Designing AI Experiences (AI UX)
Design AI features people trust and use: interaction patterns (copilot, chat, agent, inline), setting expectations, explainability and citations, feedback and control, failure and handoff design, and testing AI UX.
Designing AI Experiences (AI UX)
Design AI features people trust and use: interaction patterns (copilot, chat, agent, inline), setting expectations, explainability and citations, feedback and control, failure and handoff design, and testing AI UX.
In one line
AI UX is designing products where outputs are probabilistic: helping users understand what AI can do, trust it the right amount, correct it easily, and recover when it's wrong.
The 6 classes
Principles of AI UX
Design for uncertainty and the right level of trust.
- Design for being wrong sometimes.
- Aim for calibrated trust.
- Always provide control and recovery.
Interaction patterns
Pick the right pattern for the job.
- Match pattern to task size and risk.
- Prefer in-workflow buttons over generic chat.
- Agents need approvals and progress views.
Trust, explainability and citations
Help users judge AI output quickly.
- Citations and highlights speed verification.
- Explain decisions in user language.
- Label AI content clearly.
Feedback, control and personalisation
Turn user actions into learning signals.
- Implicit signals are richer than thumbs.
- Give users meaningful controls.
- Feed signals into evals and improvements.
Failure, safety and handoff design
Design what happens when AI can't or shouldn't answer.
- Every AI feature needs a fallback path.
- Preview and confirm before actions.
- Stream and show progress for long tasks.
Testing AI UX and final project
Validate AI experiences with users and evals.
- Test with imperfect outputs.
- Measure trust calibration, not only satisfaction.
- Design and evals go together.
Interview questions
Key principles of AI UX?
Set expectations, design for uncertainty, transparency with sources, user control and recovery, graceful failure with handoff.
When not to use chat?
When a task is small and in a workflow; an inline button or suggestion is faster and clearer.
How do you build trust?
Citations, plain explanations, labels for AI content, consistent quality, and easy correction.
What feedback signals matter?
Accept/edit/regenerate/abandon rates and edit distance, plus explicit ratings with reasons.
How do you test AI UX?
Wizard-of-Oz, prototypes with real and bad outputs, task success and trust calibration metrics.
Design for an AI agent action?
Preview, confirmation, progress, undo window, and clear logs.
