Skip to main content
Next Step PM
Back to the library

Design · Add-on

Design for uncertain AI output

AI features can give different answers to the same question, and sometimes wrong ones. Design so people know what to expect, can check results and can recover easily.

About 3-8 daysEvergreenLast reviewed 5 November 2026

Who this applies to

Skills it builds
Design and user experienceTechnical and AI fluency
PM types
AI
Product types
All
Starting points
All
Ways of working
All
Experience
All
Size of organisation
All

Steps, by experience

Beginner (under 1 year)

  1. Write down what the AI feature can and cannot do, in plain words.
  2. Show this to users before they rely on it.
  3. Make it easy to check, edit or undo what the AI produced.
  4. Give a clear way to report a bad result.
  5. Test the design with real examples of good and bad output.

Intermediate (1 to 4 years)

  1. Design for the common failure cases, such as wrong, missing or slow answers.
  2. Show confidence or sources where it helps people judge results.
  3. Use the human-AI interaction guidelines as a review checklist.

Experienced (5+ years)

  1. Set design standards for AI features across products.
  2. Balance automation with human control based on the risk of errors.
  3. Review failure reports regularly and feed them into design.

Who can do this

Owner: Product manager

Can be delegated to

  • UX Designer: Design the interaction, including failure states
  • Content Designer: Write how the feature explains itself
  • Data Scientist / ML Engineer: Provide examples of real failures

Reviewed by: UX Designer

Tool examples

Figma
Design every failure state next to the happy path.
Claude
Collect real examples of good and bad output to test with.

Definition of done

  • Limits explained to users
  • Check, edit and undo designed
  • Failure states tested with users

Common pitfalls

  • Designing only the happy path
  • Hiding limits
  • No way to report a bad result

If you lead the team

What good looks like: AI features are designed so people know when to trust the output and what to do when it's wrong.

Ask your PM

  • What happens when the AI is wrong?
  • How do we show users how confident to be?

Warning signs

  • The design assumes the AI is always right.
  • Users can't easily report or fix bad answers.

Do these first

Sources