Between November 2022 and September 2025, the job-finding rate for the most strongly attached segment of the U.S. workforce fell 13 percentage points. For the churn-prone segment that normally absorbs cyclical shocks, it fell 2 (Federal Reserve Bank of Richmond, 2026). That gap has no precedent in the data the authors examined โ not in the Great Recession, not in any expansion. And the sharpest declines since 2023 sit in the occupational quartile most exposed to AI: programmers, financial analysts, engineers.
If your 2026 workforce plan assumes the entry-level pipeline is the fragile part and your experienced bench is replaceable at market rate, the plan is built on an inverted premise.
The Entry-Level Collapse Is Real. It Is Not Where the Damage Is.
Economic Brief 26-26, published August 12 by Katarรญna Boroviฤkovรก and Claudia Macaluso, does something most labor-market commentary skips: it disaggregates. The recent rise in unemployment, the authors had already established, is almost entirely an outflow story โ people are not losing jobs at unusual rates, they are failing to find new ones. The open question was which people.
The first cut is the familiar one. The Current Population Survey separates job losers, job leavers, and entrants. The dominant narrative of the past two years โ that AI is closing the bottom rung of the career ladder โ predicts that entrants take the worst of it. The data says otherwise. Job-finding rates for new entrants have declined, but the authors describe the drop as modest relative to the declines among job losers and job leavers, and they explicitly set entrants aside for the rest of the analysis (Richmond Fed, 2026).
That is a narrow finding and worth stating precisely: it does not mean early-career hiring is healthy. It means entrants are not the margin explaining the aggregate collapse. The action is elsewhere.
Attachment, Not Tenure: What "Primary Type" Actually Measures
The second cut is the one that should change how you read your own attrition dashboard.
The authors apply a latent-type framework from "The Dual U.S. Labor Market Uncovered" by Hie Joo Ahn, Bart Hobijn and Ayลegรผl ลahin (NBER, 2023), which sorts workers by how attached they are to the labor market rather than by any observable credential:
- Primary type โ roughly 55% of the population, almost always employed, stable and long-tenure
- Secondary type โ roughly 14%, strongly attached but unemployed more often
- Tertiary type โ roughly 31%, weakly attached, frequently out of the labor force
In every ordinary cycle, the secondary type carries the adjustment. It accounts for the bulk of unemployment in essentially every year of the data, despite being the smallest population share. Recessions hit the primary type harder in rate terms โ in the Great Recession, primary-type job-finding fell 19 points peak-to-trough against 10 for secondary โ but the ratio stayed within recognizable bounds.
The current episode broke the ratio. Thirteen points against two. The people who almost never need to look for work are the ones who now cannot find it.
One caution on translation: "primary type" is a statistical classification of employment stability, not a synonym for seniority. It correlates with the profile of your long-tenure analysts, engineers and finance staff, but it is not a direct measure of them. Treat the finding as a strong directional signal about that layer, not a coefficient you can apply to a specific headcount.
The AI-Exposure Divergence Has a Start Date: 2023
The third cut is where the mechanism gets a name. Following Michael Webb's method of scoring occupations by how much of their task content overlaps with the technical capabilities described in AI patents (Webb, 2019), the authors sort workers into AI-exposure quartiles.
Historically, outflow rates across those quartiles moved together with the business cycle, separated only by small level differences. Since 2023 they have diverged, and workers in the most exposed occupations show the largest declines in job-finding rates.
The composition of that top quartile is the part mid-market operations leaders should sit with. Highly exposed occupations include computer programmers, financial analysts and engineers. Least exposed: construction, food service, personal care. This is not a story about the periphery of your org chart. It is a story about the roles that carry your systems, your close, and your product.
There is a prior literature that anticipates this shape. Acemoglu and Restrepo's work on industrial robots found displacement concentrated in exposed occupations alongside reallocation toward less-exposed work โ a pattern in which the aggregate employment number stays legible while the composition underneath it shifts hard (Acemoglu and Restrepo, 2020). What the Richmond Fed data adds is that the current episode is showing up in the outflow margin rather than the separation margin. Firms are not shedding these workers unusually fast. They are simply not absorbing them once they are out.
