The attrition buffer is 3.2%/yr for a median occupation and 1.4%/yr for software developers, and BLS projects no occupation crossing it
The question
Verbatim: "What is the empirical distribution of annual occupational exit rates (retirements + voluntary separations) across US occupations, and at what decline rate does measurable incumbent harm begin?"
A calibration follow-up to [[2026-07-28-technological-displacement-career-vs-generation]], which concluded that the predictor of within-career vs across-generation absorption is an occupation's annual decline rate measured against natural workforce attrition, then supplied only an ordering and explicitly flagged the missing benchmark: "the benchmark attrition rate is the number the whole hypothesis pivots on, and I did not fetch a source for it." This fetches it.
Short answer: the exit-rate distribution is now pinned precisely, with the caveat that BLS's rates are model projections for 2025-35 rather than observed exits. The cut point is a band, roughly 3-4%/yr for a median occupation, and it must be computed per-occupation because the spread is wide. A mechanical route and an econometric route both land in that band, though the econometric one's position depends on a harm anchor I chose. The firmest result is that zero of 831 US occupations are projected to cross the mechanical line over 2025-35.
What we already know (from the vault)
- [[2026-07-28-technological-displacement-career-vs-generation]] established the mechanism and the ordering: agriculture 1910-40 at -0.77%/yr produced no visible incumbent harm; tellers 2016-24 at -4.5%/yr moderate harm; newspaper publishing at -5.5%/yr severe; travel agencies 1998-2003 at -12.4%/yr severe. Those are realized rates from that brief's sources, not from the series used here.
- The same brief's second predictor was whether the move to the successor job was administrative (same employer, no job search) or market-mediated. That channel is orthogonal to the rate test and is not re-examined here.
- [[2026-07-27-psf-pyramid-economics-post-agent-analyst-tier]] is this question applied to one occupation: the scarce layer moved rather than the pyramid collapsing, which is the signature of entry-cohort rather than incumbent harm.
- [[2026-05-18-every-start-career-when-ai-doing-entry-level-job]] carries the Stanford finding of a 13% employment drop for 22-25-year-olds in AI-vulnerable roles, i.e. observed harm concentrated in the entry cohort. That note records only the 13% figure, so the usual accompanying claim that incumbents held is not sourced here and would need the underlying Stanford Digital Economy Lab paper to stand.
- [[2026-09-01-cpa-bookkeeper-aging-cohort-tooling-spend]] documented the opposite case: an occupation whose incumbent cohort is old enough that attrition is unusually high, which on this brief's logic widens the buffer.
What the web says
All figures below are from BLS Employment Projections, Table 1.10 "Occupational separations and openings, projected 2025-35" (bls.gov/emp/tables/occupational-separations-and-openings.htm; page last modified 2026-08-27; employment in thousands). Rates are BLS's own 2025-35 annual averages. I pulled all 1,112 rows and analysed the 831 rows BLS flags as Line item (detailed occupations), discarding the 281 Summary aggregate rows to avoid double-counting.
- BLS-published, all occupations (SOC 00-0000): employment 170,280.8 (2025) to 176,198.5 (2035), +3.5% over ten years. Labor force exit rate 4.2%/yr. Occupational transfer rate 5.5%/yr. Total occupational separations rate 9.7%/yr.
- Derived by me, distribution across the 831 line-item occupations. Labor force exit rate: min 1.3, p10 1.8, p25 2.35, median 3.2, p75 4.3, p90 5.8, p95 7.05, max 13.2 (%/yr). Occupational transfer rate: median 4.9, p10 2.9, p90 7.4. Total separations rate: median 8.3, p10 4.9, p75 10.4, p90 13.0. Parse validation: my employment-weighted means are 4.26 / 5.50 / 9.75, reproducing BLS's published 4.2 / 5.5 / 9.7 to within 0.06 points, so the extraction is sound.
- Derived by me, projected decline rates. Annualising each occupation's published 2025-35 percent change as
(1 + pct/100)^(1/10) - 1: 232 of 831 line items (27.9%, holding 43,431.5 thousand workers = 25.5% of line-item employment) are projected to decline at all. Among those decliners the median decline is only -0.47%/yr; p25 -0.87; the fastest 5% exceed -1.89; the single fastest is -4.13%/yr (word processors and typists, 43-9022). - Derived by me, the central negative result. Zero of 831 occupations are projected to decline faster than their own labor-force-exit rate. The narrowest margins are foundry mold and coremakers (51-4071, -2.58%/yr against a 3.1% exit rate, margin 0.52 points) and carpet installers (47-2041, -1.78 against 2.3, margin 0.52). The four fastest decliners all clear their own exit rate with room: word processors and typists -4.13 against 5.0; telephone operators -3.18 against 4.6; switchboard operators -2.97 against 7.1; data entry keyers -2.90 against 4.1. Against the looser total separations benchmark the count is also zero.
