Semi-structured interviews: how AI changes them

Qualitative research has always interviewed too few people. Interviews are slow and expensive: they have to be scheduled, conducted, transcribed and analysed, and every step ties up people's time. In over 20 years as a consultant and entrepreneur I saw interview rounds cut from the budget many times, even when they were exactly what would have explained why people do what they do. In the end the client didn't want to pay for them, even though they mattered to the client and to the consultant or researcher.

In this piece I go through what qualitative research and semi-structured interviews are, how a semi-structured interview is run in practice, and what AI changes about it. I'll say up front that I'm biased: at Valogent we build an AI interviewer. I'll still try to be honest about where a human is better.

What is qualitative research?

Qualitative research tries to understand a phenomenon: why people choose, switch, experience and justify things the way they do. The material is usually interviews, observation or written text, not numbers.

Quantitative research tells you how many. Qualitative research tells you why.

	Qualitative research	Quantitative research
Asks	Why? How? What does it mean?	How many? How much?
Material	Interviews, observation, text	Surveys, measurements, registers
Sample	Usually small and purposive	Large and representative
Output	Themes, explanations, typologies	Distributions, differences, correlations
Strength	Depth and unexpected findings	Generalisability and comparability

Good research often uses both. The qualitative phase finds the right questions, and the quantitative phase measures how common the answers are.

What is a semi-structured interview?

A semi-structured interview is one of the most common interview formats in qualitative research. The topics are decided in advance, but the exact wording and order of the questions follow the conversation.

It sits between two extremes:

• A structured interview is essentially a spoken questionnaire: the same questions in the same order for everyone.
• An unstructured interview is an almost free conversation around a topic.
• A semi-structured interview makes sure the same themes are covered with everyone, while leaving room for follow-up questions and the participant's own meanings.

In Finland the format has its own name and tradition, teemahaastattelu (the theme interview), built on Robert Merton and colleagues' focused interview.

Follow-up questions are the heart of the method. The first answer is often the one the participant assumes is expected. The interesting material comes from "why?", "what happened next?" and "can you give an example?".

How to run a semi-structured interview, step by step

1. Define the research question. What do you want to understand, and which decision will the result inform? The research question drives everything else.
2. Write the interview guide. Break the research question into 3–6 themes and note a few prompts for each. The guide is a checklist, not a script.
3. Choose the participants. Qualitative research selects participants purposively, not at random. Think about whose experience matters for the question, and whose voice is easily left out.
4. Decide how many interviews. The usual principle is saturation: you keep interviewing until new interviews stop adding anything essential. In Guest, Bunce and Johnson's widely cited study (2006), the main themes in a homogeneous group emerged within about 12 interviews. The more varied the group, the more interviews you need. Saturation tells you when you have found the main themes. It doesn't tell you how common those themes are, or how they differ between groups. That has needed a survey, because interviews have been too expensive. AI-moderated interviews change this: with hundreds of interviews, the same material shows both the themes and how common they are.
5. Pilot the guide. One or two test interviews reveal unclear themes and leading questions.
6. Interview. Listen, give people time to think, and ask follow-up questions. Avoid leading the participant or offering your own interpretations mid-interview.
7. Transcribe. Interviews are turned into text for analysis. Transcribing one hour by hand easily takes several hours.
8. Analyse. Common approaches are thematic analysis and content analysis: code the material, group the codes into themes and compare the themes across participants.
9. Report and assess trustworthiness. Explain how the material was collected and analysed, and back your interpretations with direct quotes.

Where the traditional semi-structured interview struggles

The method itself works. The problem is its cost and speed.

• An interviewer can only be in one conversation at a time.
• Scheduling, transcription and coding take weeks.
• The budget often stretches to 10–20 interviews, even when hundreds of people have a view that matters.
• In multilingual studies every language means another interviewer or another translation pass.

So many useful questions never get asked: nobody can afford them.

What an AI interviewer changes

In an AI-moderated interview (AIMI), an AI moderator takes the place of the human interviewer. At Valogent it is a voice AI interview: the AI moderator holds the conversation by voice, from the researcher's interview guide. A good AI moderator understands the objective of the study and decides after every answer whether a follow-up question is worth asking.

An AI-moderated interview is different from synthetic respondents. In an AI-moderated interview the participants are real people. With synthetic respondents, the AI generates the answers too.

In practice that changes three things:

• Scale. Interviews run in parallel by the dozen or the hundred, so instead of 12 interviews you can hear from everyone the question concerns. A larger dataset is worth it: you can compare groups such as teams, countries or customer segments, find segments nobody knew to look for, and catch views that are rare but important.
• Consistency. Every interview follows the same probing logic, which makes comparison easier.
• Time. Transcription and a first analysis happen automatically. Weeks of work compress into days.

