Never lose a user interview insight again

Interviewing well is a full-time cognitive task. You are tracking what was said, deciding whether to probe or move on, noticing what was avoided, and keeping the conversation from becoming a survey. Adding transcription to that list means one of them gets dropped, and it is never transcription — typing has an obvious output, so it wins the competition for attention.

What you lose is specific and predictable:

• The follow-up question. The best probe is a response to something that was just said. Two seconds late, the moment is gone and the participant has moved on.
• Everything paraphrased. Notes are compression, and you compress toward what you already expect. The sentence that would have surprised you is the one most likely to be smoothed into something familiar.
• The exact words. Participants' own vocabulary is the raw material of research — how they describe the problem is frequently the finding. Notes preserve meaning and discard phrasing.
• Rapport. A researcher looking at a keyboard gets shorter answers than one making eye contact, and the difference is largest on the questions that matter most.

The usual workarounds each fail in their own way. A second note-taker doubles the cost per interview and changes the dynamic in the room. Recording and re-listening pushes the work downstream and multiplies it, which is why the recordings mostly go unwatched. And relying on memory is worst of all: within a day you retain the interpretation you formed during the interview, not the evidence you formed it from.

The fix is to remove capture from the interviewer's job entirely. If a complete transcript exists by the time the interview ends, the researcher's job during it collapses back to one thing — listening — and the note-taking that remains is the useful kind: three or four marks about where to look again, not a record.

Earkeep records mic and system audio continuously while it runs and transcribes locally on your device, which suits research work for a reason beyond convenience. Remote interviews run through the same call as everything else, so system audio and mic both matter; there is no bot to invite into a call with a customer, and nothing to start, so the interview that begins five minutes early with small talk — often where the useful aside lands — is covered too. Interviews are defined afterwards by selecting the span on a timeline, or automatically from a connected calendar. Audio is processed only in memory and never written to disk, and the transcript is a plain file on your machine.

That last point is not incidental for research. Interview recordings contain named individuals discussing their employers' problems, often under an NDA or a consent form that specifies where the recording may go. Local transcription means the audio never leaves your device and there is no third-party processor to name in a consent form. See private by design for what that involves technically.

Freeing up attention only helps if you know what to spend it on. Three interview shapes cover most product research, and they fail in different ways.

The JTBD interview is the one that most rewards not taking notes. Its method is a timeline reconstruction — walking backwards from the purchase to the first thought of change — and it depends on the interviewer noticing gaps and inconsistencies in the story as it is told. That is unavailable to someone transcribing. It also produces the most quotable material, and the exact phrasing of what pushed someone to switch is worth more than any summary of it.

One rule cuts across all three: past behaviour, not future intention. "What did you do the last time this happened?" produces evidence. "Would you use this?" produces courtesy.

A single interview is an anecdote. The finding lives in what recurs across twelve of them — and that is where notes-based research quietly breaks down, because you are comparing twelve summaries written by someone whose interpretation was shifting as the series progressed. The theme you notice in interview nine is not searchable in the notes for interviews one through four, since it was not a theme when those were written.

Full transcripts kept as plain files change the unit of analysis from your summaries to what people actually said:

• Search the phrase, not the theme. When a word keeps coming up in later interviews, grep the earlier ones. You will usually find it there, unrecorded because it did not mean anything yet.
• Count, roughly. "Seven of twelve described the same workaround" is a different claim from "users often mention." With transcripts it takes minutes to check rather than a re-listen.
• Quote exactly. A verbatim sentence from a participant ends internal debates that a paraphrase prolongs, because the participant's words are not your opinion.
• Test the counter-case. Search for participants who said the opposite. This is the discipline notes make impossible and transcripts make cheap, and it is what stops a research readout confirming what the team already believed.

An agent can do the first pass over a folder of transcripts — group recurring language, pull candidate quotes, flag contradictions — leaving you to check its claims against the source rather than assemble them by hand. Because the transcripts are plain files in a folder you chose, that works with the tools you already use; see feeding meeting context to AI agents and agents and MCP. Keep the transcript as the evidence and treat the agent's summary as a draft, not a finding.

Research reaches a team as a deck, and a deck is a lossy format: it carries the conclusion and discards the evidence, so the first sceptical question — "did they actually say that, or is that your reading?" — has no answer available. Six months later the deck survives and the reasoning does not, which is how a finding turns into folklore.

Link the claim to its source. Each finding cites the interviews it came from, with quotes attached and the transcript one click away. It takes almost no extra effort when transcripts exist, and it changes the readout from an assertion into something a colleague can check. It also means a new team member in a year can reconstruct why the team believes what it believes, rather than inheriting the belief without its basis.

Download Earkeep free for macOS and stop choosing between paying attention and having a record.