI automations.
Sources
Kiara writes Advorize's guides on design, development and digital systems, turning the team's project work into practical, plain-language answers.