AI User Research: Plan, Analyze, and Check Findings

AI user research means using AI to help prepare studies and organize evidence from real users, while researchers check the findings and people make product decisions. Start with a specific research question, work only with approved material, and require each finding to point back to an observation or interview passage.

A plausible theme is not evidence that users have a problem. Tools like fenbs can keep the research question, approved findings, and product decisions available across assistant sessions, so a later summary does not quietly turn a guess into a settled fact.

How do you run a useful first AI-assisted study?

1. Write the decision the research will inform. Turn “understand onboarding” into a question such as “where do first-time users become unsure what to do next?”
2. Ask AI to draft a discussion guide, then remove leading questions and assumptions. Use the guide with real participants who match the question you need to investigate.
3. Agree how recording, transcription, and AI processing will work. Prepare an approved, de-identified copy of the material before sending it to an AI tool.
4. Ask AI to label relevant passages and suggest themes with participant identifiers and source locations. Require it to report contradictory evidence and missing information.
5. Review the passages yourself. Publish checked findings with their limits, then record separately what a person decided to change and what still needs investigation.

Make the first study small enough to inspect closely. A useful pilot tests whether your question, guide, and analysis process produce evidence you can use. It should not become a product roadmap simply because an assistant can generate a polished report.

What can AI help you plan without pretending to be a user?

Use AI to sharpen a research question, suggest follow-up prompts, organize a session agenda, or identify assumptions in a recruitment brief. Give it the actual decision and known constraints. Ask which questions are unsupported, rather than asking it to predict what participants will say.

Nielsen Norman Group’s interview guide↗ (opens in a new tab) explains interviews as a way to learn about people’s experiences, needs, and views through questions and follow-up. Ask about a recent concrete experience: “Tell me about the last time you tried to invite a coworker.” That gives you something to investigate beyond a prediction of future behavior.

Generated personas and simulated interview answers can help you brainstorm what to ask. Label them as invented examples and keep them outside the evidence set. The US Digital Services Playbook↗ (opens in a new tab) recommends involving actual users in understanding needs and testing services. An AI role-play cannot establish what real participants experienced.

What material should you give the analysis tool?

Give it only material your research process and organization permit it to handle. Check what participants were told, who can access the tool, and how its data settings fit your requirements. If these are unclear, resolve them before upload; a de-identified copy is not permission by itself.

Replace direct identifiers with study-specific labels such as P01. Remove details that could reveal someone indirectly, including an unusual job, exact location, or distinctive incident when they are unnecessary. Keep the mapping to identities in appropriate restricted storage, separate from the analysis copy.

• Include the research question, session identifier, and stable passage or timestamp references.
• Explain which notes are direct quotations, observer notes, or researcher interpretations.
• Mark unclear audio and missing transcript sections. Do not ask the model to reconstruct speech that was not captured.
• Provide definitions for your labels and exclusions, including whether facilitator comments should be ignored.

Keep the approved handling rules with the project. In fenbs Decisions and rules, people can record what assistants may use and what must remain outside the shared work record.

How do you turn interview passages into defensible themes?

Ask for a table of candidate observations before a narrative summary. Each row should contain a participant label, source reference, a short extract, a proposed label, and the reason that label applies. This makes it easier to spot a theme built from the researcher’s question instead of the participant’s answer.

Nielsen Norman Group describes thematic analysis↗ (opens in a new tab) as organizing qualitative material through codes to identify themes. A code is simply a label attached to a relevant observation. Let AI propose labels, then compare them with the passages and revise definitions when similar labels mean different things.

Count participants separately from mentions. One person repeating a complaint does not represent several people. Describe patterns within this study without implying population-wide prevalence. Record contrary examples and competing explanations, especially when a strong theme appears to confirm what the team already wanted to build.

What should you check before sharing findings?

Open the source for every claim that will influence a consequential choice. Verify quotations word for word, and check nearby context for qualifications or corrections. If the transcript itself is uncertain, consult the permitted recording or mark the passage unresolved rather than treating cleaner AI wording as a quotation.

Ask a second reviewer to challenge the strongest interpretation when possible. Can the evidence support a narrower claim? Did your recruitment miss a relevant group? Could the difficulty come from the test setup? Keep those limits alongside the finding so the next presentation cannot drop them accidentally.

Decide who needs which version. A broad product team may need a finding and its limitations; an authorized researcher may need the detailed source trail. Keep restricted participant material out of the shared summary. The point of a shared AI brain is useful project context, not an unrestricted archive of participant information.

How to do this with fenbs?

1. Create a free account and add a project for one research question.
2. Write the approved scope and data-handling instructions in Decisions and rules. Add tasks for interviews, analysis review, and follow-up research.
3. Keep the approved findings in project context notes, with access-appropriate source references in Links and listings. Leave recordings and identifying material in approved research storage.
4. Record the product decisions people make, then ask a supported connected assistant to save progress: what was reviewed, what remains uncertain, and where the next session should start.

fenbs does not conduct interviews, transcribe recordings, or analyze a research repository automatically. It fits the work of keeping approved context and decisions available between sessions. A dedicated research repository is a better fit for managing recordings and detailed participant evidence.

How do you stop a finding from becoming outdated?

Attach the study date, question, and relevant product version to the summary. When a design changes, record whether the finding still applies or needs a new check. A useful handoff says “the wording was revised; the revision has not been tested,” rather than suggesting the research problem is solved.

For shared upkeep, read the guides to a shared AI brain and AI memory for teams. The AI at work hub covers other practical workflows. Start this one by reviewing a single proposed finding and making its evidence, uncertainty, and next decision visible.