- The harm side comes from Edin, Evans, Graetz, Hernnäs & Michaels, "Individual Consequences of Occupational Decline" (working-paper version dated 2022-11-09, published in The Economic Journal, 2023; Swedish administrative data, outcomes 1986-2013). They define a declining occupation as one whose matched US Occupational Outlook Handbook employment growth was negative and exceeded 25% in absolute magnitude, measured on OOH employment 1984-2016 (32 years), choosing 25% because "smaller observed declines may result from measurement error." Derived by me: -25% over 32 years annualises to -0.89%/yr, or -0.95%/yr on the 1986-2016 dating used in their Table 1 note. So their classification cutoff sits at about -0.9%/yr. That cutoff is a crosswalk-error guard, not a level below which harm becomes unmeasurable, and it should not be read as a harm floor. At or beyond it, incumbents lost 2-5% of mean cumulative pre-tax earnings and 1-2% of cumulative employment versus comparable workers; those starting in the bottom tercile of their occupation's 1985 earnings distribution lost 8-11% of bottom-tercile mean earnings. 13% of Swedish employees in 1985 were in subsequently declining occupations, 8% in technology-driven ones.
- Their mechanism statement is the parent brief's hypothesis in the authors' own words: mean losses are below the mass-layoff literature's because "occupational decline is typically gradual, and can be partly managed through retirements, reduced entry into declining occupations, and increased job-to-job exits to other occupations." They also report an elasticity of earnings losses with respect to occupational employment growth of 0.08-0.13, an employment-loss elasticity of 0.04-0.05, and an elasticity of still-remaining-in-the-1985-occupation of 0.71-0.95. Three properties of those elasticities matter for anything built on them. They are computed by "taking the reduced form estimates reported in Tables 4 and 5 (expressed in percent of the outcome mean) and dividing them by the difference in occupational employment growth between declining and non-declining occupations reported in Table 1," so the denominator is Swedish employment growth relative to non-declining occupations over 1985-2013, not absolute decline. The authors state the elasticities "are likely upper bounds on the true elasticities" because measured Swedish employment change is biased toward zero by occupations being absorbed into broader SSYK96 categories. And the 0.71-0.95 figure is for remaining in the initial occupation by any route, bundling transfers with exits from employment.
Convergences and contradictions
- Two routes land in the same band, but only one of them is anchor-free. Route A, mechanical: the decline cannot be absorbed by simply not replacing people who were leaving the labour force anyway once the annual decline exceeds the occupation's own labor-force-exit rate, whose median is 3.2%/yr (IQR 2.35-4.3). Route B, econometric: applying Edin et al.'s 0.08-0.13 elasticity, a
glog-point relative growth shortfall over their 1985-2013 window implies a0.08gto0.13gpercent cumulative earnings loss, so under a 10% earnings-loss anchor (my choice, not published) reaching that loss needs 77-125 log points, i.e. 2.75-4.46 %/yr sustained over 28 years. Route B's position is set entirely by that anchor: a 5% anchor puts it at 1.4-2.2%/yr, a 15% anchor at 4.1-6.7%/yr. So the overlap with Route A at roughly 2.8-4.3 %/yr is anchor-dependent and is not independent corroboration. - No internal validation of Route B is available, and the obvious candidate for one is circular. Recovering Edin et al.'s 2-5% headline by applying their own elasticity to their own declining group is true by construction, because the authors define the elasticity as that loss divided by the declining-vs-non-declining growth gap. Anyone sanity-checking this route should not count that as corroboration. Two further complications push in opposite directions and cannot be netted out without the comparison group's growth rate: the denominator is a relative gap, so a given absolute decline implies a larger shortfall and a lower required rate, while the authors' own statement that the elasticities are upper bounds implies a smaller true elasticity and therefore a higher required rate. The net effect is that Route B's band is wider than 2.75-4.46, not that it sits somewhere else.
