Designing AI Experiences

The best model can still make a bad product if the experience is poor. People need to know what the AI is doing, when to trust it, how to fix it and how to reach a person. This lesson covers design patterns for AI features that build trust and make mistakes easy to recover from.

Set expectations

• Say what the feature is for and what it is not for.
• Tell people it is AI, in simple words, where it matters.
• Show an example of a good request, which teaches users how to ask.
• Do not overpromise. Words such as always accurate or fully automatic break trust when the first error appears.

Show your work

• Show sources for answers drawn from documents, with links that open the right part.
• Show what the AI did: steps taken, tools used, data read.
• Show what it changed, and let users compare before and after.
• Use plain labels such as drafted by AI, so users know to check.

Handle doubt honestly

Good products make doubt visible. When the AI is unsure, it says so, and offers a next step: ask a clarifying question, show likely options, or hand over to a person. Avoid fake precision, such as 87 percent sure, when the number is not reliable. Plain words work better: I'm not sure about this, please check.

Make correction easy

• Let users edit AI output directly.
• Offer undo for every action the AI takes.
• Let users regenerate, refine with a short instruction or choose from options.
• For actions with consequences, show a preview and ask to confirm.

Collect feedback in the flow

• A simple thumbs up or down, with an optional reason.
• Learn from edits: what did the user change?
• Let users report a problem in one click.
• Thank them, and tell them when you fix things.

Design for waiting

Models take time. Streaming text as it arrives makes waiting feel shorter. For long tasks, show progress and what is happening. Let users do something else and notify them when it is ready. Never leave a blank screen.

Design the handoff to people

Some cases need a human. Make the path clear and quick. Pass the whole context to the person, so the user does not repeat themselves. Tell the user what to expect, such as a reply within an hour.

[ ] Purpose and limits are clear in the first view
[ ] The user knows it is AI where it matters
[ ] Sources and steps are visible
[ ] Doubt is shown in plain words with a next step
[ ] Edit, undo, regenerate and confirm are available
[ ] Feedback takes one click
[ ] Waiting has streaming or progress
[ ] Errors explain what happened and how to recover
[ ] Handoff to a person is easy and carries context
[ ] Works with a keyboard and a screen reader; text is readable
[ ] Works in the languages and on the devices your users use

Avoid common mistakes

• Hiding that content is AI-made.
• A chat box for everything, when a button or a form would be clearer.
• No way to fix mistakes.
• Too many options, or too few clues about what to ask.
• Blaming the user for a bad answer.