TWINLADDER
TwinLadder logoTwinLadder
Atpakaļ uz apskatu

Izdevums #49

Seven and a Half Years, Down from Twelve

A manager said two things in one conversation: average tenure at her employer runs at about two years, and she would automate away twenty of her fifty people if the tools existed. The series says otherwise — US median tenure was 3.9 years in January 2024, higher than in 1983, 1987, 1998 and 2000, with voluntary quitting running below its pre-pandemic rate. What moved underneath the headline is the tail: a man aged 45 to 54 held 12.8 years with his employer in 1983 and 7.5 in 2024, the share of workers aged 25 and over with ten years or more fell from 33.3% to 30.2% in a decade, and an EDGAR search back to 2001 finds no US issuer that has ever disclosed the median tenure of its own workforce.

employee tenure
human capital disclosure
entry-level hiring
institutional knowledge
board metrics
2026. gada 15. augusts16 min read
Seven and a Half Years, Down from Twelve

TwinLadder Weekly

Issue #49 — Seven and a Half Years, Down from Twelve

15 August 2026 · Weekly intelligence on judgment, governance, and the boards accountable for both

Editor's Note

From Alex —

I met a manager this month who told me two things in one conversation, and I have been turning them over since.

The first was that average tenure at her employer runs at about two years. It is a company large enough that nobody in it knows everyone in it. The second was that she runs a team of about fifty, and that if the tools existed to do it, she would be happy to automate away twenty of them.

She was describing her budget, accurately. Twenty salaries sit on a line she owns and answers for every quarter. What those twenty people know about how the work actually goes — which complaint is the one that matters, which supplier's numbers have been wrong before, which quiet quarter is the beginning of something — sits on no line at all, in her budget or in anyone's. She read the instruments she was given. The instruments told her to cut.

So I went to check the two years, because if tenure at that scale has fallen to two years, then the argument I have been making for three years writes itself.

It has not fallen. Median tenure with a current employer in the United States was 3.9 years in January 2024, which is higher than it was in 1983, in 1987, in 1998, in 2000 and in 2002. Voluntary quitting is running below where it ran before the pandemic. And the generational story — the one where the young no longer stay — fails its cleanest test: Pew put millennials against Generation X at the same age and found the millennials slightly stickier.

Here is what did happen while the headline held. A man aged 45 to 54 had a median of 12.8 years with his employer in 1983. In 2024 he had 7.5. His slightly older colleague went from 15.3 years to 10.0. Five of the seven age bands BLS reports fell over that period while the total rose, because women's tenure rose and the workforce aged. The average was held up by composition while the long-service tail thinned underneath it.

That is a more interesting finding than the one I went looking for, and a more uncomfortable one, because judgment is a tail phenomenon. The board reading a stable median is reading the one number engineered to conceal the thing it should be watching.

Liga has the data, the two places where a two-year figure is real, and the reason no board can find any of this in a filing.

— Alex

What the Median Hides

Start with the series everyone quotes, and read the whole of it.

The US Bureau of Labor Statistics has measured tenure with a current employer since 1983, in a biennial supplement to the Current Population Survey. Its most recent release, Employee Tenure in 2024, opens like this: "The median number of years that wage and salary workers had been with their current employer was 3.9 years in January 2024, down from 4.1 years in January 2022 and the lowest since January 2002."

Read the last clause closely. Lowest since 2002. The series was lower than 3.9 in January 1983 (3.5), January 1987 (3.4), February 1998 (3.6) and February 2000 (3.5). Across forty-one years it has oscillated between roughly 3.4 and 4.6 with no downward trend, and it moves counter-cyclically: it peaked at 4.6 years in 2012 and 2014, because a frozen labour market destroys short-tenure jobs and everybody who stays gets longer in service by standing still.

The flow measure agrees with the stock measure. In the June 2026 JOLTS release, published on 4 August, the number and rate of quits were "unchanged at 3.2 million and 2.0 percent." Averaged across the first half of 2026, the monthly quits rate in that BLS series runs at 1.97% — below its 2001 average of 2.21% and its 2019 average of 2.32%, and well below the 3.0% peak of April 2022. Short tenure has also become less common: 22% of workers had a year or less with their employer in January 2024, against 24% in January 2022.

