For years, user researchers lived with a tradeoff that had no good answer. Interviews gave you depth, the real reasons behind what people do, but they were slow and expensive, so you ran a handful and hoped they were representative. Surveys gave you scale, but they flattened every answer into a checkbox and never asked the one follow-up question that mattered.
AI-moderated interviews are the first method that closes most of that gap. They pair the depth of a one-on-one conversation with the reach of a survey, and they have gone from novelty to standard practice fast. According to Maze's Future of User Research 2026, 69% of researchers now use AI in at least some projects, a 19-point jump in a single year. This guide walks through what AI-moderated interviews actually are, how they work, what the evidence says about their quality, and the situations where you should reach for one (and the situations where you should not).
What is an AI-moderated interview?
An AI-moderated interview is a live, one-on-one research session where an AI agent runs the conversation in place of a human researcher. The AI opens the session, works through a question outline you define, follows up on what the participant says, and closes with notes you can synthesize right away.
The key word is moderated. The participant is a real person sharing real experience, and only the interviewer is automated. That distinction matters more than it sounds, and we will come back to it, because it is the single most common point of confusion in this space.
A session usually runs over voice or text in the participant's browser. They join on their own schedule, talk through your questions at their own pace, and the AI adapts its questions to each answer rather than reading from a fixed script.
How AI-moderated interviews work
Under the surface, four things happen in sequence.
You define the outline, not the script
You give the AI a research objective and a set of questions or topics, the same brief you would hand a junior moderator. You are setting the destination and the guardrails, and from there the AI decides how to get through your topics based on what each participant says.
The AI probes in real time
This is the part that separates an AI-moderated interview from a survey with a chat skin. As each answer comes in, the AI weighs it for depth. A vague response gets a clarifying question. A rich one lets the AI move on. Good systems will follow an interesting thread several layers deep, the same laddering a trained interviewer uses to get from "the checkout was annoying" to the actual reason someone abandoned a cart. Industry analyses put this depth at five to seven follow-up levels per topic, which is in the range of a skilled human moderator.
It runs over voice or text
Modern voice synthesis has reached the point where pacing and inflection sound natural, so participants are not fighting an awkward robot to be understood. Text-based sessions work well for quick, structured feedback. Voice sessions tend to pull longer, more considered answers.
It hands back synthesis-ready output
When the session ends, you do not get a raw recording to transcribe and tag by hand. You get a structured transcript, and on most platforms an automatic first pass at themes. Analysis that used to eat two days of tagging can be reviewed in an afternoon.
AI-moderated interviews are not synthetic users
This is the distinction that trips people up, and getting it wrong leads to bad decisions.
Synthetic users are AI-generated personas with no real person involved. A language model invents a plausible-sounding customer and answers your questions as that invented character. With AI-moderated interviews, the moderator is AI but the participant is a verified human being.
The two differ wildly in reliability. Nielsen Norman Group tested synthetic users against real studies and concluded the synthetic responses were too shallow to be useful, and a January 2026 issue of ACM Interactions argued that AI-generated participants undermine the point of research altogether. Synthetic respondents carry built-in defects: they reflect training data from years ago rather than your current customer, they are tuned to agree with the framing of your question, and they cannot surprise you the way a real person can.
AI-moderated interviews sidestep all of that, because the data still comes from real lived experience. You get the scale that made synthetic users tempting without giving up the truth that makes research worth doing. So if a vendor blurs these two ideas together, slow down and read carefully.
What the evidence actually shows
A little skepticism about whether an AI can run a decent interview is healthy, so look at the numbers.
One comparative study found AI-moderated interviews produced 129% more words per response than a static survey, with 66% of transcripts rated higher quality. Completion rates were 61% for the AI interviews against 39% for the surveys once you adjusted for meaningful interactions, and the rate of low-effort "gibberish" answers was 26% for AI interviews compared with 56% for surveys. People give the AI more, and they give it better.