- The vault's realized cases are consistent with the band, but the comparison is weaker than it looks. Tellers at a realized -4.5%/yr in the parent brief against a 4.6%/yr exit rate in Table 1.10 is a dead heat, and the parent recorded "moderate, ongoing" harm. Newspaper publishing at -5.5%/yr and travel agencies at -12.4%/yr (flagged "[unverified]" in the parent) are past every plausible edge, and the parent recorded "severe." Agriculture at -0.77%/yr is below Edin et al.'s classification cutoff, and the parent recorded none visible. Two caveats: these mix realized declines from the 1910s through 2024 against projected 2025-35 exit rates, which is a vintage mismatch; and any threshold between roughly 0.8 and 4.5 %/yr would reproduce a four-case ordering, so this is a sanity check and not evidence for the band's location.
- The contradiction between realized and projected rates is the most important caveat in this brief. Table 1.10's single fastest projected decline over a full decade, -4.13%/yr, still falls short of that occupation's own 5.0% exit rate. BLS projections are trend-extrapolative and smooth. Realized history reaches much further: travel agencies at -12.4%/yr is three times the largest number in BLS's entire 831-occupation forward universe. So anyone citing BLS Employment Projections as evidence that AI will displace incumbent workers is citing a dataset whose maximum projected decline, in the 2025-35 vintage, is below the attrition line. Whether that holds across vintages as a property of the method is an open question and is not established here.
- Edin et al. pre-empt the obvious objection to transferring their number to an AI shock. They note their declines were "fairly gradual" and that "if, for example, machine learning improves rapidly, occupational replacement may happen faster." Their estimate is therefore a lower bound on harm for a faster shock, not a forecast of it.
Synthesis for RDCO
The cut point is now a usable number, and it is a band rather than a point: roughly 3-4%/yr for a median occupation, with the honest operating rule being to compare an occupation's annualized decline to its own labor-force-exit rate from Table 1.10 rather than to any single economy-wide constant. The spread is too wide for a constant to be safe. The labor-force-exit rate runs from 1.3%/yr to 13.2%/yr across occupations, a tenfold range, because BLS builds it from a regression on worker characteristics such as age and educational attainment applied to each occupation's current employment mix. Using the all-occupations 4.2% as a universal benchmark would overstate the buffer for a young, highly-educated occupation by a factor of three.
That matters directly, because the founder's own lane sits among the thinnest attrition buffers in the economy. Software developers (15-1252) have a labor-force-exit rate of 1.4%/yr, at the first percentile of the distribution; only 8 of 831 occupations are lower, and the minimum is 1.3%/yr. Computer systems analysts are 2.2%, data scientists 1.7%, computer programmers 1.9%. Compare clerical: switchboard operators 7.1%, customer service representatives 5.0%, word processors 5.0%, secretaries 5.0%. The clerical occupations everyone points at as AI's first casualties have buffers three to five times wider than software does. A sustained decline of 2%/yr would be comfortably absorbed by non-replacement in customer service and would force incumbent software developers out. This inverts the intuitive reading of occupational exposure: being in a young, credentialed, low-turnover occupation is a liability under the rate test, because there are almost no retirements to absorb the shrinkage. It is also a caution against reading Table 1.10's current numbers as reassurance, since software developers are projected at +0.98%/yr and data scientists at +3.02%/yr, so the buffer is not currently being tested.
The second structural finding resolves the parent brief's framing more cleanly than the parent did: the across-generation channel has no threshold at all, while only the within-career channel has one. BLS's openings arithmetic makes this exact. From the published all-occupations row, 16,870.0 thousand separations plus 591.8 thousand of annual growth equals 17,461.8 thousand openings, matching the published openings figure to the decimal, which confirms openings are separations plus employment change. So for a declining occupation, openings are separations minus the decline. Tellers at -1.41%/yr with a 9.9% total separations rate still post about 8.5%/yr in openings, meaning an incumbent who wants to stay can stay, while the entry rate is about 14% lower than it would be in a flat occupation. Entry-cohort harm therefore scales smoothly from the first fraction of a percent of decline, with no cut point to find, which is why the Stanford 13% drop for 22-25-year-olds in [[2026-05-18-every-start-career-when-ai-doing-entry-level-job]] shows up long before incumbents are forced to move. The precise statement is narrower than "incumbent harm has a threshold": forced incumbent movement has a threshold at the exit rate, whereas measurable incumbent earnings harm appears well below it, since Edin et al. detect 2-5% cumulative losses at a relative decline of roughly 0.8-1.6 log points per year. So the exit rate bounds displacement, not harm.
