On this page
AI-moderated user interviews are a research method in which AI talks one-on-one with participants following an interview guide, asks follow-up questions based on their answers, and then synthesizes many interviews into findings you can trace back to what participants actually said. You keep what interviews do best, uncovering the why, while running dozens to hundreds of sessions at the same time. By the end of this article, you'll be able to tell whether your study is a good fit for AI moderation, know the steps from kickoff to validation, and have a quality checklist you can use right away.
Definition
In an AI-moderated interview, the research lead first writes the research plan and interview guide. The AI asks questions based on the guide, and participants answer by text or voice. The AI doesn't just read through the questions and stop. It interprets each answer before deciding what to ask next: if the answer is vague, it asks about one specific experience; if the answer is already clear, it moves on to the next topic.
It isn't a customer service chatbot, and it isn't a survey with a chat interface. One test tells them apart: do the questions each participant gets change based on what they said earlier? If not, it's just a survey in conversational form.
How AI interviews work
1. Open and move through the guide
The interview guide sets which topics to cover, along with each topic's research goal and suggested questions. The AI opens with an introduction explaining that it's AI, roughly how long the interview will take, and how the transcript will be used. Once the participant consents, it starts with easy background questions and gradually moves into the core topics.
2. Ask follow-up questions based on answers
Follow-up questions are the heart of an AI interview. There are four common directions for probing: ask about one specific recent experience, the steps it involved, what they originally expected, and what they did afterward. Here's an example.
If the AI had moved on after the first answer, all you'd have is “I don't use it much anymore.” After three follow-up questions, the reason comes down to “the billing section is hard to find,” and you also learn which alternative this user switched to.
3. Know when to wrap up
The AI uses each topic's research goal to decide whether the topic is covered. Once it has a specific experience and the reason behind it, or the participant clearly has nothing more to add, it moves on to the next topic. After all topics are covered, the AI asks whether the participant has anything to add, then thanks them and closes. A good AI interview doesn't keep probing the same point just to fill time.
4. Synthesize into findings
Once the interviews are done, AI groups quotes by theme and writes one finding per theme, noting how many participants mentioned it and attaching representative quotes and counterexamples that don't fit. The research lead reviews each finding and accepts, edits, or rejects it. Only the reviewed report goes to the team.
How they compare with surveys and traditional interviews
| Criteria | Survey | Traditional interviews | AI-moderated interview |
|---|---|---|---|
| Questions it can answer | Measuring how many or how much | Why and how it happened | Why and how it happened |
| Follow-up | None; questions are fixed | Yes, quality depends on the moderator | Yes, following rules set in the guide, fairly consistent across sessions |
| Scale per round | Hundreds to thousands of responses | Usually a dozen or so sessions | Dozens to hundreds of sessions |
| Where the time goes | Fast to collect | Scheduling, moderating, and synthesis all rely on people; slowest | Many sessions run in parallel, and AI drafts the first synthesis |
| Basis for findings | Option counts and a few open-ended questions | Researcher notes and recordings | Every finding traces back to quotes |
| Nonverbal cues | None | Can observe facial expressions, tone, and how people use things in person | Voice interviews capture tone but can't see how people use things |
| Best for | Validation and quantification | Early exploration, sensitive topics | Exploring and explaining, plus analyzing reasons at scale |
The three don't replace one another. A common combination: use AI interviews to find the reasons and their types, then use a survey to estimate how common each reason is. When you need to watch people use something, or the topic is sensitive, have a researcher moderate in person.
When it's a good fit, and when it isn't
Good for
- Why users churn or stop renewing
- Concept testing for new products and features
- Whether ads and brand messages are understood as intended
- Reasons behind low satisfaction or NPS scores
- User segmentation and personas: what motivates different groups
- Purchase decisions and willingness to pay
- Running the same study across multiple markets and languages
Less suitable
- Sizing estimates that require statistical inference, such as “what percentage of users will buy,” which call for a survey with a proper sampling design
- Usability tests that require observing real use, especially in offline settings or with physical products
- Highly sensitive topics such as health, financial hardship, or family conflict, where participants need more trust and a person needs to make judgment calls in the moment
- B2B customer research where participants are a small number of key decision-makers: every session is hard to get and worth a senior researcher's time
- Children or other groups who need a guardian present
Steps in an AI interview study
- Write down the business question and the decision you need to make. For example, “Should we redesign the billing page next quarter?” rather than “Understand how users feel about billing.”
- Break the business question into 3–5 research questions, and define participant criteria and sample size. For exploratory research, you can draw on Guest, Bunce & Johnson (2006): in a homogeneous group, few new themes emerge after about 12 interviews. If you have several distinct segments, plan sessions for each one separately.
- Write the interview guide: for each topic, state the research goal, the main question, and directions for follow-up. Have the research lead review it, and if needed, run 2–3 pilot interviews and revise.
- Invite participants: select them from your own users by criteria, or recruit externally. In the invitation, state the length, incentive, and how data will be used.
- AI runs many interviews in parallel while the research lead listens in or spot-checks transcripts and excludes invalid responses.
