Qualitative market research and AI: Why people stay at the center
Qualitative market research remains indispensable, even with powerful AI. It uncovers the meaning, context and contradictions of real people. A language model can imitate this, but it cannot collect it.
Artificial intelligence (AI) is what we at CURTH+ROTH understand as augmented intelligence: It extends human understanding, but does not replace it.
AI tools transcribe interviews in minutes. They cluster statements and promise “synthetic respondents”. So the question is no longer whether we use AI. It is what we use it for, and where its limits lie.
Arvind Narayanan and Sayash Kapoor provide a sober framework in “AI Snake Oil” (2024). Generative AI is improving rapidly. Predicting the behaviour of individual people remains difficult. Synthetic respondents combine both: Generative AI is supposed to predict how people react.
What can AI do in qualitative market research today?
AI mainly takes over routine work. That frees up time for the actual understanding:
Preparation: structuring desk research, reviewing discussion guides, collecting hypotheses.
Transcription and translation: fast first drafts that researchers check.
Initial coding: suggestions for categories. As a starting point, not as a result.
Synthetic personas: ideas for questions and concept variants, which are then tested with real people.
Research is moving fast. Park et al. (2024) built AI agents from two-hour qualitative interviews with 1,052 people, conducted by an AI interviewer. The agents matched these people's answers in a large US survey 85% as accurately as the people themselves replicated their own answers two weeks later.
Ashokkumar, Hewitt et al. (2026, Nature) show that GPT-4 predicts the average effects of survey experiments well. However, it systematically overestimates the effect sizes.
Notably, the agents in this study are based on qualitative interviews. This suggests that good qualitative data becomes more important as models get better.
Where are the limits of AI and synthetic respondents?
A language model generates plausible language from existing texts. Qualitative research collects what is not yet written down anywhere. This leads to five limits:
No first-hand experience. AI derives what has been documented. New everyday practices are discovered in the field.
Answers that are too smooth. Synthetic answers vary less than real ones. In one study, almost half of the regression coefficients differed significantly (Bisbee et al. 2024).
Unstable results. Small prompt changes and model updates shift the results (Bisbee et al. 2024).
Distorted groups. Models misrepresent identity groups and flatten them (Santurkar et al. 2023; Wang et al. 2025).
Narrow cultural base. Models most closely resemble people from Western, educated and affluent societies. The greater the cultural distance from the US, the lower the similarity (Atari et al. 2023).
Who answers, who asks, who interprets?
“People at the center” refers to three roles. In each, AI can contribute something different.
Role | What AI can do | What remains with people |
Who answers (respondents) | Provide hypotheses, pilot discussion guides and concepts. | The experiences, contradictions and context of real people. |
Who asks (moderation) | Ask structured questions at scale. | The right moment to probe. Tone of voice, pauses, trust. |
Who interprets (analysis) | Transcription, initial coding, drafts. | Interpretation in context and responsibility for conclusions. |
An example from moderation: In a study on diabetes management systems, we often heard: “That's for younger patients.” The obvious assumption would be people under 30. We probed at the right moment. For the diabetologists in our interviews, “younger” often meant: under 65.
The difference is crucial. Diabetes is diagnosed less often in people under 30. With the wrong understanding, the market would have seemed too small, and the product would not have been developed further.
An example from interpretation: In studies on luxury products, participants often want to impress us. They test our knowledge of the category and stage their status. This behaviour is a finding in itself. It is partly lost in the transcript, and in an AI interview it might not show up at all. In one study, people disclosed more to a supposed computer because they felt less judged (Lucas et al. 2014). For sensitive topics, that is an advantage. Where self-presentation is part of the subject, some of the data may then be missing.
What does practice show?
The whiskey test: What lies between the lines
In October 2026, we had Claude analyse an older study. The AI received transcripts and animatics. The topic was whiskey and a quintessentially American motif: rodeo.
The AI correctly recognised that rodeo is an established cliché. What it did not recognise: This very cliché activated the brand core. The AI saw what was said. It missed what it meant.
This is not an isolated case. Hill et al. (2025, preprint) compared ChatGPT-5 and Claude with human analysts in a blinded setup. When assigning text to predefined codes, the models were as good as humans. When developing their own themes, they performed worse, especially where tone and context mattered.
The light nobody wanted
We tested vacuum cleaners with a light on the floor head. At the time, only a few premium models had one. Participants agreed: a gimmick, not worth paying extra for.
