AI for user research: preserve the evidence, improve the questions

The most useful research output is a better decision, not a larger repository. Begin with the decision your team cannot make confidently. AI can help sharpen a plan, sort material and challenge interpretations, but the researcher remains responsible for context, method and claims.

Plan around uncertainty

Write the question, who can answer it and what evidence would change the decision. Choose a method that fits. Interviews can reveal experiences and workarounds; observation can reveal behavior people forget to mention; analytics can show patterns when the events are valid. Do not ask a language model to choose the method without providing the constraints.

Use AI to review an interview guide for assumptions. Replace “How frustrating is the current dashboard?” with “Walk me through the last time you checked progress.” Ask about sequence, tools, interruptions and consequences. The goal is to learn the participant's situation rather than confirm your framing.

Prepare a controlled evidence set

Use the agreed research repository and approved processing tools. Remove unnecessary identifying details and retain consent and retention arrangements. Give every excerpt a stable identifier, such as P04-S12. Keep enough surrounding context to understand what the participant meant.

A model summary should point back to a source. If it cannot, it belongs in an assumptions column. Preserve the originals so another researcher can inspect the interpretation.

Use two passes for synthesis

First, ask for a descriptive pass: what happened, in which context, and where the source supports it. Then ask for candidate interpretations and alternative explanations. This reduces the chance that an appealing theme silently changes what the material says.

Review disconfirming cases. If four participants describe a problem and one uses a successful workaround, that difference may reveal a design opportunity. Avoid converting qualitative frequency into a claim about the whole customer base. The recruitment method determines what the sample can represent.

Make a finding inspectable

Use this structure: observation, supporting sources, interpretation, confidence, implication and next question. For example, “Participants used personal reminders after booking” is an observation only if the source supports it. “Our reminders are inadequate” is an interpretation requiring further investigation.

Ask AI to act as a skeptical reviewer: which recommendation outruns the evidence, what context is missing, and what other explanation fits? A human decides whether the criticism changes the conclusion.

Keep learning social

Bring colleagues into selected observations and synthesis discussions. A concise AI summary can help people catch up, but it should not become the only contact between the team and the people using the product. Document disagreement and revisit the original material together.

Close the loop by recording what changed because of the research. If nothing changed, explain whether the evidence confirmed a direction, was inconclusive or arrived too late. That is more useful than inflating the number of generated insights.