User research synthesis with AI: keep every finding connected to evidence
User research synthesis is the work of turning individual observations and responses into an account of what the team has learned. A summary can shorten the material. Synthesis goes further: it asks what the material means for a research question, how responses fit together, and where the explanation breaks down.
AI can help teams navigate a large set of answers, but a convincing paragraph can still overstate the evidence. The useful habit is to keep moving between the finding and the source. In Palette, summaries of written and recorded responses link back to original answers, giving reviewers a practical way to inspect the connection.
Begin with the question and the collection boundary
Before grouping answers, restate the research question. Then record who participated, how they were recruited, what they were asked, and when collection happened. These details determine the boundary of the finding. A study of experienced administrators describes a different experience from a study of first-time members, even when both contain comments about invitations.
Keep method differences visible too. A written account of a task is different from evidence of someone completing it. A microstudy on an invitation page reaches people who arrived there. If you combine sources, preserve those distinctions so that a later reader can understand what each source contributes.
If your material spans a product change, separate responses collected before and after it. An apparently contradictory answer may describe a different version. Similar care applies when people work under different permissions, account configurations, or organizational rules. Context is often part of the explanation rather than clutter to remove.
Separate observations, interpretations, and possible actions
Consider a hypothetical report-sharing study. “A participant exported a report and removed two columns before sending it” is an observation about an account of behavior. “The report contains information that is unsuitable for some recipients” is an interpretation. “Add a sharing preset” is one possible action. Those are three different statements with different levels of support.
Write them separately. The observation may be clear while the interpretation needs checking: perhaps the columns were irrelevant rather than sensitive. Even if the interpretation holds, several responses are possible. Research helps explain the situation; the team still chooses what action to take.
GOV.UK’s research analysis guidance similarly distinguishes observations, findings, and actions. That separation is especially useful when reviewing AI-generated text, where a fluent sentence can move through all three without signaling the change.
Group by meaningful patterns, not just shared words
People may use the same word for different problems. Several comments mentioning sharing might concern access, formatting, delivery, or finding the right audience. Conversely, comments about emailing, forwarding, and exporting may describe the same underlying job. Read enough surrounding material to identify the experience the language refers to.
Treat an initial set of themes as provisional. Open responses assigned to a theme and ask what they have in common. Split a broad category when it hides distinct needs. Combine categories when the difference is merely vocabulary. Do not use the number of repeated phrases as a substitute for understanding the number of participants and contexts involved.
- Check whether each theme answers the study’s research question.
- Distinguish one person repeating a concern from several people reporting it.
- Preserve differences between audiences and product versions.
- Keep a place for answers that do not fit the main explanation.
Inspect the evidence behind consequential claims
For each finding that may affect a decision, open its supporting responses. Confirm that the claim matches what participants said or did, that quoted wording is accurate, and that the surrounding question did not supply the explanation. Look at whether a finding rests on a single vivid response or appears across relevant contexts.
A citation is a route to evidence, not a quality score. It may point to an ambiguous answer, a secondhand account, or a response that does not fully support the claim. If the support is weak, narrow the finding. “Some interviewed administrators described uncertainty about permissions” may be justified where “Permissions are the main cause of onboarding abandonment” is not.
Avoid converting qualitative counts into population estimates. You can state that a concern appeared among participants in the study, with an appropriate denominator and context. Establishing how common it is across all customers requires a suitable sampling and measurement approach. More polished synthesis does not change that requirement.
Use contradictions to plan the next study
Suppose experienced administrators describe invitations as straightforward, while newer administrators report uncertainty. The contrast may suggest a learning effect. It could also reflect different roles, account setups, or task goals. Keep the competing explanations visible and write a follow-up question capable of distinguishing them.
This is where synthesis becomes useful to continuous research. A finding can close one uncertainty while exposing another. Instead of treating every contradiction as an error to resolve in the summary, use it to improve audience selection, question wording, or the next task-based test.
Review the analysis as a team
Give another teammate access to the relevant responses and ask them to challenge the main explanation. A person close to customer support may recognize a missing context; a designer may notice that two complaints concern different stages of a flow. The purpose is to improve the account, rather than secure agreement with the first summary.
In Palette, use the summaries and their source links as a starting point for that discussion. Explore user studies for collecting deeper accounts, or microstudies for a focused follow-up inside the product. When analysis raises a question about completing a task, move to usability testing and gather the kind of evidence the question needs.