How to use AI to validate and test a design

“AI validated this design” is too ambiguous to be useful. It could mean a model liked a screenshot, a script checked some elements, or a research assistant summarized sessions with people. Name the method and its limits. A layered evaluation is more credible than one confident verdict.

Layer 1: inspect the design against its requirements

Provide the intended task, audience, product rules and screens. Ask AI to identify missing information, inconsistent terminology and unsupported assumptions. Require it to cite the relevant screen or text for each observation. Treat predictions about user confusion as hypotheses until you observe users.

For an appointment flow, check time zone, location, price, cancellation terms and the consequence of confirmation. Separate a visible missing label from an inferred concern about comprehension. This distinction makes a critique actionable and prevents speculative statements from acquiring false authority.

Layer 2: exercise the implementation

Use deterministic checks where possible: broken links, required field behavior, inconsistent tokens, known component states and selected accessibility issues. AI can help draft test cases, but the checks need to run against the actual implementation. A screenshot cannot reveal whether focus order, semantic labels or backend validation work.

Test refresh, repeated clicks, slow requests, empty data and recovery. Preserve useful work when an error occurs. Record expected and actual behavior. A passing automated check covers only the condition it evaluated.

Layer 3: observe people completing the task

Choose relevant participants and a neutral scenario. For the appointment flow, ask someone to find a suitable time while respecting a realistic constraint. Watch whether they understand the final commitment and can correct a mistake. Do not tell them which controls to use.

Capture task completion, assistance, errors and meaningful comments. Small qualitative studies can reveal problems; they do not automatically establish a population success rate. Report counts and context plainly.

Turn findings into a decision queue

For each issue, record evidence type, impact, confidence, affected audience and next action. A verified blocking defect should receive different treatment from an aesthetic preference. Review severity with the team rather than letting an opaque model score decide priority.

A useful record reads: “In session P03, the participant changed the date but assumed the time remained fixed; the confirmation showed a different time. Clarify the combined selection and retest.” It gives the next designer a concrete problem without inventing prevalence.

Verify the fix

Repeat the relevant checks after revision. If wording changed, review comprehension. If a component changed, check neighboring states. If the decision affects a high-consequence action, choose validation appropriate to that consequence. Keep a record of unresolved issues rather than treating the final review as a blanket approval.