How to use AI in your design process, from brief to review

Start by deciding what needs to become clearer. A design process has uncertainty about the problem, the audience, the solution and the implementation. AI can accelerate parts of the work, but generating more material does not automatically reduce those uncertainties. Give each task an input, an output and a person who decides whether the result is useful.

1. Frame the problem

Write the user, task, context and consequence of failure. Supply verified facts and label assumptions. Ask AI to find contradictions, missing constraints and alternative explanations. Your output is a one-page brief, not a long list of features. A human owner chooses the question worth investigating.

For a workshop site, the problem might be that visitors cannot establish which class matches their experience. “Add an AI recommender” is only one possible response. Better descriptions or a comparison view might solve the uncertainty more simply.

2. Gather evidence and preserve its context

Use AI to draft neutral interview questions or organize approved research notes. Keep source identifiers attached. Review summaries against the originals and include observations that challenge the dominant story. Your output is an evidence table distinguishing what people did, what they said, what you inferred and what remains unknown.

3. Explore meaningful alternatives

Ask for three approaches that differ in interaction structure: guided selection, comparison or a searchable catalog. Require tradeoffs and failure states. Sketch enough by hand to establish the intended hierarchy and brand character. Your output is an option comparison with a reason to test or reject each direction.

4. Build the smallest useful prototype

Choose the format that can answer the question. Static screens can reveal some content and hierarchy issues; a keyboard path requires an interactive implementation. Let AI help construct repetitive states and sample data. Review every rule and label before a study. Your output is a testable flow plus a list of simulation limits.

5. Test, decide and measure

Observe relevant people doing realistic tasks. Use AI to assist transcription and initial grouping only under the agreed data-handling arrangements. A person verifies findings and chooses the next change. Before release, review accessibility, responsive behavior and implementation quality. After release, measure the user outcome and suitable guardrails.

Maintain a decision log with the evidence, alternatives, owner and revisit condition. When AI proposes a change, ask which observation it addresses. When no evidence exists, label the suggestion as a hypothesis and design a way to learn.

A brief you can reuse

Help with [bounded task]. The user goal is [goal]. Verified evidence is [source IDs]. Constraints are [rules]. Produce [artifact]. Separate supplied facts from assumptions. Include contradictions, missing states and questions that require human research. Stop before external publishing or changes to live customer data.

Evaluate the workflow by the quality of its accepted decisions and total rework, not by how much content the model produces.