Stop 12 · AI & Modern Product ManagementTopics · 3/7

Designing AI Features (UX for Uncertainty)

In short

Normal buttons do the same thing every time. AI features don't. Good design tells users how sure the AI is and makes fixing its mistakes painless.

Unlike deterministic software, AI features can be wrong — good design sets expectations, shows confidence or sourcing where possible, and makes it easy for users to verify or correct output.

The purple blob in a doctor's coat points at an X-ray on a glowing light box.

Think of it like this

The doctor with the X-ray

A good doctor doesn't just say "it's broken". She holds up the X-ray, points at the line, says "this looks like a hairline crack" and suggests a scan to confirm. You see the evidence, her confidence and the next step. AI features should behave the same way.

Watch it happen

An AI reply drafter that plans for mistakes

Support agent

App

Model

Open ticket #812

1/5An agent opens a refund complaint. The AI feature works in the background.

The big ideas

What you need to know

Set expectations

Tell users it's AI and may be wrong, in plain words, right where they use it.

Show your sources

Linking to the document or data used lets users check the answer in seconds.

Easy to edit or undo

Suggest, don't force. Make accepting, changing or rejecting output a single click.

Collect feedback

Thumbs up, thumbs down or edits tell you where the feature fails in real use.

Save yourself the pain

Common mistakes

Oops

Letting AI take irreversible actions, like sending money or emails, with no review.

Do this instead

Keep a human confirmation step for anything costly or hard to undo.

Oops

Designing only for the perfect answer.

Do this instead

Design the wrong-answer and no-answer states too. Users will hit them.

Your mission

Redesign for uncertainty

~30 min

Quick check

An AI feature drafts emails to customers. Which design is safest?

Go deeper

1Critique an AI feature you've used for how (or whether) it handles being wrong.
2Learn common patterns: showing sources, confidence indicators, easy undo/regenerate.
3Sketch how you'd design an AI feature's error and 'low confidence' states.