The resignation was visible six weeks before the letter

"Honestly? Nothing. It came out of nowhere." That is a direct quote from an engineering director on a people-analytics engagement, describing the resignation of the platform lead whose departure had just cost a quarter of roadmap. It is also, nearly word for word, what the exit interviews in the company's own HR system said managers reported in most regretted departures. Out of nowhere is the standard managerial weather report for attrition.

The data disagreed. Working with anonymized, aggregated workplace metadata, calendar density, internal mobility applications, review cycles, engagement survey trajectories, we reconstructed the months before a set of historical regretted exits. The pattern was not subtle once you looked at it as a curve rather than a memory: measurable disengagement began, on average, six to eight weeks before notice. One-on-ones got shorter, then rescheduled, then sparse. Voluntary participation, the demos, the guilds, the optional reviews, faded first. Survey sentiment dipped one cycle before the letter, in exactly the categories the departing person had previously scored highest.

"Out of nowhere" almost always means "visible for six weeks in signals nobody was assigned to watch."

Let us name the danger before the method, because this is the one domain where the method can go morally wrong. Scoring individuals for flight risk and handing managers a watchlist is surveillance wearing an analytics costume, and it poisons exactly the trust it claims to protect. It is also, conveniently, unnecessary: the leverage is not in predicting which person resigns. It is in seeing which teams, roles, and conditions are generating departure risk, early enough to change the conditions.

Done properly, this is survival analysis at the cohort level, the same mathematics hospitals use for patient outcomes, applied to tenure. Which factors shift the hazard of leaving: compensation drift against market, time since last role change, manager span, on-call load, the gap between review score and promotion outcome. The model's output is not a name. It is a statement like: senior engineers in their third year with no role change and two skipped promotion cycles carry four times the baseline hazard, and you currently employ eleven people in that cell.

(The exit interview, arriving after the decision, is the least informative document in the building. The calendar had the story six weeks earlier, and the compensation file had it six months earlier.)

The replacement cost of a senior engineer, recruiting, ramp, the projects that stall in between, runs between half and a full year of salary in most credible studies, before counting the departures that follow a respected person out the door. On that engagement, regretted attrition in the highest-risk cohorts was not a people metric; it was the single largest unbudgeted cost line in engineering, larger than the cloud bill everyone met about monthly.

We built the cohort model on the client's existing systems, HRIS, calendar metadata in aggregate, survey history, with a privacy boundary designed in from the start: no individual scores exposed, minimum cohort sizes enforced, the works council consulted before the first query ran. The output went to quarterly workforce planning, not to line managers.

The model does not tell you who is leaving. It tells you which conditions are manufacturing leavers, while the people in them are still persuadable.

The conditions, once named, were fixable with ordinary tools: a role-rotation program aimed at the stagnation cohort, compensation reviews re-anchored to current market rather than hiring-date market, and a hard look at two teams whose hazard curves said management problem in every language except words. Four quarters later, regretted attrition in the targeted cohorts had fallen by a third. No one was ever on a list.

Pull three years of departures and compute one table: hazard of leaving by tenure band and by time-since-last-role-change. Two columns, one afternoon, and it usually locates your largest leak immediately. The full model adds the other factors and the forward view; the table tells you whether the problem is worth the model. At TensorLabs we run this as a privacyfirst engagement, from that first table to the standing cohort view.

Reply with your engineering headcount and your last twelve months of regretted departures, two numbers, and we will translate those figures into replacement cost and whether your size justifies the full survival build or just the afternoon table. We will happily talk you out of the bigger version if the small one answers it.