Who Read the Transcript? A Buyer's Checklist for Qualitative Evidence When AI Does the Summarising
Photo: U.S. Army RDECOM from Aberdeen Proving Ground, MD, USA, Public domain, via Wikimedia Commons
Something has changed in the qualitative supply chain, and it has changed without much announcement. In a growing share of UK studies, the person who first reduces forty-five minutes of conversation to a paragraph is not a researcher. It is a summarisation model, applied to a transcript, producing a condensed account from which themes are then drawn.
Summarisation is a legitimate tool. It is also a lossy compression performed by a system with no stake in the decision the research is meant to inform. The procurement question for 2026 is not whether AI is used — it almost certainly is — but whether the trail from finding back to evidence survives the compression.
What summarisation actually removes
A transcript carries more than content. It carries hesitation, sequence, contradiction and emphasis. A respondent who says the right thing about the product and then spends four minutes on the service experience is telling you something about relative salience that a clean thematic summary will flatten.
Summaries also make weighting invisible. When a model condenses twelve interviews into a page, some material is retained and some is not, and the basis for that choice is not a judgement anyone can interrogate. In human analysis, the equivalent choice is at least attributable to a named analyst with a method.
None of this makes the output wrong. It makes it unauditable unless the buyer specifies otherwise.
Verbatim integrity is the first test
The plainest question to ask a supplier is also the most revealing: can you show me the original words behind this theme, in the respondent's own speech, with an identifier that lets me find them in the transcript?
If the answer involves a paraphrase, or a quote that has been tidied for readability without a note saying so, the deliverable has already lost a layer of provenance. Agree at commissioning that verbatims are reproduced exactly, that any edit is marked, and that quotation marks are never used around text that was not said.
Watch, too, for composite quotes — sentences assembled from two respondents. They are common in AI-assisted drafting and they are indefensible in a board pack.
Coding provenance
A theme is not a discovery; it is a decision about similarity. Ask who or what decided that these six comments belong together, and whether the codebook existed before the analysis or emerged during it.
Practical asks: a codebook with definitions, a note on whether coding was deductive, inductive or mixed, the number of interviews contributing to each theme, and evidence of a second coder or a check on a sample of transcripts. Where a model performed the clustering, ask what it was asked to do and whether the instructions — the prompt — are documented and retained.
Prompts are method. A supplier who cannot produce the instruction set used to condense or cluster the data is in the same position as a quantitative supplier who cannot produce the questionnaire.
Researcher judgement, named
The value of qualitative work lies in interpretation, and interpretation belongs to a person. The deliverable should make clear who read the transcripts, who weighed the evidence, and where their judgement departed from the machine's first pass.
A useful marker in any 2026 report is a short methodological page stating which stages were automated, which were human, and who signed off the analysis. Buyers should treat the absence of that page as a signal, not an oversight.
The data protection layer
Qualitative material is personal data, and in UK research it is often sensitive personal data. Before transcripts are uploaded to any third-party service, the buyer should know: which processor is involved, where processing occurs, whether the material is used for model training, how long it is retained, and how deletion is evidenced.
Check the supplier's record of processing and their transfers position. Consent language in the recruitment materials should reflect the processing that actually happens, not the processing that happened when the template was written. This is unglamorous and it is the part most likely to be discovered later by somebody less forgiving.
A buyer's checklist
- Will verbatims be reproduced exactly, with edits marked?
- Can every theme be traced to specific respondents and transcript locations?
- What is the codebook, and when was it fixed?
- How many respondents contribute to each theme, and how many contradict it?
- Which stages used AI, for what purpose, and with what instructions?
- Who read the full transcripts, and who signed off the interpretation?
- Are negative and deviant cases reported, or only majority themes?
- Where is the data processed, retained and deleted, and on what basis?
- Is raw audio or full transcript retained and available on request?
- What happens to the analysis if the model is changed or withdrawn?
The tenth question is the one buyers forget. A method that cannot be reproduced once the vendor's model version moves on is not a method that will survive a challenge two years later in a boardroom or a procurement review.
Setting the standard in the brief
Most of this is settled at commissioning, not at delivery. Specify in the brief that the deliverable must include a provenance page, an annotated sample of coded transcript, and a statement of automated processing. Make acceptance of the final report conditional on those being present.
That is not obstructionism. It is the same discipline the industry already applies to sample sizes and questionnaire wording on the quantitative side. Qualitative evidence has simply had longer to rely on the reputation of the researcher in the room — and in 2026, the room is no longer the only place where the reading happens.