Build an AI-assisted prototype that answers a real design question

In this guide

1. The useful starting point
2. Choose the uncertainty the prototype should expose
3. Build the smallest honest version
4. A fictional example: reviewing an AI-generated project brief
5. Use AI to prepare the prototype and the test
6. Run the session with room for surprise
7. Inspect the complete experience
8. Decide what to change next
9. Questions you might be asking
10. Sources and further reading

The useful starting point

Start with an uncertain design question, build only the interaction needed to investigate it, and test with relevant people using realistic tasks. Label simulated behavior, observe without coaching, and separate usability evidence from demand or business validation.

AI makes it easier to create a convincing prototype. It does not tell you whether you built the right thing. The most useful prototype has a question behind it and a plan for learning from someone's actual behavior.

Before generating screens, finish this sentence: “We need to learn whether people can…” That may be choosing an appropriate option, understanding a recommendation, correcting a mistake, or completing a first useful task.

Choose the uncertainty the prototype should expose

A prototype can explore comprehension, sequence, information hierarchy, or interaction. It can also help someone react to a proposed service. Different questions require different evidence.

Watching someone complete a task can reveal usability problems. Hearing that they like the idea does not establish willingness to pay. A clickable flow does not prove that the underlying integration is feasible. Keep these questions separate so a successful test does not become a larger claim than the evidence supports.

Write the decision you will make after the session. If no possible observation would change the design, reconsider why you are testing it.

Build the smallest honest version

Realistic content matters. Placeholder text can hide the very confusion you need to observe. Use fictional but plausible records when real data is unnecessary, and clearly identify them as examples.

A fictional example: reviewing an AI-generated project brief

Imagine a design team building a tool that drafts project briefs from meeting notes. Their question is whether a delivery lead can identify unsupported commitments before approving the brief.

They create a prototype with a sample brief, linked source notes, an “unconfirmed” label, and edit controls. One sample contains a proposed deadline that was never agreed. The team is not testing whether the model can generate the brief yet; it is testing whether the review experience makes that uncertainty visible.

During a session, a participant approves the brief without opening the notes. The team asks what they believed the confirmation label meant and observes how they inspect a second example. They avoid telling the participant where to look before the task is complete.

The finding could lead to clearer source references or a different approval step. It does not justify saying that all users will miss unsupported commitments, or that the model itself has been evaluated.

Use AI to prepare the prototype and the test

Help me prototype the interaction needed to answer this design question: [question]. Use these audience needs, constraints, and known evidence. Identify the minimum screens and states, including an error or uncertainty case. Create realistic fictional content and label assumptions. Draft neutral tasks for real participants and an observation checklist. Do not invent user findings or imply that simulated behavior is functional.

Ask the assistant to identify what the prototype cannot tell you. That list helps prevent a narrow usability exercise from becoming a broad product validation claim.

Run the session with room for surprise

1. Choose relevant participants. Recruit people whose situation fits the question you are investigating.
2. Explain the prototype. State that you are testing the design, and clarify any simulated behavior without teaching the solution.
3. Give a realistic task. Describe the situation and goal rather than naming the control to click.
4. Observe first. Record actions, pauses, errors, and recovery attempts.
5. Ask neutral follow-ups. Explore what the person expected and why they chose an action.
6. Separate evidence and interpretation. Keep what happened distinct from your proposed explanation.

Do not quietly repair the flow during the task to make the prototype appear successful. If a technical failure blocks the session, record it and explain how it affects what you can learn.

Inspect the complete experience

Include the moments that polished demos often omit: missing data, an invalid input, a slow response, an unavailable action, and a person changing their mind. For AI features, include an incorrect or uncertain suggestion and a way to correct it.

Check accessibility in the actual prototype or implementation where possible. Keyboard navigation, focus order, labels, and readable content affect whether the task can be completed. AI-generated code or design output still needs inspection.

Review the results against the original question. A participant's difficulty may come from unfamiliar terminology rather than the overall sequence. A success may depend on prior expertise. Preserve that context.

Decide what to change next

Group observations by the decision they affect. Fix obvious implementation defects, investigate ambiguous behavior, and test important revisions. Keep a record of what changed and why.

If the prototype exposes a larger problem, reduce the scope and revisit the brief. If users complete the task but show no interest in using the service, investigate demand separately. If the interface works only with constant guidance, redesign the experience instead of counting coached completion as success.

AI can make iteration faster. Use that speed to create a better next test, not to skip the uncertainty. The prototype has done its job when it gives the team evidence that changes or strengthens a decision.

Questions you might be asking

Does a successful prototype test validate the business idea?

Not by itself. Task completion provides usability evidence. Demand, willingness to pay, delivery feasibility, and economics require additional investigation.

Can I test with AI-generated users first?

You can use simulations to rehearse a script or identify possibilities, but label them as synthetic. Important conclusions about people need real participants and observed behavior.

How polished should the prototype be?

Polish the parts needed for the question. Use realistic content and enough interaction fidelity to avoid misleading participants, while keeping unneeded implementation out of scope.

Sources and further reading

• Microsoft human-AI interaction guidelines
• Anthropic evaluation guidance
• Microsoft appropriate reliance research

Sources checked October 16, 2026. Review the linked publisher information for current details.

Put one idea to work

Take the next step with a practical worksheet, a clear research question, and a standard you can use to review the work.

Explore design worksheets →Explore more from UX Design CoachGet design practice in your inbox →