Computer programmers over 55 are now leaving work at a predicted rate more than 25% higher than before ChatGPT. House painters โ bottom of the same exposure index โ moved about 2% (Sanzenbacher, CRR Issue Brief 26-13, 2026).
That is the experience drain, and it runs in the opposite direction from every automation story you have been told. The higher the pay, the higher the skill, the faster the exit.
Most mid-market operations teams have spent eighteen months worrying about the entry-level on-ramp โ whether AI is eating the junior analyst job, whether the 22-year-old hire still has a path. That concern is well-evidenced. It is also only one end of a pipeline that is now leaking at both.
What the Boston College Brief Actually Measured
The June 2026 Issue Brief from the Center for Retirement Research at Boston College, authored by economist Geoffrey T. Sanzenbacher, matched longitudinal Current Population Survey data to the Tufts Digital Planet AI-exposure index, then compared job-exit behavior for workers aged 55 and older before and after generative AI went mainstream in November 2022 (Center for Retirement Research at Boston College, 2026).
Two design choices make this worth your attention.
First, it uses observed transitions out of work, not stated retirement intentions. Survey research on older workers and AI captures anxiety. This captures departure.
Second, it ranks occupations by measured exposure rather than by intuition about which jobs "feel" automatable (Digital Planet, Tufts University, 2026). Programmers, accountants, and analysts sit high. Painters sit low. The gradient is the finding.
For the most exposed white-collar occupations, the predicted probability of transitioning out of work rose from roughly 8.7% to 11.1% โ a relative increase above 25%. At the low-exposure end, the same calculation produces about 2%. The effect is largest in the higher-paying jobs, which inverts the pattern automation has followed for four decades.
The Longevity Advantage That Just Inverted
Here is the number that should reframe your workforce plan.
Across the full period studied, high-exposure older workers historically had lower exit rates than their low-exposure peers: 11.7% versus 14.1% (Sanzenbacher, CRR Issue Brief 26-13, 2026). Skilled desk work was the thing that let people keep working. The accountant stayed; the roofer's body gave out.
That advantage is eroding, and the brief notes a significant share of the change traces to unemployment rather than voluntary retirement. This is not a wave of comfortable people choosing to leave early. It is a group whose careers were previously the most durable becoming less so.
Note also the adoption picture underneath it: only about 18% of workers aged 50โ64 report using generative AI. The exposure is landing on the segment least likely to be using the tool that creates it.
For an operations leader, that combination โ high exposure, low adoption, rising involuntary exit โ is a specific, addressable condition. It is not a demographic inevitability.
It is also worth being precise about what "exit" means here. A transition out of work at 58 is not the same event as a transition out of work at 28. The older worker rarely re-enters at the same level; the CRR framing is that these are careers being cut short, not paused (Sanzenbacher, CRR, 2026). From your side of the table, that means the person does not come back to the market for you to rehire in two quarters when the pilot underdelivers. The knowledge leaves the sector, not just the payroll.
The Experience Drain Hits the Roles You Backfill Slowest
Set the Boston College finding next to the best evidence from the other end of the age distribution.
Stanford's Digital Economy Lab, working with high-frequency payroll data covering millions of US workers, found a roughly 13% relative decline in employment for workers aged 22โ25 in the most AI-exposed occupations after generative AI adoption, with employment for older workers in those same occupations holding up or growing (Stanford Digital Economy Lab, 2025).
Read naively, the two results contradict each other. They do not. Stanford measures headcount in the occupation; Boston College measures individual transitions out of work for the 55+ cohort. One is a hiring signal, the other an exit signal โ and they can both be true at once. The occupation can hold its aggregate headcount while the specific people at the senior end of it churn out faster.
The operational consequence is what matters, and it is uncomfortable: fewer juniors entering the funnel, faster senior departures out of it, and the mid-market's thinnest layer in between. A 20,000-person company absorbs this. It has a bench, a rotation program, and three people who could plausibly step into any given role.
At 200 people, the person who knows why the reconciliation logic has that exception is one person.