The first question I always get is: do people actually want to talk to an AI? Together with Norstat, we ran a study for the Pretty Boy Chicken restaurant chain. Norstat recruited the respondents from its panel, and Valogent's AI interviewer held the conversations by voice.

• 30 interviews completed in two days
• 15–20 minutes each
• A respondent experience score of 4.9 / 5, measured by Norstat with its own panel members
• 68% of the material came from the AI's own follow-up questions

The last number is the one I'd focus on. Two thirds of what we learned didn't come from the interview guide. It came from the interviewer noticing something and asking "why?".

I see the same in our own studies. Many participants have tried talking to an AI before, and it wasn't a good experience, so they come in with a lot of preconceptions. When they then talk to Valogent, they are almost always surprised by how good the experience is. One participant even said the interview felt therapeutic. Norstat's independently measured respondent score shows the same.

University research points the same way. At Curtin University in Australia, Billy Sung and Patrick Duong ran a randomised experiment comparing interviews by an AI and by a trained human moderator, using the same interview guide. According to a summary of the study, there was no significant difference in trust or disclosure. The human was still better at emotional presence.

Is it still qualitative research?

This is a live debate in the industry, and experienced qualitative researchers make good points. An experienced researcher reads the room and builds trust over an hour-long depth interview, and AI doesn't replace that. So in the traditional sense, AI-moderated interviewing isn't qualitative research, and I'm not going to argue otherwise.

I see it as a third method, sitting between qualitative and quantitative. It has the structure and follow-up questions of a semi-structured interview, with sample sizes that used to be possible only with surveys. Interviewing hundreds of people simply wasn't possible before, so the industry doesn't even have a settled name for it yet: AI-moderated research, qual at scale and conversational surveys are all in use.

What AI doesn't change

AI takes over the craft of data collection: interviewing, transcribing and coding. The most valuable part of research stays with people.

• The right question. Writing the research question and the interview guide takes an understanding of the business or the phenomenon.
• Interpretation. Themes and quotes aren't an answer yet. Someone has to tell a real signal from a confident-sounding summary.
• Ethics and privacy. Participants need to know who is interviewing them, what the material is used for and where it is stored.

When to use AI, and when a human

An AI interviewer fits best when:

• you need 30, 100 or 500 interviews, not eight
• every conversation must follow the same probing logic
• you want to understand reasoning: why people choose, switch and justify
• the timeline or budget is tight
• the study runs in several languages

Keep a human interviewer when:

• the emotional connection with the interviewer is itself part of what you study
• the topic is especially sensitive or the participants are vulnerable
• there are only a few interviews and each one is unique, such as executive depth interviews

Often the best answer is a blend: the AI hears everyone, and a human goes deeper where it matters.

In practice, AI interviews are used for consultants' stakeholder interviews, companies' employee and culture research, and research agencies' customer and brand studies.

If you're writing a thesis

If you're using semi-structured interviews for a thesis, always agree on the use of AI with your supervisor and check your institution's guidelines. Research ethics apply to AI-collected material too: participants must give informed consent. The EU AI Act also requires telling people when they are talking to an AI.

Frequently asked questions

What is the difference between a semi-structured and an unstructured interview? A semi-structured interview covers the same themes with everyone and adapts the questions. An unstructured interview is an open conversation with no fixed themes.

How many interviews does qualitative research need? There is no single right number. In a homogeneous group the main themes often repeat after about 10–15 interviews. If you want to compare groups, such as teams, countries or customer segments, you need interviews from each group.

Can AI run a semi-structured interview? An AI interviewer conducts the interview by voice from an interview guide and asks its own follow-up questions, so the structure is the same. It doesn't replace an experienced researcher in a depth interview. A person still writes the guide and interprets the results.

Do people open up to an AI as much as to a human? In a randomised experiment at Curtin University there was no significant difference in trust or disclosure. The human was still better at emotional presence.

How is AI interview material analysed? Interviews are transcribed automatically, and the AI proposes themes that trace back to the original quotes. The researcher checks and approves the interpretations.

Does it work in languages other than English? Yes. Valogent's AI interviewer works in 70 languages, and the same interview guide runs in all of them.

Where is the interview data stored? Valogent is built in Finland for EU customers. Data stays under EU data protection, and interviews are never used to train models.

---

What's the one question you've never been able to afford to ask? Tell me. We're collecting them.