Colay / Guides
Synthesize customer interviews without depending on one model's interpretation
Several AI interpretations can help you notice different patterns in 20 interviews, but repeated themes across models are not additional participant evidence. Use Colay to compare proposed codes and themes, then verify every important finding against the permitted source excerpts. The result should be an evidence table that another researcher can inspect, not a vote on which narrative sounds convincing.
Prepare the evidence before asking for themes
Confirm that the research permissions and your organization's rules allow the material to be shared with the selected service. Remove identifiers and unnecessary sensitive details. Assign each interview a stable ID and each excerpt a location such as paragraph or timestamp. Keep the identifying key outside the material you submit.
If the material does not fit the available context, work in explicit batches with the same coding instructions. Save the batch membership and preserve links to the original excerpts. Do not assume that a model has read an omitted transcript or that separate chats automatically remember one another.
Keep observation, interpretation and action apart
GOV.UK's research analysis guidance distinguishes what was observed from the finding derived from it and the action that follows. In an AI-assisted workflow, make those three levels explicit in the requested output. A customer quote should not silently become a product requirement.
First ask available agents through Ask separately to propose codes from the same excerpts. Then compare their definitions, supporting IDs and exceptions. Different labels may describe the same behavior; matching labels may conceal different meanings. Review the underlying material before merging either.
Example: repeated words are not repeated participants
Fictional dataset: 20 interviews about appointment booking. Five participants mention difficulty changing a booking. One participant, I07, describes the same incident three times. Another, I12, says changing a booking was easy but understanding the confirmation was difficult. These are invented inputs for illustrating the method.
Do not turn I07's three excerpts into three independent participants or merge I12 into the same problem without reading the distinction. Even a verified count of five among these 20 participants describes this sample; it does not establish prevalence across all customers.
A useful disagreement would be one model coding the material as navigation difficulty and another as uncertainty after the change. Return to the excerpts and decide whether these are separate themes or stages of one experience. Preserve a contradictory case rather than deleting it to make the story cleaner.
| Proposed finding | Evidence to attach | Check before using it |
|---|---|---|
| Changing a booking can be difficult | Distinct participant IDs and exact excerpt locations | Do the excerpts describe the same difficulty? |
| Confirmation can remain unclear | I12 and its supporting excerpt | Keep separate from difficulty making the change |
| Proposed design action | Finding IDs and an explicit hypothesis | Test the change; do not present it as a proven solution |
A prompt for traceable interpretation
For synthesis, provide Consensus with the original excerpts and labelled coding proposals. Ask it to identify overlaps and disagreements without treating model agreement as additional evidence. Keep source IDs through every rewrite. If a generated quote cannot be found exactly, remove it and return to the source rather than repairing it from memory.
Research question: [question]. Permitted de-identified excerpts with interview IDs and locations: [material]. Propose codes and themes. For each theme, return its definition, supporting interview IDs, exact excerpt locations, contradictory cases and alternative interpretation. Count distinct participants only when the supplied material supports that count; never infer missing interviews. Separate observation, interpretation and proposed action. Do not invent quotes. Flag claims that require checking the original transcript.
Finish with a research artifact the team can challenge
Deliver a codebook, evidence table, unresolved interpretations and proposed next research questions. A reader should be able to move from a claim to the relevant excerpt and understand why you accepted that interpretation. Review all consequential findings, including ones on which every model agreed.
Compare the work by the quality of that traceability and the effort of review, not the number of generated insights. Colay uses credits and limits; start with a manageable permitted excerpt set before expanding. Several models can broaden the candidate interpretations, but your research judgment and participant evidence remain the basis of the conclusion.
Questions, answered
Are several AI analyses independent sets of evidence?
No. They are interpretations of the same participant material and may share errors. Evidence comes from the actual research, with its sampling and methodological limits.
Can I upload all 20 interviews at once?
Check permissions and the available input limits first. If needed, use documented batches and preserve source IDs; do not assume omitted material was considered.
Should I keep a theme only one model found?
Keep it for review if the source material supports it. Frequency among model outputs is not the criterion for accepting or rejecting an interview finding.
Sources and methodology
- GOV.UK — Analyse a research session
Primary guidance for separating observations, findings and actions. It does not validate AI-generated coding or the fictional interview dataset.
Bring your next question to Colay
Choose a model, use Auto, or bring several perspectives together with Consensus.