FreeScore your product against our 48-point UX audit checklist48-point UX audit checklist
25 checks8 groupsAbout 10 minutes

Check whether people can understand and trust your AI product

Most AI features fail on product design rather than model quality: unclear expectations, no way to verify, no control over what an agent does. Answer each check and see where to start.

What makes an AI product easy to trust?

People trust an AI product when they understand what it can do, can see when it might be wrong, can check its sources, stay in control of consequential actions, and can recover easily when it fails.

Answer the checks for the AI features in your product as they work today. The groups follow the patterns on our AI product design page: suggest and accept, draft then edit, ask to clarify, streaming with control, sources and grounding, and scoped automation.

The checklist: 25 checks

Answer each check for your product as it is today. Your score and a short start-here list update as you go.

1. Setting expectations

Does the product say clearly what the AI can and cannot do?

Do new users see examples of good requests instead of a blank prompt box?

Is it clear which content was produced by AI?

2. Input

Can users give context (files, selections, settings) without writing long prompts?

Does the system ask a clarifying question when a request is ambiguous?

3. Output and uncertainty

Does the interface signal when the system is unsure or might be wrong?

Are sources or reasoning shown where users need to verify an answer?

Can users edit, refine or regenerate a result without starting over?

Do slow responses stream or show progress, with a way to stop?

4. Control and approval

Is the level of autonomy matched to the risk of each action?

Do consequential actions (sending, paying, deleting, publishing) need explicit approval?

Can users see what an agent did, step by step, after the fact?

Can users undo or roll back an agent's actions?

Are an agent's permissions visible and adjustable?

5. Errors and recovery

When the AI fails, does the user get a useful next step rather than a generic error?

Is there a non-AI path to finish the task?

Can users reach a person when the AI cannot help?

6. Trust, privacy and data

Do users know what data the AI uses, and whether their inputs are used for training?

Are sensitive inputs handled with clear consent and controls?

Are outputs checked for harmful, biased or off-brand content before users see them?

7. Feedback

Can users flag a bad result in one step, with the context captured?

Is that feedback reviewed and used to improve the product?

8. Measurement

Do you measure whether AI features are used repeatedly rather than tried once?

Do you track how often outputs are edited, rejected or regenerated?

Do you know where users over-trust or under-trust the output?

Answer the questions above to see your score.

Frequently asked questions

How should an AI agent ask for approval?

Match approval to risk: low-risk, reversible actions can run with a visible log and undo, while consequential actions such as sending, paying, deleting or publishing should show what will happen and wait for an explicit yes.

How is the score calculated?

Yes counts as 2 points, Partly as 1 and No as 0. N/A answers are left out. The score is the share of available points you scored, so it stays fair if some checks do not apply to your product.

Is my data stored?

Your individual answers are saved only in your own browser, so you can come back to them. If you have accepted analytics cookies, Designit records that the checklist was started and completed and the overall score, never your answers.

Tell us what your AI should do, and what happens when it's wrong

We'll show you how we would design for trust, control and recovery in your product.

Discuss your AI productAI product design

Last updated 5 October 2026