22 August 2026
User research needs the participant's exact words, not a cleaned-up version of what they meant, because the argument for changing a product rests on what someone actually said. A summary swaps their sentence for the model's guess at it, and that guess is not evidence a stakeholder can push back on. The fix is to treat the transcript as the source and the summary as a shortcut for finding your way back into it.
When a note taker writes 'user found the flow confusing,' it reads clean and it is almost certainly wrong. The participant might have said 'I wasn't sure it saved,' which is a different problem with a different fix — one is a comprehension issue, the other is a feedback issue. Summarization collapses both into the same adjective because the model is scoring sentiment, not tracking the specific claim.
The cost shows up later, in a readout, when a stakeholder asks 'did she actually say that?' and nobody can produce the sentence. A finding that can't be traced back to a specific person saying a specific thing at a specific moment gets treated as opinion, and the design decision built on it loses its footing.
A sentence on its own, lifted out of a transcript, is not yet evidence. It needs enough context that someone reading the report six months later, who wasn't on the call, can tell what question produced it and who was answering.
One interview is a story. Twenty interviews are only useful once someone can pull every instance of a theme across all twenty, and that only works if the tags were decided before the tagging started, not invented session by session.
A team that lets each researcher name a theme on the fly ends up with 'onboarding friction,' 'signup confusion' and 'first-run trouble' describing the same nine participants, split into three buckets that nobody thinks to merge until the analysis is already written. Searching every past transcript for a fixed phrase — a feature AVAY uses its own transcript search for — only pays off if the phrase was fixed first.
An employee signed an employment agreement that says a work meeting can be recorded. A customer, a candidate, or someone recruited off a panel signed no such thing, and the consent for their interview has to be asked and answered on the record, not assumed from a calendar invite.
Say it out loud at the start of the call — 'I'm recording this for research and it'll be quoted internally, is that alright?' — and it lands in the transcript as proof consent was given, not just as a checkbox in a screening form. If a participant asks for anonymity, that promise has to survive into the report: a participant ID, not a name, and no LinkedIn-searchable detail attached to the quote.
One limitation worth planning around: a recording made on a call is saved to the machine that made it, not automatically archived somewhere a compliance reviewer can find it later, so a research team that needs a durable audit trail for consent should decide where recordings live before the first interview, not after.
None of this makes the automatic summary useless — it just means it's a tool for the pass before the analysis, not the analysis itself. Twenty transcripts is too much to read closely before you know which five are worth reading closely.
Skim twenty auto-generated summaries in the time it takes to read one full transcript, and use them to decide where the interesting sentences probably are. Then go back to the source and pull the actual words. The summary is a map to the transcript, never a citation from it — the moment a summary sentence gets copied straight into a findings deck as if a participant said it, the report is quoting software instead of a person.
| Fit for a finding | Fit for a first pass | |
|---|---|---|
| Verbatim quote | Defensible — traceable to a person and moment | Too slow to produce across dozens of sessions |
| Auto-generated summary | No — wording is invented by the model | Fast way to triage which transcripts to read |
| Researcher's paraphrase | Still not the participant's words | Fine for private orientation, not for citing |
Use the summary to decide which transcripts to read closely, not as the finding itself. The words in an automatic summary are the model's paraphrase, and a design decision built on a paraphrase can't be defended when someone asks what the participant actually said.
The exact wording, a participant ID, the session and timestamp it came from, and the question that prompted it. Without those four things a reader six months later has no way to check the quote or tell whether it generalizes beyond one person.
Decide the tag vocabulary before the first interview and write it down, because a theme named differently by each researcher splits one finding into several that nobody merges later. A tool that can search every past transcript for a fixed phrase only helps once that phrase is fixed.
Yes — an employee's meeting is covered by their employment agreement, but a customer or panel participant hasn't agreed to that, so ask on the record at the start of the call and let the transcript capture the answer. If anonymity was promised, keep a participant ID out of the report instead of a name.
It's the wrong tool for the finding, not for the search. Use it to triage twenty transcripts down to the five worth reading in full, then quote from the transcript itself — never present the summary's sentence as something the participant said.
A research finding is only as strong as the exact sentence behind it — tag the theme before you start, attach the source to every quote, and get consent on the record before you ever run the interview.
AVAY is a video meeting platform that transcribes the call itself — no bot joins, because there is nothing to join. Start one at avay.ai, read how each part works in the documentation, or see what it costs.