Replace the survey with a conversation.
Interviewer is an agentic interviewer and researcher that gathers information at scale through natural conversation. It adapts in real time — optimizing what it asks against the study's data targets, the respondent's previous answers, their tone, and the pace of the exchange — so every session reads like a genuine interview and still resolves into clean, comparable data.
The researcher's dilemma.
Every research team faces the same fork. A web form is cheap to field and easy to tabulate, but it's rigid: respondents skim, satisfice, and drop off, and every answer is bounded by the question you happened to write. A live interview is rich and responsive — a good interviewer hears something surprising and follows it — but it doesn't scale, costs a fortune, and leaves you with transcripts that resist counting.
You normally pick one and live with the cost. Interviewer's premise is that a language model, used carefully, can collapse the tradeoff: conduct a genuine conversation with every single respondent, and still hand you a tidy table at the end.
The research backs it.
We didn't invent the idea that a conversation beats a form. It builds on a decade of survey-methodology research, and a few findings shaped the approach:
- Traditional surveys are in trouble. Response rates for a typical telephone survey fell from 36% in 1997 to 9% by 2012, and to just 6% by 2018 — part of a broad decline across nearly every survey mode, alongside well-documented satisficing and straightlining as rushed respondents stop differentiating their answers.
- A conversation measurably improves the data. In a roughly 600-person field study across more than 5,200 open-ended answers, an AI chatbot survey collected 39% more information and 25.7% higher-quality responses than the identical questions on a standard web form (Xiao et al., ACM Transactions on Computer-Human Interaction, 2020).
- Dynamic probing is what makes the difference. A 2025 randomized experiment with 1,800 respondents found that LLM chatbots which dynamically probe for elaboration produced more detailed and informative open-ended answers than fixed questions alone (Barari et al., 2025).
- It scales without losing rigor. Economists at the LSE show AI-led interviews can be run with thousands of respondents and benchmarked against human experts, scoring highly across decision-making, political, and subjective-experience domains (Geiecke & Jaravel, 2025).
The research is honest about the tradeoffs, too: that same 2025 experiment found the richer data came at a slight cost to respondent experience, and live-coding open-ended answers is imperfect. Getting those details right — the pacing, the probing, knowing when to stop — is exactly what this experiment is for.
One conversation, two artifacts.
Under the hood the design is two layers. An open conversation layer carries the persona, the opening questions, and natural-language follow-up rules — the model improvises within a topic the way a thoughtful interviewer would. A closed data layer sits underneath: each topic declares the specific fields it must yield, as an enum plus a supporting verbatim quote.
Every session therefore produces two things simultaneously:
- A warm transcript — a real back-and-forth the respondent experienced as a conversation, not a form.
- A structured record — one clean row per respondent, one column per data point, ready to count and cross-tab across the whole panel.
Capture happens twice: in real time, as an answer becomes clear during the chat, and again as a single structured pass over the full transcript once the interview ends — the second pass is the canonical record. A point that never came up is recorded as exactly that: "did not come up" is a distinct outcome from a "no."
Structure is enforced by the system, not requested of the AI.
The single biggest quality improvement came from taking control away from the model. The server owns the interview state — which topic we're on, how many turns it has taken, whether the required fields are captured. The model only ever sees the *current* topic, never the ones ahead, so it can't skip forward. It can *propose* that a topic is finished; the server *decides* whether to advance.
The same discipline governs pacing and the tappable answer chips: both are driven by server state, not model judgment. When a topic runs long, the server force-advances it rather than letting the conversation wander. "The model proposes, the server disposes" turned out to be the difference between a demo and something you can put in front of thousands of people.
One engine, any study.
A study is defined in plain language, not code. Rather than scripting rigid questions and fixed answer choices, a researcher describes what they want to learn — the topics to cover, the specific data each one must yield, and the tone the interview should carry — and the engine conducts it. Standing up a new interview is an authoring task, not an engineering project.
That separation is what makes this a platform rather than a one-off. The same engine can field a contract-review study for lawyers today and an employee-experience study tomorrow, each with its own voice and its own data targets.
In the field.
Interviewer is being evaluated in the field, with two live research projects running real interviews with real respondents right now:
- A large-scale study of how lawyers are experiencing AI in their everyday contract work — going out to roughly 5,000 invited participants.
- A strategic-diagnostic interview for large-firm partners, surfacing the competitive advantage of a practice and where it's exposed.
Both put the same question to the test — can an AI interview be better than a form *and* more countable than a conversation — and the early signs say yes.
Systems that ask professionals to self-report their expertise die of neglect. Every time. What survives is extraction: deriving the knowledge from work that already happens. We stopped designing forms and started designing listeners.