- Review the report: accept, edit, or reject each insight, and confirm the participant count, quotes, and counterexamples behind every finding.
- Hand findings to the teams that own them, and agree on which behavior data you'll use to validate them once changes ship.
How to judge interview quality
To judge whether an AI interview study was done well, look at both the interviews themselves and the final findings. Use the checklist below item by item during review.
- The research plan states the business question, the decision to be made, and the research questions
- Every topic in the interview guide has a research goal, and there are no leading questions
- The introduction stated that the interviewer is AI, the expected length, and how data will be used, and obtained the participant's consent
- Spot-check 5 transcripts: when participants gave vague answers, did the AI ask about one specific experience?
- Follow-up questions didn't put answers in the participant's mouth, such as “Was it because it's too expensive?”
- No topic was probed so many times that the participant got impatient
- Off-topic, clearly low-effort, or unqualified responses were excluded, with the reasons recorded
- Each finding states how many participants mentioned it out of the total
- Each finding links to its quotes, the participant, and where it came up in the interview
- Each finding lists counterexamples or dissenting views
- Findings separate what participants said from what researchers inferred
- The research lead reviewed every finding and recorded reasons for rejected ones
- Each finding maps to a specific decision or owning team
Limitations and risks
Informed participants
Participants have the right to know they're talking to AI. The introduction should say so clearly and explain how the transcript is stored, who can see it, whether it's anonymized, and whether they can ask for it to be deleted. Hiding that the interviewer is AI might get more people to finish, but it erodes trust and may violate local regulations.
Where AI probing should stop
AI is good at turning vague answers into specific ones, but it has blind spots: it may miss hesitation in what someone says, or dig too deep into something mentioned in passing. Your interview guide should spell out what not to probe, such as income, health, and other people's private information. When a participant says they'd rather not talk about something, the AI should move on to the next topic right away.
Biased findings
When AI synthesizes findings, it can overgeneralize, presenting what two or three people said as a widespread pattern, or miss views that come from only a few people but matter a lot. The fix: always label participant counts and counterexamples, have the research lead review every finding, and check important findings against the original quotes.
When you need a human moderator
- The topic is sensitive, and participants may become visibly emotional
- You need to observe participants' actual behavior or usage environment
- Participants are a small number of key customers or senior decision-makers
- Very early-stage research where even the interview guide is hard to pin down and the researcher needs to adjust direction during the interviews
- Studies that need a long-term relationship with participants, such as kickoff interviews for co-creation or diary studies
Common mistakes
- Turning survey questions straight into a conversation: the AI will just ask them one by one and won't uncover reasons. Organize the guide by topic and research goal, and leave room for follow-up questions.
- Starting interviews before defining the decision: you end up with lots of findings you can't use. First write down which decision the study is meant to support.
- Not telling participants they're talking to AI: this damages trust and creates compliance risk. State in the introduction that the interviewer is AI and how the data will be used.
- Reading only the report summary, not the raw transcripts: overgeneralized findings can easily mislead you. Check important findings against the original quotes at least once.
- Extrapolating proportions from interview counts: interview samples aren't random. When you need proportions, validate with a survey.
How to do it in heari
In heari, you enter a business question, and AI generates a research plan and interview guide that the research lead reviews before publishing. AI runs many interviews at once by text or voice and asks follow-up questions based on the answers, while researchers can listen in and exclude invalid responses. Every finding in the report comes with quotes and counterexamples and is labeled with participant counts, and you can accept, edit, or reject each one. Connected to ThinkingAI Agentic Engine, heari also lets you select participants by behavior and validate findings with data.
FAQ
How are AI-moderated interviews different from online surveys?
Online surveys have fixed questions: everyone answers the same ones, which makes them good for measuring proportions. AI-moderated interviews follow up on each participant's answers, so they surface specific experiences and the reasons behind them. The two often work together: use interviews to find the reasons, then a survey to estimate how common each reason is.
How many participants does an AI interview study need?
Sample size depends on your research question and how much your participants differ. For exploratory research, you can refer to Guest, Bunce & Johnson (2006): in a homogeneous group, few new themes appear after about 12 interviews. If your study covers several clearly different segments, plan enough sessions for each one.
Do participants need to know they're talking to AI?
Yes. At the start of the interview, tell participants it's moderated by AI, and explain how the transcript is stored, how it will be used, and how to have it deleted. Begin only after they agree. It's a matter of respect for participants, and in many regions it's also required by data protection law.
Can AI interviews fully replace user researchers?
No. AI can moderate interviews and produce a first pass at synthesis, but defining the research question, reviewing the interview guide, checking the sample, and reviewing findings still fall to the research lead. Sensitive topics, key accounts, and studies that require in-person observation should still be moderated by a researcher.
How do you know the findings AI synthesizes are trustworthy?
Check each finding for a participant count, traceable quotes, and listed counterexamples, then spot-check a few raw transcripts to see whether follow-up questions were specific and free of leading. For findings that will drive major decisions, verify them line by line against the original quotes.
