You did 12 customer interviews this month. You have 12 Otter transcripts. You probably listened to 6 of them and skimmed notes on the rest. The problem is not the interviews. It is synthesis. Individual interviews are full of noise, and the signal only shows up when you connect dots across 8 or more of them. But manually reading 12 transcripts and building a theme map takes 6 hours you do not have.
Feed it transcripts plus segment metadata
The agent processes three inputs in sequence. First, transcript parsing: it reads every transcript (Otter, Fireflies, manual notes) and pulls who was interviewed, key quotes, pains mentioned, and features discussed. Second, theme extraction across all interviews. Not just "they mentioned onboarding" but "8 of 12 said onboarding was slow AND tried workarounds that failed AND took 2+ weeks to get production-ready." Third, hypothesis formation.
The setup needs Otter or Fireflies for transcripts, Notion or Google Docs as the research repository, your CRM to map each interview to segment and company size, and your historical interviews so the agent can tell new themes from recurring ones. It runs every Wednesday at 10 AM on the past week's interviews.
The report runs six sections
You get one report with recurring themes (anything in 3+ interviews, with 2 to 3 representative quotes each), top pains vs. gains ranked by frequency and severity, feature mentions, testable hypotheses, contradictions between interviews, and segment insights comparing enterprise, mid-market, and SMB.
The section that matters most is testable hypotheses. The agent converts each theme into a statement specific enough to test. "We should make onboarding faster" is not a hypothesis. "Small engineering teams abandon API onboarding because our docs assume a dedicated integration engineer" is. And the report keeps quotes throughout, so you hear the customer voice instead of a flattened summary.
Why the 8-interview threshold matters
One person saying onboarding is slow is an anecdote you should not build a roadmap on. Eight people saying it, across different company sizes and roles, with the same workaround pattern, is a signal. The whole reason to automate this is that no human reliably holds eight transcripts in their head at once and spots the cross-cutting pattern. The agent does, and it flags contradictions early too. "Enterprise loves our API, SMB finds it confusing" is a messaging fix you want to catch now.
To start, feed in your last 10 interview transcripts with segment metadata for each interviewee, run it, and ask for the top three testable hypotheses. That output is your research backlog.
FAQ
What does the agent do with the output? It gives you synthesis in one place, actionable hypotheses specific enough to test, and conversation velocity so a new interviewer can read the synthesis instead of listening to all 12 recordings.
What data sources does it connect to? Otter or Fireflies, Notion or Google Docs, your CRM for segment mapping, and your historical interviews so it distinguishes new themes from recurring ones.