PM Mapped
Working with AI QUALITATIVE · Advanced User Interviews

Advanced User Interviews with AI

At advanced level, AI extends the interview at its edges — transcription you can query, first-pass synthesis across many sessions, and critique of your own technique. The synthesis is where insight is made or lost, and AI clustering is a starting point, not the answer. Read the transcripts; never let it tidy a quote.

← Back to the Advanced User Interviews tool
1How to use AI for this

A step-by-step way to work

1

Transcribe and make it queryable

Use transcription so you can later ask 'where did people mention pricing?' across all sessions — you analyse instead of typing.

2

Cluster across interviews

Ask AI to find themes spanning multiple sessions, with mention counts, then verify each against the raw quotes.

3

Hunt for disconfirming evidence

Ask 'what did people say that contradicts the emerging theme?' — AI is useful for finding the counter-examples you're motivated to ignore.

4

Critique your technique

Paste a transcript and ask where you led the witness, asked a closed question, or missed an obvious follow-up.

5

Keep quotes sacred

Never accept a 'cleaned up' quote. The exact words are the data.

2Worked examples

Real prompts, and what good output looks like

Finding disconfirming evidence
Your prompt

My emerging theme is 'users find the reporting feature too complex.' Across these 6 transcripts [pasted], what evidence contradicts that theme?

What good output looks like

2 of 6 actively praised the depth of reporting ('I switched from a competitor because their reports were too basic').

1 said complexity wasn't the issue — discoverability was ('once someone showed me, it was fine').

So the theme may be too broad: the friction could be 'hard to discover' rather than 'too complex', and a power-user segment values the depth. Worth segmenting before concluding.

Why this matters: confirmation bias makes you skim past the two who disagreed. Asking AI for the counter-evidence is a cheap guard against a wrong conclusion.
Critiquing your own interviewing
Your prompt

Review this transcript excerpt and tell me where I, the interviewer, made mistakes [excerpt pasted].

What good output looks like

You asked 'so the slow load time was annoying, right?' — leading; it suggested the answer. Better: 'how did the load time affect what you were doing?'

When they said 'it was fine I guess', you moved on — a missed follow-up. A simple 'tell me more about 'I guess'' might have surfaced a real reservation.

Otherwise your open-ended openers were strong.

Cheap coaching: AI won't catch everything, but it reliably flags leading questions and missed probes you can fix next time.
3Copy-paste template

A prompt you can reuse

Fill in the highlighted parts and paste it into your AI tool of choice. Edit the output — it's a starting point, not a finished answer.

Reusable prompt
Help me synthesise advanced discovery interviews. Treat clusters as drafts to verify, and never alter a quote.

Transcripts / observations from [N] interviews:
[paste]

Please:
1. Cluster observations into themes that span multiple interviews; label each and give mention counts.
2. Explicitly list evidence that CONTRADICTS each emerging theme.
3. Flag any theme supported by only one participant.
4. Suggest where I should segment rather than generalise.
Keep all quotes verbatim.
4Common pitfalls

What AI gets wrong here

Confirmation bias, amplified

If you prompt toward your hoped-for theme, AI will helpfully find support for it.

Do this instead: Always ask for disconfirming evidence and single-source flags as part of the same request.

Tidied quotes

Summarising can silently reword what participants said.

Do this instead: Demand verbatim quotes; use AI for grouping and counting only.

Premature generalisation

AI will state a theme as universal when it's really one segment's view.

Do this instead: Ask it to identify segments and to mark how many participants support each theme before you generalise.
The judgment that stays yours

Synthesis is the heart of discovery, and it's where a confident AI summary can do the most damage — merging distinct issues, smoothing over contradictions, generalising from a couple of voices. AI earns its place by transcribing, counting, and surfacing counter-evidence faster than you can. The insight — deciding what the research actually means — comes from your own reading of the words people really said.