Designing AI experiences
Users trust AI products that are honest about what they can do and easy to correct. This module covers the design patterns that make that possible.
Learning objectives
- Set accurate user expectations
- Design for noticing and recovering from errors
- Choose the right level of human oversight
- Collect feedback that improves the product
3.1 Set the right expectations
Users form a mental model of what your AI can do. If it's too optimistic, they over-trust and get burned; too pessimistic, and they ignore a useful feature. Be clear about what the feature does, what it doesn't, and how sure it is.
- Explain capability in plain language at the point of use ("Drafts a reply from your order history — check before sending")
- Show the sources behind answers, so users can verify
- Signal uncertainty where it helps ("I couldn't find this in the policy")
- Label AI-generated content where people would want to know — and where regulation requires it
3.2 Design for when it's wrong
Every AI feature will make mistakes. Good design makes them cheap to notice and easy to fix:
- Make outputs editable rather than final
- Offer alternatives (several suggestions) when one guess is risky
- Provide an easy route to a human or to the non-AI path
- Confirm before irreversible actions — sending, paying, deleting
- Fail gracefully: "I'm not sure" beats a confident wrong answer
3.3 Human in the loop
| Level | Pattern | Suits |
|---|---|---|
| Suggest | AI proposes; human decides | High stakes, early trust |
| Act with approval | AI prepares the action; human confirms | Actions with consequences |
| Act and notify | AI acts; human can review or undo | Reversible, lower-stakes actions |
| Fully automated | AI acts; humans monitor samples and metrics | High volume, low stakes, proven quality |
3.4 Feedback that improves the product
Collect feedback that's useful, not just thumbs. Implicit signals — accepting, editing, regenerating or abandoning an output — are often more reliable than ratings. Ask for a reason when users reject an output. Close the loop: show users their feedback changed something.
3.5 Conversational interfaces aren't always best
A chat box is flexible but puts the burden on users to know what to ask. Often a button ("Summarise this thread"), inline suggestions or a pre-filled form works better. Choose the interface for the task, not the technology.
For your feature from Module 2, list the three most likely ways the AI will be wrong. For each, describe how the user will notice and how they'll recover.
See a worked example
Example — expense categorisation: (1) Wrong category → shown as an editable suggestion with the top three options. (2) Can't read receipt → 'We couldn't read this — please choose a category' with the normal form. (3) Confidently wrong on a new merchant → finance sees a weekly sample of AI-categorised claims, and corrections feed back into the model.
Key takeaways
- Calibrate trust: explain capabilities and show sources.
- Make errors cheap to notice and easy to fix.
- Match human oversight to stakes and evidence.
- Implicit signals are often better than ratings.
- Chat isn't always the best interface.