AI in Qualitative Research: What the Moderator Keeps

AI in qualitative research now transcribes and sorts focus groups. What the evidence says it still misses, and where the moderator's judgment decides.

A qualitative moderator writes the discussion guide with the client and runs the focus groups. Then the moderator turns hours of talk into a debrief, a topline and a final report. That last stretch, the analysis, is where AI in qualitative research has moved fastest. Transcripts come back in minutes. Software files quotes by theme. Some platforms now run the interview itself with an AI moderator. So when a moderator asks whether their job is next, it is a fair question. The honest answer is in the analysis: what a machine already does well, and what the published evidence says it still does poorly.

Key Takeaways

• In MRII's AI in Focus 2025 survey of 426 market researchers, 47 percent expressed some concern about losing their own job, up 6 points on 2024, while 71 percent expected AI to improve their work.
• When David Morgan ran two published focus group studies back through ChatGPT, it reproduced concrete, descriptive themes and was less successful at locating subtle, interpretive ones.
• In a 2025 statement in Qualitative Inquiry, 419 qualitative researchers from 32 countries rejected generative AI for reflexive analysis, which they say needs a positioned human researcher.
• Transcription, filing by guide section and descriptive coding can already be automated. Framing the reading, weighing what the room meant and signing the implication stay with the moderator.

What a qualitative moderator does besides moderating

The job title says moderator. In most market research agencies, the same person also designs the study and writes the analysis. Three terms carry most of that work.

A moderator is the researcher who leads a focus group or an in-depth interview. They follow a discussion guide and probe when an answer needs more depth.

A discussion guide is the document that sets the topics, their order and the stimuli for each session. It is agreed with the client before fieldwork.

A debrief is the first reading of the sessions. It usually goes to the client right after fieldwork, before the full analysis.

PERSONAL EXPERIENCE In twenty years of market research operations, I rarely saw a moderator stop when the session ended. They watched the client behind the mirror and took notes between groups. They changed the guide for the next session. They carried the whole project in their head until the report went out. The moderation is the hour people see. The analysis is the week nobody sees.

Greenbook's 2025 GRIT report finds the same pattern in its data. Suppliers who live mostly from focus groups and in-depth interviews seem to be differentiating on their core skills, interviewing and moderating. The tools they use to run the sessions matter less.

How focus group analysis is done today

Most teams still run the analysis by hand, in a spreadsheet, in the days between the last session and the debrief.

PERSONAL EXPERIENCE The recordings go out for transcription. The moderator reads every transcript, often twice. The moderator builds an analysis grid in Excel, with one row per participant and one column per topic in the guide. Then the quotes get pasted into the cells. The slow part comes next. The moderator reads the grid across and groups what repeats. One participant said something nobody else did, and that gets marked too. Then it is back to the video to check a tone of voice. The topline comes out of that grid, and the deck is built slide by slide, quote by quote. A chatbot open in another tab speeds up a summary here and there, with the limits we listed in why a general chatbot is not a research platform. The grid is still built by hand.

The cost is time, and it lands at the worst moment. Fieldwork ran late. The debrief date did not move. Every hour spent sorting quotes is an hour taken from thinking about them. In front of the client, the moderator answers for it. Their name is on the reading. We covered the coding side of this problem in qualitative coding with AI without losing the evidence.

Why the question about the job is fair

The work around the moderator is changing, and the profession's own surveys show that people see it.

An AI agent is software that carries out a series of steps toward a goal within instructions a person wrote. In an interview, the steps are asking a question, reading the answer and choosing the next question.

An AI-moderated interview is a qualitative session in which an AI agent asks the questions and records the answers, with no human moderator in the conversation. It follows probing instructions set in advance. Discuss, which sells the method, describes an AI moderator that asks exactly what it was told to ask. It probes for exactly what it was told to probe for.

MRII's AI in Focus 2025 states its methodology plainly: 426 online interviews with market researchers, conducted in February 2025. In it, 47 percent expressed some level of concern about losing their own job, up 6 points on the year before. Job cuts from automation worried 46 percent, up 8 points. That was the largest rise of any item in the study. In the same sample, 71 percent expected AI to improve their job, and 63 percent were concerned about depending on AI at the expense of human judgment.

Researchers expect help from AI. They also watch their jobs.

A bar chart of four results from MRII, AI in Focus 2025, an online survey of 426 market researchers worldwide conducted in February 2025 in partnership with QuestionPro. Unit: percent of respondents. 71 percent believe AI will improve their job in the coming years. 63 percent are concerned about increased dependence on AI at the expense of human judgment. 47 percent express some level of concern about personal job loss, up 6 points on 2024. 46 percent see potential job reductions due to automation as a concern, up 8 points on 2024, the largest increase of any item. A closing line says hope and worry sit in the same sample, and the worry is about judgment as much as jobs. Every number belongs to MRII and none of them is a Cassi.ai measurement.

The same researchers who expect AI to help are the ones counting the risk to their jobs and to human judgment.

Greenbook's 2025 GRIT Insights Practice Report adds the workforce side. It is based on 992 completed surveys from the first quarter of 2025. It says 50 percent of professionals are proactively reskilling for AI integration, even as attrition and hiring freezes strain traditional roles.

Neither study can tell one moderator what will happen to their job. Nobody writing about it can either. What the evidence can do is show which parts of the analysis still need the moderator. That is a more useful answer than a promise.

What the evidence says AI does well in qualitative analysis

AI handles the descriptive part of qualitative analysis reasonably well. It handles the interpretive part less well. The same split shows up in experiments and in the objections of method experts.