On the arithmetic trap the dispatch flagged, the answer is that labor-force exits are the right denominator and total separations are not, and the reason is endogeneity rather than definitional tidiness. BLS splits separations into labor-force exits (leaving the occupation and the labour force, "due to retirement or other reasons") and occupational transfers (moving to a different occupation, counted as a permanent separation even though the worker stays employed). The case for exits rests on real-world worker behaviour, not on how BLS builds the numbers: BLS's methodology page describes both the exit model and the transfer model the same way, as a regression on worker characteristics such as age and educational attainment applied to the occupation's current employment distribution, so inside Table 1.10 the projected transfer rate is no more mechanically tied to projected decline than the exit rate is. What separates them is that leaving the labour force is driven mainly by age and life stage, whereas moving to another occupation is exactly the behaviour a demand decline induces. Edin et al. measure that response directly: their 0.71-0.95 elasticity for no longer working in the 1985 occupation shows decline pushing people out of the occupation, though by any route, since the measure bundles transfers with exits from employment. Using the 9.7%/yr total separations rate as the natural-attrition benchmark would inflate the benchmark precisely in the occupations under stress and would build the displacement being tested for into the test's own denominator. Neither series is a JOLTS-style quit rate, and neither counts job changes within the same occupation, so they must not be mixed with JOLTS turnover figures.
Confidence, calibrated. High on the extraction of the exit-rate distribution: it is a complete census of the 831 line items and my employment-weighted means reproduce BLS's published aggregates to within 0.06 points. But note what the distribution is: these are BLS regression projections of exit rates for 2025-35, not observed exits, so the question's request for an empirical distribution is answered with modelled rates and that substitution should travel with the number. High on the negative result that no projected occupation crosses its own exit rate over 2025-35, since that is a complete enumeration of the table. Low-to-medium on the 3-4%/yr band, downgraded from medium because Route B turns out to carry no independent validation. Low on any claim that the band transfers to an AI shock, for the reason Edin et al. give themselves.
"Measurable incumbent harm" is not a BLS construct and I had to operationalize it, which is an assumption and not a finding. I used two definitions. Mechanically, that the decline exceeds the occupation's own labor-force-exit rate, which is a necessary condition for forced incumbent movement and says nothing about welfare. In welfare terms, that mean cumulative career earnings losses reach about 10%. That 10% anchor is mine and is not published. The tempting justification, that 10% is the floor of the mass-layoff literature, does not survive checking: the mass-layoff figures Edin et al. cite are 4-6% of annual earnings over the 5-10 years after displacement in Sweden and 7-14% in the US, which are not in cumulative-career units and therefore cannot pin a cumulative-career anchor. The honest justification is weaker, that 10% is a round number comfortably above their 2-5% mean. The band is correspondingly sensitive: a 5% anchor moves it to 1.4-2.2%/yr, a 15% anchor to 4.1-6.7%/yr. That sensitivity is the real width of the answer and is why this brief reports a band and a per-occupation rule rather than manufacturing a single threshold. Three further limits on the elasticity route: the authors state their elasticities are likely upper bounds, which on its own pushes the required rate above 4.46%/yr; the denominator is a relative rather than absolute growth gap, which pushes it the other way; and it is Swedish data for pre-2013 technology. The harm function may also be convex, so that the linear extrapolation understates harm at high decline rates, but that is my conjecture and is not in the source.
Why this is in the vault
This supplies the one number [[2026-07-28-technological-displacement-career-vs-generation]] flagged as missing, which converts that brief's rate-vs-attrition ordering into a computable test, and it changes the read on the phData cert-escalator bet ([[project_phdata_cert_escalator_path]]) in a specific direction: the parent justified that bet on administrative-position grounds, and this brief adds that the software and data occupations the bet sits inside are among the thinnest-buffered in the economy (1.4-2.2%/yr, spanning the 1st percentile to roughly the 25th), so the administrative channel is doing more of the work than previously credited and skill adjacency even less. It also hands the credibility-for-phData-sales positioning argument a defensible number to lead with, in place of the usual move of citing BLS projections, which this brief shows do not support an incumbent-displacement claim anywhere in the 2025-35 vintage.