That distinction matters operationally, because it means the signal will not appear in your layoff tracker or your industry's headcount announcements. It appears only in how long it takes a competent, experienced person to land โ a metric almost no operations team measures, and one that describes your own re-hiring conditions as much as anyone else's.
The authors are careful about causality, and so should we be. The pattern is consistent with automation displacing exposed occupations, and consistent with field evidence that generative AI raises task-level productivity โ particularly for less-experienced workers โ in ways that could reshape hiring composition (Brynjolfsson, Li and Raymond, 2025). Whether the decline is driven by AI specifically or by aggregate conditions that happen to load on the same occupations is the explicit subject of the next brief in the series. The correlation is documented; the decomposition is pending.
The Objection: "This Is Macro Data. My Attrition Is Fine."
It probably is. That is the problem.
Two mechanics operate here, and both invert standard reading.
Flat attrition in your senior layer is a weak outside option, not an engagement win
If your experienced, AI-exposed staff are staying, the base rate now offers a second explanation alongside the flattering one: they cannot leave. A 13-point drop in job-finding probability for strongly attached workers means the external market that normally validates your retention is not clearing. Your engagement survey will not distinguish between "committed" and "stuck." Your regretted-attrition number will look excellent right up until job-finding rates recover, and then it will not.
This is not a reason to distrust your people. It is a reason to stop using attrition as your primary evidence that the senior layer is healthy, and to instead read something with independent signal โ internal mobility applications, manager-reported flight risk, or the ratio of external offers your people disclose.
Severance is now more expensive than it prices
The standard restructuring model treats a role as re-acquirable: cut now, rehire at market when conditions change. That model prices re-acquisition off historical job-finding dynamics, which is exactly what this brief says has broken. A worker who is hard to re-hire externally is, symmetrically, hard to re-acquire โ the same friction that traps them also traps your ability to reconstitute the capability later, because the pool you would rehire from is the pool that is not moving.
For an AI-exposed senior role in a 200-FTE company, the question this quarter is not "can we operate without this seat." It is "if we need this capability back in eighteen months, what does the market look like then, and what did we pay to find out." Internal redeployment โ moving the person to an adjacent process, a data-governance function, or an AI-oversight role โ beats severance on that math far more often than the spreadsheet suggests, because the spreadsheet has a stale re-acquisition cost in it.
Three Decisions Available Before the Quarter Closes
- Re-price the re-acquisition cost in every AI-exposed severance case. Not a philosophical shift โ a number. If your model assumes a 60-day backfill at market for a senior analyst or engineer, ask what evidence supports that assumption given a 13-point drop in job-finding for that profile. Where it does not hold, redeployment moves ahead of separation on cost alone. Run it on the two or three roles you have already flagged for review; you do not need a policy, you need a corrected number in the cases actually in front of you.
- Stop reading flat senior attrition as retention performance. Add one question to your quarterly review: what is our independent evidence that this layer is staying by choice? If the answer is only "they haven't left," you do not have evidence, you have a market condition.
- Audit which roles sit in the exposed quartile before you plan headcount, not after. Webb's classification is public and occupation-level. Mapping your own org chart against it takes an afternoon and tells you which parts of your workforce plan are exposed to a re-hiring market that is not functioning normally.
None of these require a position on whether AI is causing the divergence. They only require accepting that the divergence exists and that it is concentrated in the layer most plans treat as fungible.
The Window Is the Finding
The uncomfortable feature of this data is that it is temporary in a way that punishes patience. A depressed job-finding rate for AI-exposed workers gives you an unusual amount of holding power over exactly the people who are hardest to replace โ and it gives you that power for free, invisibly, without any decision on your part. It also expires. When outflow rates normalize, the retention you thought you had earned resolves back into a market, and it resolves fastest in the quartile you are least prepared to lose.
The decision for this quarter is not whether to cut. It is whether the roles you are treating as fungible are the ones the labor market has quietly stopped supplying โ and whether you would rather find that out now, at the cost of an afternoon's mapping, or in eighteen months at the cost of a search.