Then the generational claim, which is the version of this story that circulates most and survives least. Richard Fry at Pew compared millennials in 2016 with Generation X at the same age in 2000, using the same CPS tenure supplement. On thirteen months or more with a current employer: 63.4% of millennials, against 59.9% of Gen X. On five years or more: 22.0% against 21.8%. Fry's conclusion is that millennials "are just as likely to stick with their employers as their older counterparts in Generation X were when they were young adults." Among college-educated workers the millennials had the longer records.

Now the age table, which is where the story changes.

Age January 1983 January 2024
All, 16 and over 3.5 3.9
25–34 3.0 2.7
35–44 5.2 4.6
45–54 9.5 7.0
55–64 12.2 9.6

Five of seven bands fell. The total rose. Split by sex and the movement becomes plain: men aged 45–54 went from 12.8 years to 7.5, men aged 55–64 from 15.3 to 10.0, while women aged 25 and over went from 4.2 to 4.6. Rising female tenure and an ageing workforce carried the headline upward over a within-cohort decline.

BLS said as much itself in 2000, before anyone was arguing about AI: "All other things being equal, this age shift would have been associated with an increase in median tenure because, until retirement age, tenure tends to increase with age. However, a decline in median tenure for men in most age groups offset the aging effect."

The tail measure is the one to put in a board pack. Among US workers aged 25 and over, the share with ten years or more of service fell from 33.3% in 2014 to 30.2% in 2024. For women it fell from 32.6% to 28.2%. In the 40–44 band, the people an institution is counting on to run it in a decade, it fell from 33.1% to 29.8%.

Europe runs the same arithmetic more cleanly. On the OECD measure — "the length of time workers have been in their current or main job or with their current employer" — EU27 average employee tenure rose from 9.74 years in 1995 to 9.98 in 2025. Underneath it, average tenure for prime-age employees aged 25 to 54 fell from 10.06 years to 8.59, a decline of about 15%. The reason the headline held is visible in one line of the same dataset: employees aged 55 to 64 were 12.0% of EU27 employment in 2010 and 19.5% in 2025.

A European board looking at a flat national tenure figure and concluding that its own bench is intact is reading an average of a workforce that got older, not a measure of how long its people stay.

Where Two Years Is Real

The manager's number exists. It exists in one industry at scale, and in two arithmetic accidents.

The industry is contact centres. Teleperformance, which employed 485,606 people at the end of 2025, discloses the figure in its own registration document: attrition for advisors, "who represent over 80% of the workforce," ran at 5.6% a month in 2025, "i.e. 67% over the year." Group attrition was 5% a month, "i.e. 60% for the year 2025." At 60% annual attrition, implied average tenure is about 1.7 years. At 67% it is about 1.5.

Note the second half of that disclosure, because it is the shape of most large service employers. Supervisors, support and management functions at the same company run at around 2% a month — roughly 24% a year, an implied five years or so. One company, one year, two populations, and a blended average across them that describes nobody in either.

Against that, BLS's industry cut puts the lowest median of any major industry group at 2.1 years, in leisure and hospitality, and the public sector at 6.2 years against 3.5 for the private sector. A two-year median is a front-line service number. It is not what a professional or technical workforce produces, and BLS attributes much of the gap to age distribution rather than to a difference in retention culture.

Set that against the sector most large service employers actually resemble. The IT-services majors publish attrition every year, and in 2025 they ran between roughly 13% and 15%, which implies a tenure of seven years and upward. Capgemini is the useful case, at 423,400 people: it publishes a voluntary figure and a total figure, and the two sit about four and a half points apart, in the region of 15% and 19.5%. Which one gets quoted moves implied tenure by years. The firms also work to different denominators — some exclude involuntary exits, some exclude whole business lines — so a league table built from these numbers compares very little.

The first arithmetic accident is growth. Average tenure in a steady state runs at roughly one divided by the attrition rate, and once a firm is growing, the denominator becomes attrition plus the growth rate. A company that doubles headcount in four years has half its people at under four years of service by construction, with nobody leaving any faster than before. Any employer that has recently grown into a very large headcount will show short average tenure for that reason alone. The same identity runs the other way: an employer that is shrinking shows lengthening tenure while its people are simply unable to move.

