BRENNEN

Industries / Finance and business services / Legal and professional services / Consulting

Interviews and fact-finding

Interviews and fact-finding is how consultants gather what is not written down: talking with a client's staff, customers and outside experts, running surveys, and watching the work being done.

6

Milestones achieved, from 1928 to 2026.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1928

    Interviews that listen

    At Western Electric's Hawthorne plant, Elton Mayo and Fritz Roethlisberger of Harvard Business School oversaw more than 21,000 interviews with employees from 1928 to 1930 and trained the researchers who ran them. They found that employees spoke more candidly when encouraged to talk freely than when answering set questions, so the program moved to "nondirected" interviews, which grew from about 30 minutes to 90 minutes or two hours. A research report noted that "the interview must be a listening rather than a questioning process."

    library.hbs.edu

    Credit

    Elton Mayo, Fritz Roethlisberger and Western Electric

  2. 1951

    Experts asked in rounds

    RAND's Project Delphi questioned a panel of seven experts, four of them economists, on how many atomic bombs a Soviet planner would need to cut US munitions output by a set amount. They answered five questionnaires at about weekly intervals, with interviews after the first and third. The method, the memo says, "avoids direct confrontation of the experts with one another." The ratio between the largest and smallest estimate fell from 100 to 1 to about 3 to 1. The original memo is dated 14 November 1951 and was published in shortened form in July 1962. RAND says the Delphi method has since been used in health care, education, management and environmental science.

    rand.orgrand.org

    Credit

    Norman Dalkey and Olaf Helmer, RAND Corporation

  3. 2003

    One question for customers

    In 2003 Fred Reichheld, a partner at Bain & Company, proposed that companies track a single survey question: "How likely are you to recommend us to a friend or colleague?" The Net Promoter Score is the percentage of customers who answer 9 or 10 minus the percentage who answer 0 to 6. In July 2007 Timothy Keiningham and three co-authors used data from 21 firms and more than 15,500 interviews in Norway's customer satisfaction survey, and could not replicate the claim that the score predicts growth better than other measures.

    LessonA claim that one number predicts growth needs testing on data from outside the firm that proposed it.

    netpromotersystem.comnetpromotersystem.comdoi.org

    Credit

    Fred Reichheld, Bain & Company

  4. 2011

    Hawthorne data rechecked

    The Hawthorne experiments gave their name to the "Hawthorne effect", in which people in a study change their performance because they are being observed. Economists Steven Levitt and John List found the data from the first experiments, on factory lighting, which had been thought lost, and reported in January 2011 that the striking patterns in published accounts "prove to be entirely fictional", though they found subtler signs of such an effect.

    LessonA famous result from fieldwork is only as good as its records, and old records can be found and checked.

    aeaweb.orglibrary.hbs.edu

    Credit

    Steven D. Levitt and John A. List

  5. 2011

    Expert calls charged as tipping

    Expert networks arrange paid calls between clients, such as investors, and people who know an industry. On 3 February 2011 the US Securities and Exchange Commission charged six expert-network consultants and employees with tipping hedge funds and other investors with inside information about companies including AMD, Apple, Dell, Flextronics and Marvell, for nearly $6 million in illegal gains. Two of them were employees of the network Primary Global Research.

    LessonA paid interview with an insider can pass on confidential information, so expert calls need rules on what may be asked and said.

    sec.gov

    Credit

    US Securities and Exchange Commission

  6. 2026

    Chat interviews, AI against people

    In a randomized experiment, Semra Guven and four co-authors had 41 university students interviewed by text chat, either by trained non-expert human interviewers or by an AI system running on a small open-weight model with 7 billion parameters on the researchers' own hardware. Interviews took a similar median time. The AI asked more than twice as many questions and drew out about 65% more text, answers were almost always on topic in both groups, and no difference in how specific answers were was detected, though the test could not rule out one of up to three quarters of a point. Only respondents interviewed by the AI skipped questions.

    arxiv.org

    Credit

    Semra Guven, Tobias Gårdhus, Andreas Bjerre-Nielsen, Anna Rogers and Hjalmar Bang Carlsen

Watch for

  1. Interviews run by Claude

    Anthropic says its Anthropic Interviewer, a tool built on Claude, interviewed 1,250 professionals (1,000 from the general workforce, 125 scientists and 125 creative workers) for about 10 to 15 minutes each, then passed the results to human researchers. Of the participants, 97.6% rated their satisfaction 5 or higher on a scale of 1 to 7. The study did not compare the tool with human interviewers.

