How to run user interviews with AI without losing the why

Start with the decision

A good interview round answers one decision. “Should we ask for the card before or after the trial?” is a decision. “Learn about onboarding” is not.

Write the decision at the top of the script. Every question should help settle it; anything else belongs in a later round.

Recruit from behaviour, not from a mailing list

The most useful people are the ones who did the thing you want to understand: they left at step 2, they cancelled, they never invited a teammate. Analytics tells you who they are.

An AI teammate can pull that cohort, write a short personal invite and send it where the user already is. Most people will ignore it, and that is fine: small rounds are enough.

Write the script once, let follow-ups do the work

Keep the script short: a warm-up, three to five open questions and a closing question about what they would change. Avoid questions that suggest the answer.

The real difference between an interview and a survey is the follow-up. When someone says “I just wanted to try it”, the next question is “what made you want to try it first?”. An AI interviewer can be told to always ask one follow-up before it moves on.

Synthesis: themes, counts, clips

A useful summary has three layers: the themes, how many people mentioned each one, and the exact quotes and clips behind them. Without the third layer nobody can check the first two.

Tie each theme back to the behaviour you started from. “Five of eight people who dropped at step 2 said the card request felt early” is a finding. “Users dislike friction” is not.

Where humans stay in the loop

Approve the script and the cohort before invites go out. Watch a few clips yourself. Decide what to change as a team: an interview tells you what people experienced, not what the product should be.

AI interviews also raise consent questions. Tell people they are talking to an AI, say how the recording is used, and let them stop at any time.