AI Workflows · Guide
Turn Customer Interviews Into Themes Without Inventing Consensus
Use AI to organize observations while keeping quotes, participants, disagreements, and human interpretation connected to the evidence.
AI can help cluster interview observations, but it should not decide what participants collectively believe. Extract evidence with participant references, keep observations separate from interpretations, review clusters as a team, and report exceptions as well as common patterns.
Protect consent and source records
Use only interview notes or transcripts covered by the participant agreement and the team’s data policy. Remove unnecessary identifiers before any approved AI step. Keep the original recording or transcript in its controlled location and give each session a neutral identifier. AI should receive the minimum text needed for the analysis task, not the entire research archive by default.
Extract observations before themes
Create one evidence note per observed behavior, statement, need, or difficulty. Preserve the participant ID and enough context to understand it. GOV.UK guidance distinguishes observations from what researchers think the observation means. Ask AI to format or deduplicate candidate notes, then compare them with the transcript. Do not let a summary replace the evidence layer.
Cluster without counting people twice
Group related observations, but preserve which participants contributed to each cluster. Repeated comments from one interview are not multiple independent participants. Keep contradictions and isolated but consequential observations visible. Ask the model to propose neutral cluster labels and alternative groupings, then let researchers move notes and agree on names. A theme is a human interpretation supported by evidence, not a word-frequency result.
Write findings with limits
A useful finding states what happened, who it affected in the sample, why it matters, and what remains uncertain. Avoid “users want” when only one person said it. Include an example quote only after checking it against the source and consent. Report disagreement when it changes the design decision. Separate the finding from the proposed action so a persuasive solution does not distort the evidence.
Choose the next research action
Use themes to decide what to test, clarify, or change next. Invite observers to challenge the interpretation and inspect the linked notes. Save the cluster, evidence references, reviewer, and date. If the AI grouping hides important variation or cannot maintain traceability, return to manual affinity sorting. The workflow succeeds when it makes evidence easier to discuss—not when it produces more confident summaries.
Use an evidence table
Create one row per finding with the research question, observation IDs, participant count, contradictory evidence, interpretation, confidence note, and proposed next step. The participant count is not a population estimate; it only describes the sample. Ask AI to check whether every cited observation exists and whether the same session has been counted twice. A researcher should verify the table before it is shared. This small structure makes it harder for a memorable quote to become a universal conclusion.
Example: a confusing onboarding step
Suppose four of eight interviewees pause at an optional profile question, two continue without concern, and two never reach the step. A weak summary says “users find onboarding invasive.” A grounded finding says that four observed participants hesitated because the purpose of the question was unclear; two explicitly asked why it was needed; the remaining sessions do not establish the same reaction. The next action might be to explain the purpose and test the revised step, not remove all profile questions. AI can locate passages about the step, but a researcher decides whether the evidence supports that action.
Share evidence without exposing participants
Use neutral identifiers and only the minimum quotation needed to support a finding. Remove contact details and incidental personal information. Check whether a distinctive quote could identify the participant inside the organization. Share links to controlled source material only with approved colleagues. The final presentation should focus on the research question, pattern, exceptions, and next action rather than creating a searchable archive of private stories.
Keep a human interpretation note
Add two sentences explaining why the cluster became a finding and what alternative explanation was considered. This forces the team to distinguish a pattern from an assumption. When new sessions disagree, update or split the finding instead of asking AI to smooth the evidence into one story. Research remains useful when uncertainty is visible and can direct the next round.
Audit the final summary against the evidence
Before sharing, select each headline claim and follow its links back to the observation notes and sessions. Confirm that the wording matches the sample, includes material exceptions, and does not count the same participant twice. Ask a researcher who did not write the summary to challenge one alternative explanation. Fix the evidence links or narrow the claim rather than asking AI for a more persuasive sentence. The result should help a decision maker understand what was observed, what the team inferred, and what still needs research.
Sources and update note
Follow the linked official source before a product, price, plan, or policy decision.