Regretted, and benchmarked

The list these pages use ticks five categories as regretted: Better Opportunity, Career Growth, Compensation, Management/Culture and Work-Life Balance. The idea behind the line is "could we have kept them": those five are reasons an employer might have answered, and the rest are reasons it could not (Retirement, Further Study, Relocation), did not want to (Involuntary/Performance), or cannot classify (Unknown, Personal/Health). Five ticks give 137 of 210 = 65.2 percent.

Every one of those placements is arguable. Relocation is unavoidable unless remote work was on the table; Personal/Health is unknowable unless the exit interview said more. Tick both and the share is 171 of 210 = 81.4 percent. Decide that only the three market reasons count, the ones a counter-offer could have touched, and it is 95 of 210 = 45.2 percent. Keep the five but take the 17 Unknowns out of the base, because they might be anything, and it is 137 of 193 = 71.0 percent. All four are computed on the same constructed rows; none is wrong.

So the deliverable is not a share. It is a share with its list printed beside it, and the list is what a colleague can disagree with. "Regretted attrition is 65 percent" is an assertion; "65.2 percent, counting these five categories as regretted" is a finding, and it tells the reader exactly which tick to argue about.

Apex's reason_category is itself a coding of a free-text column, and the free text is what the exit interview actually captured: "offer i couldnt refuse", "underpaid for the work", "bad manager", "no growth path here". A coding step sits between those phrases and the eleven categories, and it was a person or a rule that put "offer i couldnt refuse" under Better Opportunity rather than Compensation. Print the raw phrases under each category once a year. If the coding drifts, the regretted share drifts with it, and nobody will see the drift in the category counts. On a real table, the coding rule belongs in the six decisions from deep-dive session 1, as a seventh line.

All eight rates use the same recipe from deep-dive session 1: leavers in the window over headcount on the start date. That makes them comparable in construction. It does not make them comparable in precision. Sales at 50.0 percent is 54 exits over 108 people; one more exit would read 50.9. HR at 10.7 percent is 8 exits over 75; one more exit reads 12.0, one fewer reads 9.3. The HR rate is a number about one person wide, and a ranking that puts HR at 10.7 above Finance at 8.3 is a ranking of noise.

The habit is to print the base under the rate, every time, in the table and on the slide. "HR 10.7 percent (8 of 75)" lets the reader do the one-person arithmetic in their head; "HR 10.7 percent" invites them to compare it with Finance. On a dashboard the same discipline is one row per rate with the numerator and denominator visible, which Foundations session 6 builds without code.

The two big rates are the ones worth a meeting: Sales at 50.0 percent on a base of 108 and Customer Service at 37.6 on 93 are far from the rest and wide enough to survive a person or two. That is what a base tells you: which differences are worth a conversation and which are the arithmetic of small numbers.

Cross regretted with department and the cells shrink: HR has 8 exits in total, so its regretted count is at most 8 and its regretted share is a fraction of a handful. Two cuts on 210 rows produce a table where most cells are under ten, and a share computed on a cell under ten should be shown as the count, not the percentage. The same limit applies to reason categories by entity, by tenure band, by anything. A useful rule for a people table: below about 20 in the base, print the count and the base and no rate at all; the reader can divide, and will be reminded not to.

A quits rate of 2.0 percent means that in that month, quits equalled 2.0 percent of employment. Twelve such months are about 24 percent of employment in quits over a year, if the same people never quit twice and the base never moves. Neither holds exactly, so 24 is a rough scale, not an annual rate, and the honest sentence carries both conditions: "roughly 24 percent a year, multiplying the monthly rate by 12, which assumes no repeat quits and a fixed base". Written that way it is useful. Written as "the US annual quit rate is 24 percent" it is a number nobody measured.

Apex's side has the same shape with a different clock. 22.3 percent for the half year, times two, is 44.6 percent a year, if the second half repeats the first. Now the two numbers can sit on one line: "Apex roughly 45 percent a year (H1 doubled), the US roughly 24 (monthly times 12)". Both are labelled scales, and the labels say why a reader should not subtract one from the other and call it a gap.

Then the bases. JOLTS counts voluntary quits only, across every US employer, in every industry, on the employer's own employment count. Apex's 210 include involuntary exits, retirements and study leave, in one constructed group of four entities. Remove the 22 involuntary, retirement and study exits from Apex and the comparison is a little closer; it is still a different country, a different mix of industries, and a different definition of a person. The benchmark answers "is this order of magnitude unusual", and nothing finer.

The total hides a spread. In December 2025 the quits rate was 4.4 percent in leisure and hospitality, 1.9 in professional and business services, 1.9 in private education and health, 1.4 in manufacturing; across 2025 the means were 3.92, 2.34, 1.93 and 1.39. A company in hospitality comparing itself with the total is comparing itself with the wrong line, and the API returns the industry line for the same seven lines of code. Which industry Apex "is" is itself a choice: four entities, one of them retail and one logistics, and no single JOLTS series describes the group.

One number that does not appear on any page of the track is the cost of a departure. The figure that circulates, "replacing someone costs 1.5 to 2 times salary", has no measured source; it is attributed to an analyst rather than to a study, and the range SHRM itself publishes is 50 to 200 percent of salary, which is wide enough to mean anything. A cost model can be built for one employer from its own recruiting spend, ramp time and vacancy cover, and that is a finance exercise with the employer's own inputs. It is not a benchmark, and quoting the folk number as one is the fastest way to lose the room when someone asks where it came from.