But their tone and emotionality did not match the rejection. Three motives lay behind it: People did not want to question their last purchase, did not know the feature from their own experience, and were worried about the price. We recommended the light. Today it is widespread.
A motorcycle has to smell of petrol
In an ethnographic study with home visits, ride-alongs and an online community, motorcyclists initially rejected electric motorcycles. Part of the rejection was justified: short ranges, high prices. Another part concerned identity: sound and smell.
We explored what riding a motorcycle means emotionally: self-image, flow, community. One finding: The feeling of sovereignty, untamed power on demand, can be experienced just as much on an electric motorcycle, if not more. We did not fight the riders' identity. We built on it.
Enduro riders appreciated the acceleration early on. Acceptance is growing from this niche. In 2025, an enduro was the best-selling electric motorcycle in Germany. Serge Moscovici described how consistent minorities change attitudes in this way.
Vacuum cleaners and motorcycles show: Rejection is not always the final word. But not every rejection is a pretext. The art lies in telling them apart. This is how we protect clients from false negative and false positive findings.
Why does good brand strategy need people?
AI works with existing data. So do people, because experience is accumulated knowledge. The difference: AI pulls towards the average.
Doshi and Hauser (2024) show this with short stories. With AI assistance, the stories became more creative individually, but more similar to each other. If everyone uses the same models, brands will converge too.
Brands that stay in memory often surprise with skilful rule-breaking. AI can also supply ideas for this. The hard part is the choice: Which rule-break can the brand carry?
On Black Friday 2011, Patagonia ran the ad “Don't Buy This Jacket”. The brand broke a rule of retail, but not the rule of its own brand. According to an IMD case study, sales then rose by around 30%. The cause, however, is not proven.
In 2009, Tropicana gave its orange juice new packaging without the familiar recognition cues. Within two months, sales fell by around 20%. The brand returned to the old design.
This choice requires trained gut feeling and the willingness to test it. Intuition is only reliable where you can learn from feedback (Kahneman & Klein 2009). In brand management, this feedback often comes late.
That is why our rule is: Gut feeling provides the hypothesis. Qualitative research shows whether it holds.
Which biases are a risk when using AI?
Biases also exist in traditional research. But AI can amplify them and make them harder to see, because its results sound so convincing.
Bias | When using AI | Remedy |
Confirmation bias | Prompts ask for what you want to hear. The model delivers matching quotes. | Specifically ask the AI for contradicting evidence. |
Framing effect | Small prompt changes shift results measurably. | Have prompts reviewed, test variants. |
Selection bias | Training data mainly reflects Western groups with a strong online presence. | Do not use synthetic data for underrepresented groups. |
Law of small numbers | Hundreds of synthetic answers look like a large sample, but are based on a single model. | Never present a synthetic “n” as a sample. |
Halo effect | A fluent AI summary seems more competent than the evidence supports. | Trace every key statement back to original quotes. |
Conclusion: What does this mean for clients?
AI makes qualitative research faster, and the models are getting better. We see this as an opportunity.
In our projects, however, we believe AI could not have delivered some of the central insights. And it is precisely these insights that matter for distinctive, successful marketing.
More suitable for synthetic data: piloting discussion guides and concepts, collecting hypotheses, pre-sorting variants.
Less suitable: questions where identity plays a role. International studies and small target groups. New categories without documented experience.
Our guidelines:
People at the center: AI delivers drafts, not findings.
Synthetic means hypothesis: We test it with real respondents.
Traceability: Every key statement can be traced back to original material.
Bias check as routine: Discussion guides, prompts and summaries are systematically reviewed.
Data protection: Interview material only goes into AI tools with consent.
Transparency: Where AI was used, we say so, including its limits.
At CURTH+ROTH, projects are in the hands of senior researchers from planning to analysis. We use AI where it makes us faster. The understanding stays with us.
How do you use AI in your market research? Get in touch: research@curth-roth.de or +49 (0) 40 3600690.
Frequently asked questions
Can synthetic respondents replace real interviews?
No. They provide hypotheses and help with piloting. They do not reliably capture the experiences, contradictions and context of real people.
What is AI suitable for in qualitative market research?
For preparation, transcription, translation and initial coding. In a recent comparative study, it was as good as humans at assigning text to known categories. When it came to discovering new meanings, it fell behind.
How can bias be prevented when using AI?
With counter-hypotheses, reviewed prompts and by tracing every key statement back to original quotes. Synthetic results are always labelled as hypotheses.
Comments