# AI Focus Groups and AI-Moderated Interviews: The Complete Guide

> How AI-moderated interviews and AI focus groups work, what the early evidence shows, where AI moderators fall short, and when to use a human moderator.

Source: https://www.pulselake.co/blog/ai-focus-groups-and-ai-moderated-interviews-guide-2026
By PulseLake Team · PulseLake

An AI-moderated interview is a one-to-one conversation where an AI asks the questions and picks the follow-ups. An "AI focus group" usually means that same interview run with many people, one at a time, or else a panel of AI-generated personas. The first uses real participants and has early published support for short interviews. The second simulates people and works best as a drafting aid. For sensitive, exploratory or group-interaction work, a human moderator is still the better choice.

We make PulseLake, a research platform. It comes up once in this guide, in the section on choosing a method.

![Chat between an AI moderator and a participant with a follow-up probe and a human researcher reviewing.](https://www.pulselake.co/blog/img/production/08bc15220945fe976e4595158ce235633ad0cb4d-1200x840.png?w=1600&fit=max&auto=format)

## What do "AI focus group" and "AI-moderated interview" actually mean?

Vendors use these terms loosely, and they stretch them over three different methods. Pull the three apart and most of the confusion goes away.

| Method | Who answers | Who moderates | Evidence base |
| --- | --- | --- | --- |
| AI-moderated interview | Real people, one at a time | An AI | Early academic studies on short interviews |
| "AI focus group" (real participants) | Real people, often hundreds | An AI, in separate conversations | Same as above; no group interaction |
| Synthetic focus group | AI-generated personas | A person or an AI | Mixed; see the section on synthetic groups |

A traditional focus group is a moderated discussion in a small group, typically about 6 to 10 people, where participants react to each other ([Simply Psychology](https://www.simplypsychology.org/what-is-a-focus-group.html)). That back and forth is the whole reason the method exists. Most products sold as AI focus groups with real participants actually interview people individually. You get the breadth of a big sample. You don't get people reacting to one another live.

If you only need the basics, start with our short definition at [What is an AI-moderated interview and how does it work?](https://www.pulselake.co/answers/what-is-an-ai-moderated-interview). This guide goes further.

## How do AI-moderated interviews work?

Most tools follow roughly the same workflow:

1. Write the guide: the objective, the topics, the questions that have to be asked, and the rules for probing.
1. Recruit real participants from a panel, a customer list or a link, and tell them an AI will be interviewing them.
1. Let the AI run each session, by text or voice. It asks one question at a time and decides on a follow-up from what the person just said, and sessions can run in parallel.
1. Code the transcripts into themes, usually with AI help, while a researcher checks a sample by hand.
1. Have people interpret the findings and decide what to do about them.

Great Question describes a similar sequence: write questions, recruit, let the AI run and follow up in real time, then analyze the transcripts ([Great Question](https://www.greatquestion.com/ux-research/ai-moderated-interviews-guide)).

### What is conversational AI in market research?

It's software that can hold an open-ended exchange in natural language. A chatbot serving fixed answers doesn't count, and neither does a survey with fixed questions. In research, conversational AI does two jobs. It can act as a virtual moderator that interviews participants, and it can be an assistant that helps researchers write guides and code transcripts. This article is about the first job. For the second, see our pieces on [AI-first qualitative analysis](https://www.pulselake.co/blog/ai-first-qualitative-analysis) and [qualitative coding at scale](https://www.pulselake.co/blog/qualitative-coding-at-scale).

### How should you write prompts for a virtual moderator?

Quality gets won or lost in the guide. Be specific. Ask one question at a time, and tell the AI when to probe and when to move on. The block below is a generic illustration of the kind of rules a guide might contain. It isn't a template from any particular product.

```text
Objective: understand why trial users stop using the product in week two.
Rules:
- Ask one open question at a time. Never suggest an answer.
- If an answer is abstract, ask for a specific recent example.
- If a point is unclear, ask the person to explain in their own words.
- Do not give your own opinion or praise the product.
- Cover these topics in order: first use, a moment of friction, alternatives tried.
- If the person seems upset or asks to stop, end politely.
```

The Geiecke and Jaravel interviewer, which we discuss below, works on a similar logic: non-directive questions, follow-ups to clarify, and requests for concrete evidence over abstract opinion.

