Can an AI interview your customers? Yes, it can. And they like it.
We put Stella, our AI interviewer, to the test: 98 real conversations, the CFI playbook, no human in the room. Nobody broke off. They rated it in the top fifth of the scale, felt understood, and shared more than we expected: stress, family friction, their own mistakes. Per minute, the AI found two to three times as many insights as an experienced human interviewer typically does.
Our answer: yes, it works.
Why we ran this test
Qualitative interviews are among the strongest ways to learn what customers think. But they do not scale. A typical project needs 20 to 30 interviews per customer group, which takes six to eight weeks and many hours of senior staff. Many companies skip them, and then they guess.
AI promises to change this: an agent that leads conversations on its own, with hundreds of people at once. The open question was whether people would accept it, and whether the output would be good enough.
Meet Stella. Stella is our AI interviewer. She runs a customer interview the way we would, following our Customer-Focused Innovation (CFI) interview guide, in voice or in chat, in several languages at once.
Together with moweb, which recruited the respondents from its panel, and Cartorga, whose platform ran the voice interviews, we tested her on an everyday decision: how people choose a holiday. Half the respondents spoke with Stella, half typed with her.
The study at a glance
98 leisure travellers from Germany, Austria and Switzerland
Stella by voice: 49 · Stella by chat: 49
Fieldwork 22–26 September 2026
Four languages: German / Swiss German, French, English, Italian
Job studied: "Choose a holiday trip", in 7 job steps
Do people accept an AI interviewer? Yes, clearly.
Every respondent knew from the screening that Stella is an AI. In all 98 interviews, nobody broke off and not one person refused to answer a question.
They also liked it. Voice respondents rated the interview 8.4 out of 10, and chat respondents 4.2 out of 5, both in the top fifth of their scale. Chat respondents said they felt understood, at 4.4 of 5. Asked what to improve, most chat respondents wrote "nothing", "Top Umfrage", "Alles bestens".
Even the sceptics stayed. A few respondents said up front that they do not like AI. None of them broke off, and they answered in full. People judged the conversation, not the technology.
The few low scores almost all pointed to fixable problems: audio, speech recognition for Swiss German, or a question that felt unclear. That is engineering, not rejection.
In their own words
Why people liked talking to Stella:
"Zehn, weil es einfach und unkompliziert war." Ten, because it was simple and uncomplicated.
"Neuf, parce qu'il y a juste à parler, pas besoin d'écrire." Nine, because you just talk, no need to write.
"Zehn, weil alles gut geklappt hat und du sehr sympathisch bist." Ten, because everything worked and you are very likeable.
"Je trouve que tu as eu des questions très pertinentes." I found your questions very relevant.
"Elf von zehn Punkten, weil es eine neue Erfahrung war." Eleven out of ten, because it was a new experience.
People open up to the machine
Acceptance went beyond politeness. People shared moments of stress, family friction and their own mistakes with Stella, without being asked twice. One respondent described a partner falling ill before the trip and a plan B to travel alone. Another admitted: "Ich hätte besser organisieren müssen." A chat respondent explained how autism in the family shapes every trip.
Why? Our working hypothesis: in front of a machine there is nobody to impress, so people gloss over things less. Earlier research points the same way: people disclosed more to a virtual interviewer when they believed it was fully automated (Lucas et al., 2014).
In practice: say clearly that it is an AI and that answers stay anonymous. It did not stop anyone. And invest in understanding: being understood is what earns trust.
"Interviews got cheap. The method did not."
Is the output good enough? Yes, and fast.
Each voice interview with Stella produced 8.1 Job Metrics in 7.6 minutes, about one per minute. Job Metrics are the measures customers use to judge whether a step went well, such as keeping any budget overrun on the total trip cost as small as possible. From past CFI projects, an experienced human interviewer finds about 0.3 to 0.4 per minute over a 45 to 60 minute interview.
Per interview, a human still finds more. But AI interviews do not need to be long, because they can be many: they run in parallel, at any hour, in several languages from one setup. Across all 98 interviews, Stella surfaced 683 statements, which merged into 176 distinct customer needs, each traced to a customer's own words.
The human figure is an estimate, not a controlled comparison. A direct test with the same guide is next.
Can it scale to everyone? Yes, and scale adds depth.
All 98 interviews ran within five days. And more interviews kept finding more: after 49 voice interviews, each additional one still added about 1.8 new needs. More than half of all needs came up in only a single interview.
The usual rule of 20 to 30 interviews per customer group was made for hour-long human interviews. With AI, extra interviews cost little, so the saturation curve, not a rule of thumb, can decide when to stop. That turns a handful of conversations into a whole customer base.
Voice or chat?
Both work; they serve different tasks. Both formats were accepted and rated alike. Speaking takes less effort than typing, so voice respondents said about three times as many words, and the voice interviews surfaced more distinct needs in shorter conversations. Chat keeps the question on screen and the answer exact, which makes it strong for screening and collecting figures.
The likely best of both: a hybrid where people read and hear each question and answer by talking or typing. We will test that next.
What makes it work
An AI interviewer cannot decide on its own what to listen for. It needs a behavioural model: a clear idea of what a good answer is and how answers add up to a result. Stella's model is CFI, our applied Jobs-to-be-Done method.
Why not build your own?
We tried. With AI coding tools, we built a working chat interviewer in two weeks, and it ran well as a prototype. Running it in client projects is another matter: data protection, tracing every interview, stable quality as the AI models change, and fixing errors all keep costing time. The voice service by Cartorga ran reliably from the start.
Building our own interviewer showed us where the effort lies. Our conclusion as a consultancy: we build the method, not the tool. We recommend the same to clients and to other consultants.
What this means
AI can take over the interviewing. People accept it, finish it, rate it highly and open up to it.
Hear everyone, not a handful. A whole customer base, in its own language, at any hour, in days instead of weeks.
Let saturation set the sample size. Run AI interviews until new ones stop adding new needs.
The method stays with the consultant. We design, steer and judge. Stella runs the conversations.
What we test next
Human benchmark: 10–15 human interviews with the same guide and analysis
Voice-chat hybrid: read and hear each question, answer by talking or typing
Client appetite: will corporate clients accept AI interviews with customers and employees?
Different jobs: the same design on a B2B job with B2B customers
Sensitive topics: does the openness hold for money, health or work?
Novelty effect: do ratings hold when the same people are invited again?
About the partners
Vendbridge (Zurich) is an expert consultancy for customer-focused growth. Its Jobs-to-be-Done approach, Customer-Focused Innovation (CFI), draws on 20 years of practice, more than 200 growth missions and over 8,000 measured customer discoveries. In this test, Vendbridge provided the method, built the chat interviewer and did all the analysis.
Cartorga is an AI platform on which consultants build their own AI interviewer with their own method, deploy it to many participants at once in several languages, and analyse the findings. Cartorga designed and ran the voice interviews.
moweb (Düsseldorf) has run online panels since 2004 and today reaches more than 3 million members in over 30 countries. moweb recruited both groups and ran the fieldwork.
Hear your whole customer base, not just twenty of them.
Let Stella talk to your customers. A CFI discovery sprint: 100 or more AI interviews per customer group and a finished long list of Job Metrics in about two weeks, every need traced to a customer's own words.
Comments