The second is the sample. The most widely circulated two-year figure comes from a résumé database — the Indeed Hiring Lab reports that "the median job seeker on Indeed in the US has been with their current employer for two years and three months." That is a median for people who keep a CV on a job board, which is to say people who are looking. It measures something real and it does not measure the workforce.

So the question to put back to a manager quoting two years is not whether she is wrong. It is: two years of what? Mean or median, across which population, over what geography, and how fast has headcount grown? All four move the number by years.

The Manager Is Reading a Correct Instrument

Now the part that matters more than the number.

Her decision is locally rational and she should be assumed to have made it in good faith. The twenty salaries appear in her budget. The tooling appears as capital or as a subscription. The saving appears this year, on a line with her name on it, and she will be assessed on it. Nothing in her reporting pack carries the other side.

What the other side consists of is the accumulated capacity of a group of people to know when something is wrong before they can say why. It is built the only way it is ever built: by doing the ordinary work, repeatedly, in one place, long enough that the ordinary acquires a texture and the wrong thing stands out from it. Automate the ordinary work and the capacity stops forming. Nothing on any statement records that it stopped, because it was expensed the moment it was paid for and was never carried as the asset it was.

IBM is the most useful case available, because IBM automated and then said out loud what the automation could not reach. Its HR system now handles roughly 94% of routine requests. It could not handle the remaining 6%, which is where the judgment calls sit. And IBM tripled its entry-level hiring. Its chief human resources officer, Nickle LaMoreaux, gave the reason: "If we don't continue to invest in entry-level hires, what happens in three to five years. There's no pipeline; the well simply dries up."

That is the argument of this issue, made by a practitioner, at the company that did the automating. It should also be read for what it concedes. The 6% is the residue, and the residue is the hard part — the part nobody knew how to write down. The people who will be asked to handle it in five years are the people being hired, or not hired, now.

Which is where the AI evidence actually sits, and it does not sit where most commentary puts it. The August revision of the Stanford Digital Economy Lab's Canaries in the Coal Mine, published on 12 August with ADP payroll records through June 2026, finds that "employment of young workers (ages 22–25) in AI-exposed occupations now stands 19% below where it would be had it kept pace with that of their less-exposed peers; experienced workers show no comparable gap." The authors are explicit about the mechanism: separation rates fell for both groups, and among young workers fell at least as much in the most exposed occupations as in the least exposed. The divergence runs through hiring.

Read that against the tenure data and a coherent picture forms, on a longer clock than the one the manager was working to. The observable AI shock so far is a hiring shock. It closes the entrance to the pipeline rather than accelerating the exit from it. That is not a milder finding. It is a slower one, and it will surface fifteen or twenty years after the decisions that caused it, in a cohort of people who were never formed and cannot be back-filled at any price.

There is an answer to this in circulation, and it deserves to be taken seriously because it comes from the field best placed to have thought about it. The answer is that people will form themselves across a portfolio of settings, arriving able to read an organisation quickly, and will pull whatever technical knowledge each setting needs from a machine that holds it. The social half of that is right. The technical half assumes that domain judgment is a stock of retrievable facts. What a domain expert carries is the formed capacity to feel that the fluent answer in front of them is wrong — and that capacity cannot be fetched from the machine, because judging the machine is its entire job. An organisation hiring the finished article has bought judgment formed at another firm's expense, and is betting that the other firm is still running the apprenticeship.

The Asset the Disclosure Regime Declined to Make Visible

If a board wanted to manage this, it would start by looking at what its peers report. There is almost nothing to look at.

Gibson Dunn's five-year review of Form 10-K human capital disclosure, published in January and covering the S&P 100, found that while general statements about attracting and retaining talent appear almost everywhere, "only 19% of companies surveyed provided specific employee turnover rates (whether voluntary or involuntary), consistent with the 20% of companies that did so in the previous three years." Roughly one large US issuer in five puts any number on it, and the share has been flat for four years.

