AI interviewers double customer needs on the same budget

AI voice interviews can make adaptive customer research cheap enough to repeat at large scale, without turning each customer into a predictable digital twin. In a pre-registered study of 317 home-improvement decision-makers, AI moderation recovered 238 unique needs on an approximately $5,000 budget, compared with 109 for human moderation. A company could search thousands of spoken conversations for worries about windows, doors, and installation, provided the system remains accurate and safe beyond this test.

Companies that want to know why customers buy face an awkward choice. A cheap, standardized survey can ask many people the same questions, but it may miss the explanation behind an answer. A live interview can ask a useful follow-up when someone mentions a worry or an unexpected need, but one moderator can only speak with so many people. That makes large-scale, adaptive interviewing slow and expensive.

Why the follow-up matters

Market researchers use qualitative interviews to find the needs and words that a fixed questionnaire may miss. The important part is not only asking a prepared question. It is noticing which answer deserves another question.

This matters for a second reason. A digital twin in this setting is a language-model persona built from a person’s data. Researchers can ask it a new question and have it answer as that person might. An interview may reveal fairly stable preferences, such as concerns about price or installation. But it may not capture the quick visual or emotional reaction that shapes a response to an advertisement. Describing a person plausibly is easier than predicting exactly how that person will react next.

A voice interviewer that keeps probing

The study tested whether an AI voice interviewer could make the useful part of a live interview cheaper. It included 317 people who make home-improvement decisions. They discussed windows, doors, and a partner brand in one of three formats: an AI-moderated spoken interview, a static spoken interview with fixed questions, or a human-moderated interview. Each format lasted about 30 minutes and covered the same topics.

The AI could respond to what a participant had just said instead of waiting for the person to guess every relevant detail. AI-moderated participants spoke for 8.1 minutes on average and produced 952 words. Static interviews produced 609 words, while human-moderated interviews produced 1,404. AI interviews covered more coded themes than static interviews in three of four open-ended sections.

The difference became clearest when the researchers counted distinct customer needs. After 24 interviews, AI moderation recovered an average of 111.5 unique needs, compared with 109 for human moderation and 75.9 for static interviews. Under a market-research budget of about $5,000, AI moderation recovered 238 unique needs. Static interviews recovered 188, and human moderation 109. The calculated cost per unique need was $19 for AI, $27 for static interviews, and $36 for human interviews.

That does not mean the AI sounded like a person. Participants speaking with human moderators were more positive, activated, and assertive. On the study’s speech-based measures of affect, AI and static interviews were statistically indistinguishable. The advantage was the ability to ask adaptive questions repeatedly without paying for a human moderator each time.

More conversations, not a crystal ball

The researchers also used the interview transcripts to build digital twins. These twins predicted held-out responses to mailers, commercials, and marketing claims better than twins built from demographics alone. But the richer AI transcripts did not produce significantly better quantitative predictions than static transcripts.

The twins also gave more analytical explanations than real people did, especially when reacting to image and video marketing. In the AI-interview condition, adding a participant’s response to another similar stimulus improved rating accuracy from .743 to .787. That is a useful improvement, but it points to a limit: a longer conversation does not automatically reproduce an immediate reaction to a picture or video.

In a home-improvement company’s research program, a homeowner might record a conversation about drafty windows after dinner instead of waiting for a scheduled focus-group slot. The next morning, researchers could search thousands of conversations for repeated worries about installation, cost, and comfort. That depends on AI interviewing remaining accurate, safe, and useful beyond the high-involvement purchase tested here.

At a marketing review, a team could use an AI-generated customer twin to organize questions, not to decide that a real person will like a new commercial. It would still need to test closely related ads while models miss people’s immediate reactions to images and video. The technology makes many adaptive interviews affordable; it does not make customers predictable.

Sources

AI-Moderated Interviews for Market Research and Digital Twins Calibration