What Walks Out the Door With Them
The cost of a senior exit is not the salary line you stop paying. It is two things you never had on a line at all.
The first is tacit process knowledge โ the undocumented reasons a system is configured the way it is. Every mid-market company runs on a layer of this, and it is invisible precisely until the person holding it leaves. It is the exception in the revenue-recognition logic that exists because of a 2019 contract, the vendor who must be called rather than emailed, the query that looks redundant and is not. None of it is in the handbook. Most of it is not in anyone's head twice.
There is a second-order version of this that AI makes sharper rather than softer. Documentation and code assistants are good at describing what a system does. They are poor at recovering why it was built that way, because the reasoning was never written down to be retrieved. The tooling that raises your exposure does not close the gap it opens.
The second is measurable, and it is the one most operations plans ignore. Working alongside more-skilled colleagues produces real human-capital gains for the less-skilled worker โ effects large enough to show up in subsequent wages, and consistent across independent research groups (Herkenhoff, Lise, Menzio & Phillips, Econometrica, 2024). Your senior people are not just producing output. They are a training input for everyone in proximity to them.
Lose that input while simultaneously hiring fewer juniors, and the compounding runs the wrong way. You are not merely down a headcount. You have reduced the rate at which the remaining team gets better.
Where the exposure concentrates
Programmers, accountants, financial analysts, and technical writers are the profile most affected โ which in a 50โ500 FTE company means finance, engineering, and the small internal-systems function that keeps the stack coherent. These are also, not coincidentally, the roles with the longest time-to-productivity for a replacement.
Where This Argument Could Be Wrong
Three caveats, stated plainly, because the author states them himself.
This is an early analysis of a recent inflection. The post-ChatGPT window is short, and estimates over short windows are fragile.
Confounds exist and are named in the brief: cuts to government R&D spending and unusual hiring dynamics at AI-native startups could both bias the estimates for exactly the high-exposure occupations under study (Center for Retirement Research at Boston College, 2026). Some of the measured exit may belong to those forces rather than to AI displacement.
And exposure is not the same as replacement. A high score on an exposure index means a role's tasks overlap with model capability. It does not establish that a model did the job.
What survives all three: the direction is the load-bearing claim, not the decimal. A longevity advantage for high-skill older workers that has held for decades is narrowing, and it is narrowing fastest where exposure is highest. Even a substantially attenuated version of that finding still points your risk at the same segment.
What Changes in a 50โ500 FTE Company
Three moves. None require budget approval.
Run an exposure-by-tenure crosstab, once. Pull your headcount, tag each role high or low AI-exposure using the occupational categories in the brief, and cross it with age or tenure. You are looking for a single number: how much of your most-exposed work is held by people in the last decade of their career. Most operations leaders have never computed this and are surprised by it.
Convert proximity into documentation before it becomes urgent. The coworker-learning effect is real but requires presence (Herkenhoff, Lise, Menzio & Phillips, Econometrica, 2024). Where your exposure crosstab flags a concentration, fund the extraction now โ a documented runbook, a recorded walkthrough, a named second person on the process. This is cheap while the person is still employed and impossible afterward.
Fix the adoption gap in that cohort specifically. With around 18% generative-AI usage among workers 50โ64, the group carrying the most exposure has the least fluency with the thing creating it. Targeted enablement for your senior technical staff is the rare intervention that reduces displacement risk and raises output at the same time โ and it is considerably cheaper than replacing them.
One Decision This Quarter
Look at your org chart and find the roles that are simultaneously high-exposure and held by someone within five years of retirement. Then answer one question honestly: if that person gave notice on Friday, how long until someone else could do the job โ not the tasks, the job?
If the answer is measured in quarters, you have an unpriced liability, and the Boston College data says the probability attached to it just went up.
Every AI workforce plan written in the mid-market this year protects the on-ramp. Almost none protect the off-ramp. The experience drain is not a retirement problem for HR to manage in 2030 โ it is an operations problem, and it started in 2023.
Budget for the exit you are not expecting, from the person you cannot replace.
