You have a folder of exit interviews. Read end to end, they say three things: compensation, growth, and work-life balance.
They said the same three things last year, and the year before that, and the folder has grown without the answer changing.
That stability is usually read as a signal. It is a property of the instrument.
What the responses encode
An exit interview samples a person at the single moment when they have the most to lose from candour and nothing at all to gain from it. There is a reference outstanding. There may be a final payment, a rollover, a network they will need in four years. The conversation is conducted by the employer, recorded by the employer, and kept by the employer.
What gets written down is therefore not what happened. It is the overlap between what happened and what the person judged safe to put in writing on their way out of the building.
That overlap has a predictable shape. Compensation sits at the top of nearly every exit file in the world because compensation is the reason that costs nothing to give. It is impersonal. It implies no criticism of anyone still employed. It flatters the departing person slightly and it ends the conversation politely. Growth is second for the same reason.
I have written those answers myself. When you are leaving to get away from a manager or a culture, that is not what goes on the form. You smile, you shake hands, and you get out with the reference intact. In The Perception Revolution I called these the polite fictions, and the translations are not subtle. "Seeking new opportunities" means any environment has to be better than this one. "Better compensation elsewhere" means no amount of money was worth what this was costing.
So the frequency ranking your file produces is real. It is a ranking of what was safe to say. It is not a ranking of cause, and treating it as one is how organizations spend a retention budget on the thing their leavers were most comfortable naming.
Is an exit interview a trap?
That question sits in Google's People Also Ask box for this topic, which is worth more to you than any answer to it.
It tells you what a meaningful share of your respondent population believes about the instrument before they ever sit down with it. You are not running a survey. You are running a survey whose respondents have a theory about the survey, and are answering the theory.
It is worth saying plainly how thin the evidence base is here. When I went looking for peer-reviewed work on whether exit interviews produce accurate reasons for leaving, on departing-employee candour, or on how exit data compares against separations data, the most recent sweep of the organizational journals returned nothing. The pages that rank for this topic are almost entirely produced by companies selling exit interview software. That is not proof the instrument fails. It does mean that a practice most organizations run on autopilot has been examined mainly by the people who sell it.
This is not an argument for abandoning exit interviews. It is an argument for reading them as what they are: a low-fidelity signal produced under a known incentive, which is still useful once you stop asking it a question it cannot answer.
Read the file as a distribution
The content of the free text is the weakest data in the folder. The shape of the folder is the strongest, and almost nobody looks at it.
Four cuts, none of which require reading a single response:
Completion rate by manager. Who declines to sit the interview at all is data, and it is data that has not been filtered for safety. A team where half the leavers skip the conversation is telling you something the other half's answers are not.
Response length by manager. Three-word answers and three-paragraph answers come from different conditions. Word count is a crude proxy for candour and it is available in every system you own.
Elapsed time from resignation to interview. A conversation held eleven weeks after notice, when the person is already gone and the reference has been given, is a different instrument from one held on day two.
Regret against non-regret separations. Split the file by whether you wanted the person to stay, then compare the stated reasons across the two groups. If they are identical, your instrument is not discriminating between two populations you know are different. That is a measurement failure and it is visible in an afternoon.
The two questions worth asking anyway
A compromised instrument still returns something if you ask it what people can safely answer.
When did you first seriously consider leaving? This asks for a date rather than a grievance. It costs nothing to answer honestly, nobody is accused, and it is the highest-value field in the whole exercise. If the median answer across your file is well over a year, then your notice period is not thirty days. You had eighteen months and no mechanism that could see them.
What would have had to be different? Counterfactual and forward-looking. It lets someone describe a system without indicting a person, which is the only register in which most people will tell you anything true on their last day.
Both are safe to answer. That is the entire design criterion.
Triangulate against the separation data
Your exit file is the weakest evidence you hold about why people left. Your HRIS is the strongest evidence you hold about who left, when, and from where. They are rarely joined, and joining them takes an analyst a day.
Pull separations for the last twenty-four months with manager, team, tenure at exit, date, and time since last promotion or level change. Then overlay the exit reasons. What you are looking for is not a correlation with the stated reason. You are looking for clusters that the stated reason does not explain.
Departures concentrate. They concentrate by manager most often, and by team, by shift, and by the eighteen months after a reorganisation. A cluster is visible in the HRIS whether or not anybody wrote down why, which is precisely what makes it better evidence than the free text. Nobody had to feel safe for a date to be recorded.
Here is the check that usually settles it. MIT Sloan researchers analysing thirty-four million employee profiles found toxic workplace culture to be 10.4 times more predictive of turnover than compensation. Your exit file almost certainly says compensation. Both cannot be describing the same mechanism.
One of them is measuring what people do. The other is measuring what people say on a form, to their employer, with a reference outstanding.
The 10.4 is the number to bring to a budget meeting, because it is a ratio and ratios survive scrutiny better than totals. For scale only: SHRM has put the cost of turnover linked to workplace culture at $223 billion over five years. That figure is too large to act on and I include it to establish that this is not a soft problem. What a single departure costs you is a separate calculation, and the house multiplier probably has it wrong.
What you could notice this week
Take one manager whose team has had three or more departures in two years, and count how many of those exit interviews were completed at all.
You do not need the transcripts. The count is the finding.
Read the spread rather than the average. The perception survey metrics on our tools page cover what to measure when the instrument you have is measuring the wrong population.
Sources
- Sull, D., Sull, C., & Zweig, B. (2022). Toxic Culture Is Driving the Great Resignation. MIT Sloan Management Review, 11 January 2022. sloanreview.mit.edu
- Society for Human Resource Management (2019). The High Cost of a Toxic Workplace Culture — the $223 billion over five years figure.
- Mann, P. A. (2026). The Perception Revolution, Chapter 2, "The Pattern Hidden in Exit Interviews". Fast Company Press.