## How good are AI moderators compared with human ones?

The best published evidence we found is a single academic study, and it has clear limits. Geiecke and Jaravel of the London School of Economics built an open-source AI interviewer and had sociology PhD students grade its transcripts. Twenty transcripts, each graded twice, were scored on a 1 to 5 scale where 3 meant an average human expert working in an online text chat. The average grade was 2.95 ([Geiecke and Jaravel, 2025](https://conference.nber.org/conf_papers/f221279.pdf)). In a separate comparison, eight sociology PhD students who weren't told which modality each transcript came from ranked an AI voice interviewer at 3.9, human face-to-face experts at 3.6 and an AI text interviewer at 3.2, with an online text interview as the benchmark. Against a face-to-face benchmark, human experts came out highest at 3.6, the AI voice interviewer at 3.4 and AI text at 2.8 (same paper).

Read those numbers slowly. In the first exercise, graders compared transcripts with a hypothetical expert, not with named human moderators covering the same topics. In the second, the human interviewers were four PhD students or recent graduates. Every interview was one-to-one, and the sample is small. The authors also report that graders disagreed with each other noticeably: two graders scoring the same transcript correlated at 0.41.

So what does it support? Something like "AI interviews can reach a reasonable standard in short, structured settings." It doesn't support "AI matches human moderators."

A second study, by Wuttke and colleagues, randomly assigned a small group of university students to an AI or a human interviewer using the same political questionnaire. It reports that AI conversational interviewing can produce data of comparable quality to traditional methods. The abstract itself describes the study as small and run in a controlled setting ([Wuttke et al., arXiv](https://arxiv.org/abs/2410.01824)).

## What are the limits of AI moderation?

Several practitioner sources land on the same weaknesses. Keep in mind that these are vendor and practitioner views, not controlled studies. Treat them as hypotheses to test in your own pilot.

- Unspoken cues. CleverX says text and voice AI can't observe facial expressions, posture or hesitation ([CleverX](https://cleverx.com/blog/what-ai-moderators-cannot-do-limitations-and-risks)), and Great Question also rates AI as limited at picking up unspoken signals.
- Unexpected tangents. Both CleverX and Typeform say an AI tends to stick to its predefined path and may miss a theme the guide didn't anticipate ([Typeform](https://www.typeform.com/blog/ai-moderated-interviews-how-they-work-where-they-fail)).
- Sensitive topics. CleverX says an AI can't recognise distress or decide to pause or end a session, which is a real problem for health, money or bereavement.
- Context such as sarcasm. Typeform says AI may miss irony and cultural nuance, and could log a sarcastic remark as positive.
- Knowing when to stop. Typeform says an AI has no sense of whether it has reached saturation or is just collecting redundant data, which can inflate the number of interviews.
- Fraud and quality. Great Question states that AI moderation "does not guarantee trustworthy participants", so recruitment and screening still matter.

CleverX also points out a trade-off. Identical probing removes moderator inconsistency, but it can flatten responses when the script is too rigid. A human moderator brings judgment and rapport. An AI brings consistency and scale.

## When should you use an AI moderator, a human one, or both?

Great Question and CleverX suggest similar splits. The table is our own reading of them:

| Situation | Our read |
| --- | --- |
| Testing a specific hypothesis with a defined audience | AI moderation is a reasonable fit |
| Discovery at volume on low-sensitivity topics | AI moderation, then human review of themes |
| Follow-ups after a survey or after human sessions | AI moderation |
| Open-ended exploration where you do not know the questions | Human moderator |
| Emotionally sensitive topics | Human moderator |
| Participants who need accommodation (accents, low reading confidence) | Human moderator |
| Live group reactions and debate | Human-moderated focus group |

Here's a hypothetical to show the scale argument. A team wants 200 interviews of 15 minutes each. That's 200 x 15 = 3,000 minutes, or 50 hours of conversation. One human moderator would need 50 hours of live sessions, plus all the scheduling. An AI can run many at once, which is why the approach suits breadth.

What the arithmetic can't tell you is whether those 200 conversations go as deep as 12 human-led ones. We like a combined design for exactly that reason: a few human-led sessions to find the themes, then AI interviews to see how widely they hold.

## Can AI moderate a real group discussion?

We couldn't find a published evaluation of an AI moderating live interaction among several participants. The LSE and Wuttke studies are both one-to-one.

Group moderation is a different skill. The moderator has to make sure everyone gets a turn, draw out views that differ, and notice when a dominant voice or groupthink is taking over ([Simply Psychology](https://www.simplypsychology.org/what-is-a-focus-group.html)). Until there's evidence an AI can do that reliably, we'd use a human moderator whenever group interaction is the reason you picked a focus group in the first place.

Want some of the benefits without the scheduling headache? One option is a set of individual AI-led interviews, followed by a human-led group session to test the main reactions. For a broader look at the method, see [Qualitative vs quantitative research: beyond the basics](https://www.pulselake.co/blog/qualitative-vs-quantitative-beyond-the-basics).

