Make the relationship clear
Microsoft's human-AI interaction research and HAX resources focus on how AI systems should behave during interactions, including failures. A useful product explains what the feature can help with and leaves people able to correct its behavior.
Do not make users discover the system's limits only after an important mistake. A writing assistant can say that it creates an editable draft from the supplied facts. An order assistant should distinguish reading an order from changing it. These explanations are more helpful than a generic AI-powered badge.
Technical foundation: Microsoft Research: Guidelines for Human-AI Interaction
Design states, not only a chat bubble
Specify the empty, editing, pending, partial, complete, failed, and cancelled states. Preserve the user's input across errors. Give each state a meaningful action, such as retrying a failed read, adjusting the request, or returning to a manual workflow.
Streaming needs careful presentation. Avoid moving focus or repeatedly shifting the layout as text arrives. Keep Stop and other important controls reachable. Screen-reader announcements should communicate meaningful status changes without reading every arriving token as a new alert. Test the interaction with assistive technology rather than assuming a visually smooth animation is accessible.
Put review where consequences begin
A generated draft and a sent message should look different. Before an action with external effects, show the actual recipient, content, and operation being approved. A button labeled Continue may be too vague when it sends an email or commits a change.
For generated code or copy, offer a clear comparison and a way to accept only the desired changes. Do not replace the user's work silently. A useful assistant reduces effort while preserving authorship and control; it should not make recovery depend on remembering the exact previous prompt.
Use uncertainty as interface information
If evidence is missing, explain what is missing and offer a next step. Avoid decorative confidence percentages unless they have a defensible, calibrated meaning. Source links, input assumptions, and explicit unresolved questions often communicate uncertainty more honestly.
Keep core actions available without requiring natural-language input. Forms, filters, and ordinary navigation can be faster and more discoverable for repeated tasks. Support keyboard access, visible focus, readable contrast, and reduced-motion preferences. These are part of the product contract, not optional polish after the AI feature ships.
Review the complete interaction
- Explain capability and limits before consequential use.
- Preserve input and provide cancellation and manual alternatives.
- Keep focus stable and status announcements meaningful.
- Show exact effects before asking for approval.
- Test correction, failure, keyboard use, and assistive technology.