The cost and speed gap is just as stark. A human-moderated study typically runs $150 to $300 per 30-minute session once you count recruiting, incentives, moderator time, and analysis, which lands a 100-interview study somewhere around $15,000 to $30,000 and several weeks of calendar time. AI-moderated sessions run closer to $5 to $20 each, and a study can go from launch on a Friday to 50 synthesized interviews on Monday morning. Across teams that made the switch, the average cost per insight fell by roughly 60%.
None of this means more is always better. It means the old reason to run only five interviews, that each one is slow and costly, no longer holds.
When to use AI-moderated interviews
Reach for an AI moderator when the work rewards consistency, scale, or speed.
- Continuous discovery. If you want research running in the background all the time rather than in occasional bursts, AI moderation is what makes that affordable. You can keep a study live and collect interviews as users hit a moment that matters.
- Product and feature feedback. Beta feedback, post-launch reactions, and concept tests are well-defined enough that an AI can probe them reliably.
- Larger, more representative samples. When you need 50 or 200 conversations instead of 8, the economics finally work. A bigger sample also means your themes rest on more than a few loud voices.
- Multilingual research. An AI moderator can interview in many languages without translators or bilingual staff, which widens who you can actually hear from.
- Screening and recruitment. Use a short AI conversation to confirm eligibility before a person ever reaches a human session.
- Consistency across sessions. Human moderators fatigue after four to six interviews in a day and start leading the witness or chasing pet theories. Every participant in an AI-moderated study gets the same tone, pacing, and follow-up logic, which removes a quiet source of bias.
When not to use them
An honest case for AI moderation has to include where it falls short. Nielsen Norman Group's 2026 evaluation, run with ten research leaders across eight countries, is blunt about the limits, and so are we.
- Sensitive or emotional topics. Studies that touch trauma, health crises, discrimination, or identity need human empathy. NN/G found participants held back sensitive information with an AI interviewer because the trust was not there. Do not automate those conversations.
- Brand-new problem spaces. When you do not yet know what to ask, you need a human who can abandon the guide and chase a surprise. AI moderators follow your outline well but cannot reinvent it mid-session.
- Complex observation. Anything that depends on screen-sharing, watching someone struggle through a task, or reading body language is still human work.
- Specialized domain depth. Highly technical interviews that hinge on industry jargon and expert judgment can outrun an AI moderator's understanding.
The pattern is straightforward. AI moderation is strong when your questions are reasonably well-defined and your goal is breadth, consistency, or speed. Human moderation wins when the work demands empathy, real-time reinvention, or deep domain expertise. Most mature teams use both, which is the subject of our companion guide on choosing between AI and human moderation.
How to run a good AI-moderated interview
The method is only as good as the brief you give it.
- Write a tight objective. One or two sentences on what a successful study tells you. The AI uses this to decide which threads are worth chasing.
- Ask open questions. "Walk me through the last time you..." beats "Did you like the feature?" The point of the method is the follow-up, so give it room.
- Cap the scope. Five to eight core questions is plenty. Long guides produce rushed answers, the same as with human interviews.
- Pilot with a few participants first. Read the early transcripts, adjust the framing, then open it up.
- Read the raw transcripts, not just the summary. The auto-themes are a head start, not the finish line. The good stuff is usually a sentence someone said that no summary would surface.
For a refresher on the fundamentals that apply to any interview, automated or not, see our guide on how to conduct effective user research.
Where this fits at User Evaluation
AI-moderated interviews are built into User Evaluation, so you can run adaptive voice or text conversations with real participants, then move straight into analysis without the manual transcription and tagging step in between. You set the objective and the questions, the AI handles the moderation and the follow-ups, and your transcripts land ready to synthesize alongside the rest of your research.
Where this leaves you
AI-moderated interviews will not replace the craft of research, and they are nothing like the synthetic users that deserve all the skepticism they get. What they give you is a way to run real interviews with real people at a scale and speed that used to be out of reach. Use them where your questions are clear and you want breadth, consistency, or speed, and keep a human in the room for the sensitive, the exploratory, and the truly complex work. Run it that way and you stop choosing between depth and scale, and start getting a fair amount of both.