## What are synthetic focus groups, and how do they differ?

A synthetic focus group swaps the participants for AI-generated personas. That makes it a different method, with different risks.

The Decision Lab notes that systematic errors cluster around lived experience, that simulated respondents can fill gaps in thin data with stereotypes, and that building reliable agents may itself require interviewing a real panel. Its recommendation is to use synthetic groups as a supplement that sharpens questions before real research, and then to validate with actual people ([The Decision Lab](https://thedecisionlab.com/insights/technology/focus-simulated-the-promise-and-limitations-of-ai-market-research)). We agree with that ordering: synthetic for speed and early direction, real people for validation.

One result gets cited a lot. In a study of 1,052 Americans, Park and colleagues built agents from two-hour interviews. On held-out General Social Survey items, those agents reached 83% of participants' own two-week test-retest consistency, and 86% when combined with survey data ([Park et al., arXiv](https://arxiv.org/abs/2411.10109)). Notice what that measures: agreement on survey answers about known people. It doesn't show that a synthetic panel can predict how your market will react to a new product. Our articles on [synthetic respondents and their limits](https://www.pulselake.co/blog/synthetic-respondents-and-their-limits) and [how accurate synthetic respondents are](https://www.pulselake.co/answers/how-accurate-are-synthetic-respondents) go into more detail.

## How do you run an AI interview study step by step?

1. Pick the question, and write down the decision the study will inform. See [designing research that produces actionable answers](https://www.pulselake.co/blog/designing-research-that-produces-actionable-answers).
1. Choose the method using the table above.
1. Write the guide and test it on yourself and two colleagues. Look for leading wording with the help of [writing questions that don't bias responses](https://www.pulselake.co/blog/writing-questions-that-dont-bias-responses).
1. Disclose the AI. Tell participants they're talking to an AI and how their data is processed. In the EU, the European Commission says Article 50 of the AI Act requires that people be told they are interacting with an AI system ([European Commission](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act)).
1. Pilot with a small group, read every transcript, and fix the guide.
1. Screen for quality by checking for empty, duplicated or off-topic answers; see [detecting low-quality survey responses](https://www.pulselake.co/blog/detecting-low-quality-survey-responses).
1. Code and review. Let AI propose themes, but have a person verify a sample and keep the verbatim quotes attached.
1. Pair the study with a quantitative check when the decision needs numbers, as in [concept testing](https://www.pulselake.co/blog/concept-testing-guide-methods-questionnaire-examples-2026).

## Where does PulseLake fit?

PulseLake's AI-led interview mode uses conversational and avatar-based AI interviewers. They follow a protocol, probe adaptively and interview real participants, and each conversation becomes structured evidence. There's also a separate synthetic research mode. We don't claim our interviewers moderate live multi-person groups, and every limit above applies to any AI moderator, ours included.

## Frequently asked questions

### Can AI replace a focus group moderator?

For a one-to-one interview on a structured, low-sensitivity topic, early evidence says an AI moderator can reach a reasonable standard. Live group discussion is another matter. For that, and for sensitive topics or open exploration, we'd still use a human, since we found no published evidence for AI in the group role.

### What is an AI-moderated interview?

It's a research interview with a real participant where an AI asks the questions, follows a discussion guide and probes the answers. Our [short answer](https://www.pulselake.co/answers/what-is-an-ai-moderated-interview) has the definition if you want it in a paragraph.

### How do you conduct a focus group with AI?

You've got two routes. Run individual AI-led interviews with real participants and combine the findings, or use synthetic personas as a rough pretest. Just don't treat synthetic output as evidence about real customers until you've validated it with people.

### Are AI interviews as good as human interviews?

Close, in a narrow setting. In the one academic study we cite, sociology PhD students rated AI-led text interviews close to an average hypothetical expert in an online text chat, and in a separate comparison rated an AI voice interviewer above AI text. That covers short online interviews, not every topic or every moderator.

### Is it ethical to use AI to interview participants?

It can be. Participants need to know an AI is interviewing them, consent has to cover how their data is processed, and sensitive topics should go to human moderators. Check the rules where you operate, too.

### What are the biggest risks of AI moderation?

Missed tangents, no reading of non-verbal cues, weak handling of distress, misread sarcasm and unverified participants. A pilot plus a proper transcript review will catch most of them.

## Related reading

- [What is an AI-moderated interview and how does it work?](https://www.pulselake.co/answers/what-is-an-ai-moderated-interview)
- [Synthetic Respondents and Their Limits](https://www.pulselake.co/blog/synthetic-respondents-and-their-limits)
- [Qualitative Coding at Scale](https://www.pulselake.co/blog/qualitative-coding-at-scale)
