Running AI-Moderated Interviews: A Step-by-Step Tutorial

Image: "Oral History of Egypt's Education 2.0 Reform-Nelly ElZayat, Deena Bora" by Antoineberlin (CC BY-SA 4.0) via Wikimedia Commons, recoloured.

AI-moderated interviews use a conversational agent to ask questions, listen to answers and probe with follow-ups, either by text or by voice. They do not replace the depth of a skilled human interviewer, but they make it possible to collect hundreds of semi-structured interviews in a week, in ten languages, at a cost closer to a survey than to fieldwork. This tutorial walks through a full study, from design to analysis, and flags where human judgement still has to sit.

Step 1: Decide whether the method fits

AI moderation works well when the topic is concrete, participants are comfortable talking to a machine, and you need breadth as much as depth: user experience studies, consumer decision journeys, employee pulse research, or exploratory work before a smaller set of human-led interviews. It works poorly for trauma, grief, illegal behaviour, or any topic where rapport and the ability to stop are essential. If in doubt, run five human interviews first and use them to build the AI guide.

Step 2: Write the interview guide for a machine

An AI moderator needs more structure than a human one. Write each section as an objective, not a script: what you want to learn, two or three opening questions, and the kinds of follow-up that are acceptable. Add explicit rules: how many probes per question, what to do if the participant gives a one-word answer, and which topics are off limits. Test the guide yourself as a participant at least three times and revise it until the conversation feels natural and the probes stay on topic.

Step 3: Get ethics and consent right

Participants must know they are talking to an AI, what will be recorded, where the data is stored, and how to withdraw. Add a plain-language statement at the start of the interview and a human contact for questions. Check that the platform you use has a data processing agreement, stores data in an acceptable jurisdiction, and does not use your transcripts for model training. Our responsible research guidelines cover disclosure and data protection in more detail.

Step 4: Pilot, then recruit

Run a pilot with ten participants from your target population. Read every transcript. Look for probes that go in circles, questions that are misunderstood, and points where participants disengage. Fix the guide, then recruit. Panels such as Prolific integrate well with interview links; for organisational studies, a personal email from a known sender still gives the best response rate.

Step 5: Analyse with the codebook, not the chatbot

The temptation is to ask a model to "find the themes". Do not. Instead, build a codebook the usual way from a sample of transcripts, then use an AI-assisted qualitative analysis tool to apply it across the full set, with the researcher reviewing every code that the model applies with low confidence. Report inter-coder reliability between the model and a human coder on a random sample of at least 10 percent. Themes should still be written by the researcher, with quotes anchored to specific participants.

Step 6: Report transparently

State the platform, the model behind the moderator, the interview guide, the number of interviews and their median length, the languages used, and how the analysis was done and checked. Journals in management, marketing and information systems now expect this level of detail for any AI-assisted qualitative study.

Tools

Qualitati, developed at HEC Paris, offers AI-moderated text and voice interviews in ten languages together with thematic analysis and a coding workspace, with a free tier for students. Other options in our human research toolkit include AILYZE and CoLoop. The QualiTaTi Academy has free guides on AI-moderated interviews, GenAI-assisted qualitative research and thematic analysis, and a free self-paced summer school on qualitative research in the AI era.