David L. Morgan's research method was direct. He used ChatGPT to re-analyze two qualitative datasets that had already been analyzed by hand. Then he compared its themes with the original ones, in a study in the International Journal of Qualitative Methods. ChatGPT performed reasonably well. In both cases, though, it was better at reproducing concrete, descriptive themes and worse at locating subtle, interpretive ones. Morgan concludes that coding and AI are both tools. Each has to sit inside a larger analytic process.

A descriptive theme is a pattern participants state directly, such as complaints about price.

An interpretive theme is a pattern the researcher reads underneath what was said. A product used to look competent in front of a partner is one example.

The strongest objection comes from the academic side. In a statement in Qualitative Inquiry, 419 experienced qualitative researchers from 32 countries take a firm position. Approaches such as reflexive thematic analysis, they say, are human practices. They need a subjective, positioned and reflexive researcher, and generative AI does not fit that work. They also reject it on grounds of social and environmental justice.

Reflexive thematic analysis is a method in which the researcher builds themes through repeated reading of the data. The researcher's own position is treated as part of the analysis. Virginia Braun and Victoria Clarke, who developed it, are among the authors of that statement.

Vendors draw a similar line. The same Discuss guide says AI handles volume and consistency. Humans handle the brief, the question design and the reading of what the AI surfaces. A human moderating live can also read hesitation and recover from a badly worded question.

The middle column follows Morgan's study and the Discuss guide. The split between the columns is ours.

Where the moderator's judgment stays indispensable

UNIQUE INSIGHT The judgment that matters does not wait for the end of the analysis. It happens at five points along the way. Each one needs something the transcript does not hold.

1. Set the reading frame before opening the first transcript: the decision the client has to make, the objectives, the hunches the client walked in with.
2. Add what the room showed: the pause before an answer, the participant who changed sides after someone else spoke, the stimulus that got a laugh.
3. Separate how often from how much it matters. A theme raised by many people can weigh less than one raised once by the right person.
4. Name the interpretive theme, the one participants circled without stating. That is where Morgan found the model weakest.
5. Write the implication and sign it, because the client hears it from the moderator and will hold them to it.

One analysis, two halves

Two columns describe the same qualitative analysis. The left column is the part automation can carry, in five steps: transcribe every session, so audio and video become text; file each passage under the guide section it answers; tag the descriptive codes participants stated directly; pull candidate quotes with the code of who said them; and draft the report structure, sections and slots but not conclusions. An arrow hands the organized material to the right column, the part the moderator decides, in five steps: set the reading frame from the decision the client has to make; add what the room showed, such as pauses, laughs and positions that shifted; weigh frequency against weight, because said often is not the same as important; name the interpretive themes participants meant without saying; and write the implication and sign it, because the client hears it from the moderator. A closing band says the organizing can move to software and the interpretation stays with the person who was in the room. The figure contains no measured numbers.

The left half saves the hours. The right half is the analysis the client pays for.

None of the five is a typing task. They are why a client hires a qualitative researcher instead of a transcription service. They are also what makes the debrief worth attending.

Where Cassi.ai comes in

@Cassi.ai is a software engineering company specialized in the pains of market research, innovation and insights. It was founded by Cassiano Albuquerque after twenty years in market research. We automate research projects with or without language models.

QualiLab is a Cassi.ai product that automates qualitative analysis: focus groups and interviews go in, and the report and the deck come out with a verbatim behind every finding.

QualiLab is the Cassi.ai product for the analysis half of a qualitative study. The client sends focus groups and interviews as video, audio or transcript. The project KPIs and the discussion guide go in with them, and the transcription happens inside the platform. Then 19 specialist agents, one per qualitative analysis method, read the whole project. Thematic, emotion, tensions and contradictions are four of them. Every finding follows the same chain, from the verbatim to the reading to the implication. The moderator can see where each claim came from.

A verbatim is a participant's exact words, quoted with the code of the person who said them.

The group is still run by a person, and the final analysis is still signed by a person. What the tool takes is the sorting, the prompt started from zero every time, and the work of building the structure. What it gives back is time for the part of the analysis that needs repertoire. The same line holds in online communities. There, a moderator agent runs the mechanics and the human moderator keeps every call that needs judgment.

FAQ

Will AI replace qualitative moderators?

No study can answer that for one person. MRII's 2025 survey found 47 percent of market researchers with some concern about their own job, and 71 percent expecting AI to improve it. The evidence on analysis points to interpretation as the part that still needs a human researcher.

Can AI do thematic analysis of focus groups?

For descriptive themes, reasonably well, according to Morgan's 2023 study. It was less successful with subtle, interpretive themes. Reflexive approaches reject it on grounds of method.

What is an AI-moderated interview?

A qualitative session in which an AI agent asks the questions and follows probing instructions written in advance. No human moderator takes part in the conversation.

Which parts of qualitative analysis can be automated today?

Transcription, filing passages by guide section, descriptive coding, pulling candidate quotes and drafting the report structure.

Who is responsible for the interpretation when AI is used?

The researcher who signs the report. The tool can draft. The person who presents the findings to the client answers for them.

The part of the analysis with the moderator's name on it

UNIQUE INSIGHT Automation is taking the sorting out of qualitative analysis. The interpretation was never in the sorting. It sits with the person who was in the room. That person knows why the group went quiet after the second concept. That person will stand in front of the client with the findings.

If your team still builds the analysis grid by hand after every round of groups, count the hours between the last session and the debrief. Then count how many of them went to thinking.

About the author

Cassiano Albuquerque is the founder of @Cassi.ai. He spent twenty years in market research operations before building software for research teams. To talk about qualitative analysis for your studies, contact him on LinkedIn.

Published by Cassi.ai. Read the full article at https://www.cassiai.com/blog/ai-qualitative-research-moderator.