Tenure fares worse. Our own full-text search of EDGAR's entire 10-K index, back to 2001, for the phrase "median tenure" returns seven documents. Every one of them refers to somebody else's median — a senior management team's, a board of independent directors', another industry's workforce. Not one US issuer has disclosed the median tenure of its own workforce in a Form 10-K. Boards get medians. Executives get medians. The people who do the work do not.

This was noticed, and by people who are difficult to dismiss. In June 2022 a working group including two former SEC Commissioners and a former SEC General Counsel petitioned the Commission for human-capital disclosure, and put mean tenure and employee turnover in the top rows of the grid they proposed. Their reasoning is the argument of this newsletter written in accounting language: "if a firm has investments in labor of $100,000 and employees typically remain at the firm for five years, investors might amortize that $100,000 at $20,000 per year for five years. Therefore, we include turnover and tenure here both to allow for calculation of amortization and because these metrics may well be sufficiently important to warrant disclosure on their own."

Tenure as the amortisation schedule for a firm's investment in its own people. That is what the number is for, and it is why the aggregate cannot substitute for the distribution.

The Commission's own expansion item, RIN 3235-AM88, sat on the Unified Agenda from 2022 through 2024 at proposed-rule stage with a slipping date. In the Spring 2025 edition its entry reads: "This item is being withdrawn from the Agenda." The withdrawal is dated 21 April 2025. The item is absent from the current agenda. The petition remains pending and has produced nothing.

So the position, stated plainly: an asset that determines whether an organisation can supervise its own automation is expensed as it is paid for, absent from the balance sheet, absent from the risk register, disclosed by one large issuer in five in the crudest available form, and never disclosed at all in the form that would show the erosion. The regime was asked to make it visible in 2022. It declined in 2025.

What This Means for Boards Right Now

One. Stop asking for the average and start asking for the distribution. A median or a mean tenure figure moves with headcount growth, with the age mix, and with whether the labour market is frozen — three things that have nothing to do with whether the organisation is accumulating judgment. The cut that carries information is the tail: how many people in each critical function have ten or more years here, what that share was five and ten years ago, and how it looks in the functions where a wrong call is expensive. In the US series that share fell from 33.3% to 30.2% across a decade while the median barely moved. If management cannot produce the equivalent for this company, that is itself the finding.

Two. Give the loss an owner, because the saving already has one. Every automation case that reaches a board arrives with a headcount number attached to a named budget holder. Almost none arrives with a statement of what capability the affected work was forming, who now forms it, and by when. Require the second document beside the first, signed by the same person. A manager who has to write the sentence "this work forms nothing that we need" will either write it and be right, or discover while writing it that she cannot.

Three. Watch the entrance, not the exit. The evidence through June 2026 says the AI adjustment is running through hiring, with young workers in exposed occupations 19% below their counterfactual and experienced workers showing no comparable gap. A retention dashboard will show nothing, because nobody is leaving faster. The number to bring to the table is the count of entry-level people hired into each function this year against three years ago, and the answer to the question that follows from it: who, in 2040, will be able to tell that this system is confidently wrong?

Which leaves the question for the next meeting, and it is one line: our median tenure looks stable — how many people here have been doing this for ten years, and how many had that a decade ago?

Reading List

What We Are Watching Next

  • Whether the BLS publishes a fresh Employee Tenure supplement, and what the male 45–64 bands do when it arrives — those are the two bands that have halved since 1983
  • Whether any S&P 100 issuer becomes the first to disclose the median tenure of its own workforce, in any filing
  • Whether the SEC's next Unified Agenda restores a human capital management item after the April 2025 withdrawal, and whether the 2022 petition draws any response at all
  • Whether the next revision of the Stanford entry-level series shows the 19% gap widening, closing or holding, and whether any European dataset reproduces it
  • Whether any large European employer publishes tenure by age band, which is the only cut that would show a European board what its own flat headline is concealing

The next issue goes deeper into one of these. If you want a specific function or sector covered, reply to this email.

— Liga

TwinLadder Weekly is a weekly intelligence report on judgment, governance, and the boards accountable for both. Subscribe at twinladder.ai/newsletter. Forward this issue freely.

Izdevums #48Izdevums #50
Atpakaļ uz visiem izdevumiem