    Achieved whenan outside team runs the same interview guide with AI and human interviewers and publishes a comparison of what each found, graded by people who do not know which was which.

    anthropic.com

    Claimed by

    Anthropic, December 2025

  2. Expert calls with no interviewer

    GLG says its AI Moderator asks a GLG discussion guide "instead of a human interviewer", with adaptive follow-up questions and GLG's compliance rules built in. AlphaSense says its AI Interviewer runs expert calls of 30, 45 or 60 minutes, and that a human review finds and removes material non-public information before transcripts go to clients.

    Achieved whenan outside party compares AI-led and human-led calls on the same questions, for what they find and for compliance breaches, and publishes the results.

    glg.comhelp.alpha-sense.com

    Claimed by

    GLG (page undated) and AlphaSense, June 2026

Open

  1. 01

    Planted facts, two interviewers

    A head-to-head test of AI and human interviewers. Role-players act as staff of a made-up business with a problem, such as orders that ship late, and each holds a written list of facts that they reveal only when a question reaches them. AI and human interviewers get the same brief, and graders count how many planted facts each interview uncovers.

    The brief
    Credit · unclaimed
  2. 02

    A public test for leaks

    An open set of made-up expert-call transcripts about fictional companies, in which some passages disclose material non-public information (MNPI: confidential facts that could move a share price) and others only look as if they do. Two reviewers trained on the rules label every passage, and screening tools are scored on what they catch and what they flag wrongly.

    The brief
    Credit · unclaimed

Open 01 · the brief

Planted facts, two interviewers

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A head-to-head test of AI and human interviewers. Role-players act as staff of a made-up business with a problem, such as orders that ship late, and each holds a written list of facts that they reveal only when a question reaches them. AI and human interviewers get the same brief, and graders count how many planted facts each interview uncovers.

Why it is still open

InterviewPlayground, an October 2026 simulation for testing AI interviewers, gives each simulated participant six insights but scores how much of the interview guide is covered, not the share of insights found, and tests no human interviewers. The randomized comparison by Guven and co-authors measured the amount of text, specificity and relevance, with no answer key.

arxiv.orgarxiv.org

Built with what exists

the open-source interview platform of Friedrich Geiecke and Xavier Jaravel (github.com/friedrichgeiecke/interviews) or another current AI model as the AI interviewer; volunteer or paid role-players; a video-call tool; speech-to-text such as ElevenLabs Scribe v3 for voice interviews (see the Note "Fewer word errors"); a scoring sheet and two graders.

  1. First step

    Write one scenario with 15 facts and run five AI and five human interviews with the same role-player.

  2. Done looks like

    At least 60 interviews across three scenarios of 15 planted facts each, half run by AI and half by people, scored by two graders who agree on at least 90% of facts, with the share of facts found by each kind of interviewer published.

Open 02 · the brief

A public test for leaks

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open set of made-up expert-call transcripts about fictional companies, in which some passages disclose material non-public information (MNPI: confidential facts that could move a share price) and others only look as if they do. Two reviewers trained on the rules label every passage, and screening tools are scored on what they catch and what they flag wrongly.

Why it is still open

Screening tools report their own figures with no public test set. FullStack says the AI call auditor it built for a compliance provider flags "91% of Material Nonpublic Information Disclosures", without saying how many calls were tested or how accuracy was measured, and AlphaSense says people review its AI-led call transcripts for MNPI before release.

fullstack.comhelp.alpha-sense.com

Built with what exists

a current language model to draft transcripts from fictional company profiles; the SEC's 2011 expert-network case as a pattern for what a leak looks like; open-weight models and commercial tools as the screeners under test; a labeling spreadsheet.

  1. First step

    Write 20 transcripts for one fictional company, 10 of them with planted leaks, and have two people label them independently.

  2. Done looks like

    500 transcripts with at least 1,000 labeled passages, two reviewers agreeing on at least 90% of labels, and published catch rates and false-alarm rates for at least three screening tools.