Use AI to move from research to better questions

Decide what AI is allowed to help with

Start with a narrow assignment. AI might help group de-identified observations, propose questions, or draft a research summary for review. Do not ask it to fill missing interviews, invent customer quotes, or decide that a hypothesis has been proven.

Before adding research material to any tool, check your organization’s approved use, the applicable data controls, and the consent or permission under which the material was collected. Remove unnecessary identifying information. Avoid uploading sensitive customer or business data to an unapproved service.

Give every observation an address

Prepare notes with stable IDs. For example, O-01 might identify a verified observation about a participant looking for a preview. Keep the original note accessible to authorized reviewers. Separate what someone did from your interpretation of why they did it.

A proposed instruction: “Group the supplied observations into tentative themes. For each theme, list the supporting observation IDs, contradictory examples, and unanswered questions. Use only the supplied material. Label anything else as a hypothesis.”

This instruction is a starting point to test, not a guarantee that a model will follow every constraint. Human review remains part of the workflow.

Audit the draft against the actual notes

Check every cited ID and every quotation. Remove claims that have no support. Look for minority experiences or conflicting observations that were flattened into a convenient theme. Ask whether the summary confuses frequent mention with importance.

When a theme is plausible but weakly supported, keep it in the question list. Do not quietly turn it into a finding. The useful output may be a focused follow-up interview question rather than a recommendation.

Move from a finding to several possible responses

For a verified problem, ask for several ways to address it, including a non-AI option and an option that requires little new implementation. Have the tool explain assumptions and possible failure cases. Treat the suggestions as drafts.

A reviewer should then judge relevance, feasibility, accessibility, and customer risk. Choose one response worth prototyping, not the option that sounds most impressive. Record the reason for choosing it and the evidence it depends on.

Create a reviewable experiment brief

Use the Research-to-Action Brief to connect the original observation, the proposed explanation, the change, and the measurement plan. Assign a human owner to the final claims and decisions. Keep version notes when the brief changes.

Evaluate the workflow itself as well: did reviewers spend less effort on organization, or merely more effort correcting confident errors? Did important contradictory evidence survive the summary? Were data-handling requirements followed?

The goal is not to produce more research-looking documents. It is to move responsibly from evidence to a question worth testing. A smaller, checked brief is more useful than a polished account of things no customer actually said.