AI Product UX Readiness Check
This check helps you see whether the product around your AI feature has been designed, or whether only the model has been chosen.
How to use it
Use the check before committing engineering time, and again before release. Work through it for one AI feature or use case with the people who would have to defend the decisions.
Answer for one AI interaction, not for your AI strategy. Write each answer down, because "we would probably..." is not an answer.
Score each question as decided, assumed, or open.
- Decided: written down, and someone owns it.
- Assumed: everyone thinks they agree.
- Open.
Purpose
- 01
What can the user do afterwards that they could not do before?
If the answer is a capability of the model rather than a change in someone's work, the feature has no product case yet.
- 02
What would this person do today without the AI?
You need this to know whether the AI is faster, better or merely present.
- 03
What does a good outcome look like, in the user's terms?
Not model accuracy. What the person was trying to get done.
Role and limits
- 01
What is the AI allowed to do on its own, and what needs a person?
The useful ladder is: show information, suggest something, prepare something for approval, act and report, act silently. Pick one deliberately.
- 02
Who is accountable when it acts?
If the answer is the AI, the answer is wrong. Name a role.
- 03
What is it explicitly not for?
A boundary tends to establish itself one way or another. Writing it down beforehand is preferable to discovering it in a complaint.
Basis and evidence
- 01
What is the output based on, and can the user see that?
Sources, records, documents, prior behaviour. If the user cannot inspect the basis, they cannot judge the output.
- 02
How does the user tell a confident answer from a shaky one?
If everything arrives with the same tone, users will either trust all of it or none of it. Both are failures.
Failure and recovery
- 01
What does being wrong look like here, specifically?
Not "hallucination". The actual bad outcome: the wrong customer emailed, the wrong figure in the report, the wrong record closed.
- 02
How would anyone notice?
Silent wrongness is the expensive kind. If the answer is "the customer would tell us", say so out loud and see how it sounds.
- 03
What can the user do about it?
Correct, override, undo, escalate, or nothing. If nothing, that is a product decision and it needs owning.
- 04
If a wrong output has already spread, what happens?
Something was sent, saved, approved or acted on. Containment is a design problem, not an incident-response problem.
Reading your results
Use the cluster of open answers to decide what the result means and the reasonable next move.
| Where the open answers cluster | What it means | Reasonable next move | |
|---|---|---|---|
| Purpose (1 to 3) | You have a capability, not a feature. | Settle the user problem before building. | Settle the user problem before building. |
| Role and limits (4 to 6) | The feature will work in the demo and cause arguments in production. | Decide the level of autonomy and who is accountable. | Decide the level of autonomy and who is accountable. |
| Basis (7 to 8) | Users will either over-trust or ignore it. | Design what can be inspected before designing the output. | Design what can be inspected before designing the output. |
| Failure and recovery (9 to 12) | In Tcules' experience the section teams most often leave open. | Design the wrong-answer path with the same care as the right one. | Design the wrong-answer path with the same care as the right one. |
What this does not tell you
It does not evaluate your model, measure accuracy, assess your data, or cover security, privacy or regulatory obligations. It covers the product interaction around the AI.
What to do next
Related tools, services and audits
Use these paths to take parts of the check further.
For a practitioner's view on one use case
The AI Product UX Readiness Assessment begins with a free audit.