Industries / Finance and business services / Advertising and marketing / Market research
Interviews and focus groups
Qualitative research talks at length with a few people, alone or in groups, to learn why they think and act as they do.
4
Milestones achieved, from 1941 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1941
The focused interview
In November 1941 Paul Lazarsfeld took his new Columbia University colleague Robert K. Merton to a test of a government radio program, where a little over a dozen listeners pressed buttons to record what they liked and disliked and were then asked why. Merton tried the questioning himself, and with Patricia Kendall he set out the method in "The Focused Interview" (American Journal of Sociology, May 1946), the start of what is now called the focus group, according to Liza Featherstone's history.
LessonButton presses recorded what people liked; only the conversation afterwards could show why.
Credit
Robert K. Merton, Paul F. Lazarsfeld and Patricia L. Kendall, Columbia University
1985
Taste tests miss the bond
On 23 April 1985 the Coca-Cola Company changed its formula for one "preferred in taste tests of nearly 200,000 consumers." In the company's words, "What these tests didn't show, of course, was the bond consumers felt" with the original; calls to its consumer line rose from 400 a day to 1,500 a day by June, and the original returned as Coca-Cola classic on 11 July, 79 days later.
LessonA test that measures one thing, here taste, says nothing about what it did not ask.
Credit
The Coca-Cola Company
2005
Ethnographers in industry meet
Ken Anderson of Intel and Tracey Lovejoy of Microsoft organized the first Ethnographic Praxis in Industry Conference (EPIC) in 2005, for ethnographers employed in and by companies. Ethnographers study people by watching and talking with them where they live and work, not only by asking set questions. EPIC grew into a nonprofit membership organization and annual conference applying "interpretive sciences to business."
Credit
Ken Anderson and Tracey Lovejoy
2026
AI codes a focus group
Researchers at the University of Southampton gave one in-person focus-group transcript to two general AI models (ChatGPT-5 and Claude 4 Sonnet), an AI coding application (QualiGPT) and blinded human analysts. Applying a fixed list of 10 codes, the AI models agreed with an expert panel 93.5% of the time and the humans 92.7%; only one model kept pace with humans at finding themes unaided, and 1.2% of the quoted evidence in that theme-finding did not support the code given, by the authors' hand check of every quote.
LessonApplying a code list is within reach of AI; finding themes and keeping track of who said what is not yet reliable.
Credit
C. Hill, A. Dahil, G. Simpson and colleagues, PLOS Digital Health, April 2026
Watch for
AI interviews at scale
Anthropic says its Anthropic Interviewer, built on its Claude models, ran 1,250 interviews of about 10 to 15 minutes with professionals recruited through crowdworker platforms, and it has released all the transcripts. Afterwards 97.6% of participants rated their satisfaction 5 or higher on a 7-point scale and 99.12% said they would recommend the format; Anthropic notes that its crowdworker sample may lean toward people with positive or experienced views of AI.
Achieved whenan outside team compares the AI's interviews with human-led interviews on the same guide and finds comparable depth.
Claimed by
Anthropic, December 2025
A million AI interviews
Listen Labs, a San Francisco firm whose AI interviewer questions participants before its software produces reports, says its platform has interviewed over one million people, drawn from a panel of around 30 million. It raised a US$69 million Series B led by Ribbit Capital and names Microsoft, Perplexity and Sweetgreen among its customers.
Achieved whenan independent audit or published comparison shows its interviews match human-moderated ones on depth and accuracy.
Claimed by
Listen Labs, January 2026
Open
01
Check quotes in AI reports
A checker that takes an AI-written qualitative report and the transcripts it came from, finds each quoted passage, and records whether the quote is word for word, whether it is credited to the right speaker, and whether it supports the claim it is attached to. Then run it on reports from at least five AI research tools.
The briefCredit · unclaimed02
Trial AI-moderated focus groups
A randomized comparison in which groups of six to eight consumers discuss the same topic, half moderated by an AI voice moderator and half by experienced human moderators using the same guide, with transcripts scored by analysts who do not know which was which.
The briefCredit · unclaimed
Open 01 · the brief
Check quotes in AI reports
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A checker that takes an AI-written qualitative report and the transcripts it came from, finds each quoted passage, and records whether the quote is word for word, whether it is credited to the right speaker, and whether it supports the claim it is attached to. Then run it on reports from at least five AI research tools.
Why it is still open
The Southampton study checked every quote by hand, for one transcript, and found the models' main weakness was keeping the researchers' speech apart from the participants'. Anthropic has published 1,250 interview transcripts under the MIT license, a large shared test set, and searches found no published test of quote accuracy across commercial tools.
Built with what exists
The Anthropic Interviewer transcripts on Hugging Face; a fuzzy string-matching library such as RapidFuzz; an AI model as judge of whether a quote supports a claim, checked against people on a sample; the AI research tools under test.
First step
Download the transcripts, write the word-for-word matcher, and check 100 quotes from one tool's report by hand.
Done looks like
For each tool, the share of quotes that are verbatim, misattributed and unsupported, across at least ten research questions, with 500 quotes also checked by two people to measure the checker's own error.
Open 02 · the brief
Trial AI-moderated focus groups
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A randomized comparison in which groups of six to eight consumers discuss the same topic, half moderated by an AI voice moderator and half by experienced human moderators using the same guide, with transcripts scored by analysts who do not know which was which.
Why it is still open
The main published test, Focus Agent (September 2024), ran five sessions with 23 participants, and its abstract says the work "exposes some improvements associated with LLMs acting as moderators." Anthropic's 1,250 AI interviews, above, were held with one participant at a time, which leaves out the group discussion.
Built with what exists
A video meeting service such as Zoom or Jitsi; speech-to-text with speaker labels, such as ElevenLabs Scribe v3, which ElevenLabs says makes 17% fewer speaker-labeling errors than its earlier model on recordings with three or more speakers (see the Note "Fewer word errors"); a current AI model as moderator; a text-to-speech voice.
First step
Write the guide and the scoring sheet, then run two pilot groups with volunteers to test turn-taking.
Done looks like
At least 20 groups, 10 in each arm, with pre-registered scores for coverage of the guide, follow-up questions asked, share of talk by the quietest participant and themes found, rated by three blinded analysts and published with consenting participants' transcripts.