Interviews & synthesis
Use AI before the interview (better questions), never during trust-building, and hard after it — turning raw transcripts into grounded, quote-cited themes.
Discovery interviews are where investigations are won, and most analysts under-prepare for them and under-process them. AI transforms both ends of the interview while leaving the middle — the human conversation — alone.
Before: AI as prep partner
I'm a business analyst investigating a 40% quarter-over-quarter increase in refund costs at a coffee retailer (web store + 3 physical locations). I'm interviewing the customer support team lead for 30 minutes. Draft 8 questions: start broad, then probe for when the change started, what patterns support agents see (products, channels, reasons), and what data or systems would show it. Avoid leading questions — I don't want to plant my hypotheses. Then list 3 follow-up probes I should have ready if they mention a specific product or a specific time period.
The 'avoid leading questions' instruction matters: AI happily writes questions that presume your favorite theory, and interviewees happily confirm whatever you presume.
- Ask AI to role-play the interviewee for five minutes before a high-stakes conversation: 'Play a defensive warehouse manager who thinks this investigation is looking for someone to blame.' Rehearsing the hard version makes the real one easier.
- Have it draft a 30-second framing statement for opening the interview — what you're investigating, what you're not (blame), what happens to their input. Trust up front changes what people tell you.
After: synthesis with receipts
The post-interview move that separates professionals: synthesize the same day, with AI doing the heavy lift and you doing the judgment. The critical instruction is to demand verbatim quotes for every theme — an AI synthesis without quotes will occasionally invent a plausible theme nobody actually said, and you won't catch it three weeks later when it's in your findings deck.
Here are my notes/transcript from an interview with [role]. Synthesize into: (1) 3-6 themes, each with 1-2 VERBATIM quotes from the transcript as evidence — never paraphrase inside quote marks; (2) facts asserted that I should verify against data, as a checklist; (3) what this person believes is the cause, stated as THEIR belief, not as fact; (4) open questions this interview raises. If a theme has no direct quote supporting it, don't include it. [paste transcript]
Item 3 is the discipline that keeps you honest: interviewees hand you conclusions wearing the costume of observations. Label beliefs as beliefs until data promotes them.
Across multiple interviews, run a second pass: paste all the single-interview syntheses and ask for cross-interview patterns — where sources agree, where they contradict, and whose claims are checkable. Contradictions aren't noise; they're usually the most informative thing you collect. When support says 'it started in spring' and fulfillment says 'nothing changed on our end,' at least one of those is wrong in an interesting way.
Don't type prompts mid-conversation, and be careful with AI note-takers — in sensitive interviews (and a cost investigation is sensitive), a recording bot changes what people say. Take human notes, record only with consent, and save the AI for the same-day synthesis. The interview itself runs on eye contact and follow-ups, which remain stubbornly analog.