Headcount is a stock, attrition is a flow

A stock is a count at an instant. Headcount, open positions, people on leave, people in a grade. It has no meaning without a date, and two stocks from different dates cannot be subtracted to give a flow unless nothing else moved between them, which it always did.

A flow is a count over a period. Joiners, leavers, promotions, transfers, hires, offers. It has no meaning without a start and end date, and a flow reported "for the year" in March is a different number from the same flow reported in December.

The identity is the only check that needs no judgement: start stock, plus every flow in, minus every flow out, equals end stock. At Apex, 1,022 on 1 January, 49 joiners and 210 leavers to the end of June, 861 on 1 July. It reconciles. When it does not, the usual reasons are a transfer counted as a leaver on one side and not as a joiner on the other, a contractor in one stock and not the other, or a leaver whose last day was entered in the wrong month.

The reason this matters beyond tidiness: every rate on a people dashboard is a flow divided by a stock, and if the two were not measured on consistent dates, the rate is a fraction of two unrelated numbers.

Apex has four entities and 291 of its 1,071 rows carry a transfer flag: somebody who moved from Apex Retail to Apex Logistics is a leaver on Retail's dashboard and a joiner on Logistics', and neither on the group's. Whether a transfer counts as attrition depends entirely on whose tank you are looking at, and a group-level attrition figure built by adding up entity-level figures will count every transfer as a loss.

The same applies to people on long leave (in the stock or not?), contractors converting to staff (a joiner or not?), and rehires (a joiner, but not a new person). None of these has a universal answer. Each has to be written down once, in the definition of the stock, so that the identity can be checked with the same rules on both sides.

The start base asks "of the people we had in January, what share left". Clean, but it ignores the 49 who joined and were also at risk of leaving, and it understates the rate in a growing company. The end base asks the question backwards: it divides by a headcount that is smaller because people left, so the more people leave, the more the rate inflates. The average base approximates the number of people who were exposed to the risk of leaving during the period, which is what an attrition rate is trying to measure, and it is why most standards use it.

None of this makes the average base right in every case. A company that doubled in size in a month has an average that describes nobody. The point is narrower: the base is a decision, it moves the rate by points, and it goes under the number in writing. Foundations session 2 takes the same 210 leavers by department and by entity, where the base problem gets worse because the bases get small.

Apex's 22.3 percent covers six months. Doubled, it is 44.6 percent a year, and that is the number that will be quoted. The doubling assumes the second half repeats the first, which for a company that lost 54 of 108 salespeople in six months is a guess in either direction: the exodus may be over or may be halfway. Annualised figures should carry the word "annualised" and the period they were built from, and a half-year figure compared with last year's full-year figure is not a comparison.

Take the median tenure of Apex's active staff: 5.01 years. It sounds like "people stay about five years". It is not that. It is the median tenure of the people who were still there on 1 January, which is a group that by construction excludes everyone who left earlier, and people who leave tend to leave early. The 210 who left in the first half of 2026 had a median tenure at exit of 1.52 years, and 61 percent of them had been at Apex under two years. Two very different pictures of the same company, and both are computed correctly from the table; the table itself is the problem.

The question that catches it: what would a person have to have done to be missing from this table? For Apex, the answer is "left before 2026". Any statistic about how long people stay, how many leave in their first year, or which cohort is loyal has to be read knowing that the early leavers of every year before 2026 are absent. Deep-dive session 2 measures exactly how far this pushes a survival curve, and what kind of table fixes it, which is an event log rather than a photograph.

Every people dataset is a photograph of who was around to be measured. Engagement surveys measure the people who filled them in, which excludes the disengaged in proportion to their disengagement. Performance reviews measure the people who stayed to the review date. Exit interviews measure the leavers who agreed to talk, and the reason they give is the reason they were willing to give. None of these are useless. Each has a missing group whose shape can be guessed, and the guess belongs in the sentence that reports the number. Foundations session 5 is the survey version of this problem.