Open follow-ups
- Does BLS systematically understate decline in the fastest-declining decile of occupations at a ten-year horizon? The whole cut-point test is unusable on projected rates if the answer is yes, and BLS's published projection evaluations do not appear to break out accuracy by decline decile.
- What is the realized, SOC-revision-break-adjusted distribution of annualized occupational decline over 2015-2025 from OEWS? The test needs measured rather than forecast rates, and the series breaks mean nobody can read it off the published tables directly.
- Is the labor-force-exit rate itself demand-elastic, so that decline induces earlier retirement and widens the buffer endogenously? Edin et al. observe "a slightly younger age of retirement for those in declining occupations" without quantifying it as a rate, and if exits rise with decline the cut point moves upward.
- Does the band hold for occupations where the successor role sits with the same employer, i.e. does the administrative-vs-market-mediated channel from the parent brief shift the threshold rather than just the severity?
Related
- [[2026-07-28-technological-displacement-career-vs-generation]] — the parent brief whose missing benchmark this calibrates.
- [[2026-07-27-psf-pyramid-economics-post-agent-analyst-tier]] — the same question applied to the analyst tier; the scarce-layer-moved finding is the entry-cohort signature.
- [[2026-05-18-every-start-career-when-ai-doing-entry-level-job]] — the Stanford 22-25-year-old employment drop, i.e. the no-threshold entry-cohort channel in live data.
- [[2026-09-01-cpa-bookkeeper-aging-cohort-tooling-spend]] — the high-attrition case, where an old incumbent cohort widens the buffer.
- [[2026-07-26-cobol-mainframe-ai-migration-phdata-pipeline]] — a live displacement with an unusually old incumbent cohort, the condition under which this brief's test predicts generational absorption by default.
- [[project_phdata_cert_escalator_path]] — the bet this recalibrates.
Sources
- BLS Employment Projections, Table 1.10, "Occupational separations and openings, projected 2025-35": https://www.bls.gov/emp/tables/occupational-separations-and-openings.htm (page last modified 2026-08-27). Primary source for all exit, transfer, separations and employment-change figures. Retrieved via Playwright after plain
curland WebFetch returned HTTP 403. - BLS Employment Projections, occupational separations methodology: https://www.bls.gov/emp/documentation/separations.htm — source for the exits/transfers definitions and for the regression-on-worker-characteristics method that supports treating exit rates as exogenous to the employment path.
- BLS Employment Projections, occupational openings and separations fact sheet: https://www.bls.gov/emp/factsheets/occupational-openings-and-separations.htm — definitions only; carries no rates.
- Per-Anders Edin, Tiernan Evans, Georg Graetz, Sofia Hernnäs & Guy Michaels, "Individual Consequences of Occupational Decline," working-paper version 2022-11-09, published The Economic Journal 2023: https://personal.lse.ac.uk/michaels/Edin_Evans_Graetz_Hernnas_Michaels_Occupational_decline.pdf — source for the 25%-over-1984-2016 classification cutoff, the 2-5% mean and 8-11% bottom-tercile earnings losses, the 0.08-0.13 / 0.04-0.05 / 0.71-0.95 elasticities and their Table 1 definition, the "likely upper bounds" caveat, the 4-6% and 7-14% mass-layoff comparisons in footnote 6, and the retirements/reduced-entry mechanism quote.
- Fresh-eyes verification: the BLS figures in this brief were independently re-extracted and recomputed from Table 1.10 by a zero-context critic subagent on 2026-10-09, which reproduced every BLS-derived number and the zero-crossing count. The Edin et al. corrections in the "no internal validation" bullet, the 1984-2016 dating, the bottom-tercile wording and the mass-layoff-units correction all came out of that review.
~/rdco-vault/06-reference/research/2026-07-28-technological-displacement-career-vs-generation.md— the parent brief; source for the realized historical decline rates and the harm ordering.~/rdco-vault/06-reference/research/2026-07-27-psf-pyramid-economics-post-agent-analyst-tier.md~/rdco-vault/06-reference/2026-05-18-every-start-career-when-ai-doing-entry-level-job.md~/rdco-vault/06-reference/research/2026-09-01-cpa-bookkeeper-aging-cohort-tooling-spend.md~/rdco-vault/06-reference/research/2026-07-26-cobol-mainframe-ai-migration-phdata-pipeline.md- Derivation scripts for every figure labelled "derived by me":
~/.claude/scratch/occexit/analyze.pyand~/.claude/scratch/occexit/bands.py.