Method
AI organizational ethnography
The lab practices a method that did not have a settled name. This is what the term means here, and what it does not mean.
AI organizational ethnography is the practice of using an AI interviewer to hold an open-ended, adaptive conversation with every person in an organization, rather than sampling a few of them or sending all of them a questionnaire, and then reading the resulting accounts for patterns in how the work is actually done.
Ethnography is the older discipline: you go where the work happens, you talk to the people doing it, and you write down how it is actually done rather than how the manual says it is done. Its weakness has always been arithmetic. A researcher can hold thirty good conversations in a quarter, and thirty conversations inside a company of four thousand people invites the obvious objection. The method below keeps the ethnographic posture and removes the arithmetic constraint.
Four properties
Census, not sample
Traditional ethnography studies a handful of people deeply and generalizes carefully. This method interviews everyone in scope. That removes the sampling argument from the conversation: a leader cannot dismiss a finding as unrepresentative when the finding came from the whole population rather than the fourteen people who agreed to a focus group.
Conversation, not instrument
A questionnaire fixes the questions before it meets the first respondent, so it can only collect what was anticipated. An adaptive interview chooses the next question from the last answer, which means it can follow a thread nobody knew was there. The prior still steers what is asked; it just does not close the door.
Incidents, not attitudes
Asking how satisfied someone is with a rollout produces a number that moves for reasons the number cannot explain. Asking what happened the last time they used the new system produces an account with a date, a workaround, and a name for the obstacle. Accounts of specific incidents are what make a finding actionable.
Cohort, not individual
Ethnography traditionally protects subjects through anonymization at the point of writing. Here the protection is structural: nothing is reported below a minimum number of voices, transcripts are sealed, and the findings are barred from individual employment decisions. That constraint is what buys the candor the method depends on.
How it relates to adjacent methods
CLASSICAL ORGANIZATIONAL ETHNOGRAPHY
In common. Both take seriously that work as imagined differs from work as done, and both privilege the account of the person doing the work.
Different. A human ethnographer achieves depth on a small number of informants over months. This method trades some of that depth for coverage of the entire population in days.
EMPLOYEE ENGAGEMENT SURVEYS
In common. Both try to read the workforce at scale, and both report in aggregate.
Different. A survey collects positions on questions chosen in advance. This method collects accounts, and can be surprised. See the fuller comparison on the comparison page.
CONVERSATIONAL OR AI-ASSISTED SURVEYS
In common. Both use a language model to ask follow-up questions rather than presenting a fixed form.
Different. A conversational survey is still organized around instrument items and usually still runs in text. This method is organized around eliciting an account, runs in voice, and treats how something was said as part of the data.
QUALITATIVE RESEARCH AND CODING
In common. Both end in coded qualitative data read against a codebook.
Different. The interviewing and the first coding pass are conducted by the same system, which is what makes census coverage affordable. It also means the method has to be validated rather than assumed, which is what the lab publishes on.
What is not yet established
A method is worth only what has been tested about it. The lab has published a controlled benchmark showing that an AI-conducted voice interview preserves enough signal for a blinded scorer to recover which change-readiness barrier a participant was describing, at 50.0% accuracy against a 20% chance baseline. That is a first result about signal, not a demonstration that the method diagnoses real organizations. Human-coder validation and prediction of actual adoption outcomes are still open, and the lab states so in the paper.