Quick Takeaways
What you'll learn in this article
- 1
Using a headcount-by-functional-family definition — peak employment in a family minus current employment, regardless of subsequent rehires under new titles — the United States has shed approximately 5 million white-collar jobs (predominantly clerical and administrative support), 700,000 production-occupation jobs (net of the post-COVID manufacturing recovery), and an additional estimated 3 to 4 million U.S. jobs filled offshore through corporate captive centers and BPO between 2015 and 2025.
- 2
Office and Administrative Support (BLS major group 43-0000) — the largest white-collar family in the country — lost roughly 4 million net jobs over the decade, an 18 percent contraction. The family did not recover its pre-COVID level. This is the single largest occupational hollowing in U.S. history and predates the generative AI wave by half a decade.
- 3
Manufacturing employment at the industry level (MANEMP) recovered to its December 2019 peak by late 2024, but production occupations (51-0000) sit about 700,000 below their 2014–2015 window peak — manufacturing firms are employing fewer line workers per dollar of output and more engineers and technicians, a structural shift the headline industry number conceals.
- 4
The discrete AI-attributed displacement wave (2022–2025) is small in measured headcount — roughly 70,000 to 130,000 jobs explicitly attributed to AI by firms making the cuts — but the AI curve is barely 36 months old and is already dominating hiring composition within the Computer Occupations family. The mass-displacement scenario remains modeled (Goldman 300 million globally, McKinsey 12 million U.S. shifts by 2030), not yet observed.
- 5
Against those losses, the U.S. generated roughly 17 million net new payroll jobs over the same window. But 73 percent of the gainer-family growth is in jobs paying under $45,000 per year — home health aides, food service workers, warehouse hands, personal care — with materially weaker employer health and retirement benefits than the families that shrank. Aggregate headcount is up. Compositional value is sharply down.
Keep reading for detailed implementation, code examples, and real-world results
Key Takeaways
- Using a headcount-by-functional-family definition — peak employment in a family minus current employment, regardless of subsequent rehires under new titles — the United States has shed approximately 5 million white-collar jobs (predominantly clerical and administrative support), 700,000 production-occupation jobs (net of the post-COVID manufacturing recovery), and an additional estimated 3 to 4 million U.S. jobs filled offshore through corporate captive centers and BPO between 2015 and 2025.
- Office and Administrative Support (BLS major group 43-0000) — the largest white-collar family in the country — lost roughly 4 million net jobs over the decade, an 18 percent contraction. The family did not recover its pre-COVID level. This is the single largest occupational hollowing in U.S. history and predates the generative AI wave by half a decade.
- Manufacturing employment at the industry level (MANEMP) recovered to its December 2019 peak by late 2024, but production occupations (51-0000) sit about 700,000 below their 2014–2015 window peak — manufacturing firms are employing fewer line workers per dollar of output and more engineers and technicians, a structural shift the headline industry number conceals.
- The discrete AI-attributed displacement wave (2022–2025) is small in measured headcount — roughly 70,000 to 130,000 jobs explicitly attributed to AI by firms making the cuts — but the AI curve is barely 36 months old and is already dominating hiring composition within the Computer Occupations family. The mass-displacement scenario remains modeled (Goldman 300 million globally, McKinsey 12 million U.S. shifts by 2030), not yet observed.
- Against those losses, the U.S. generated roughly 17 million net new payroll jobs over the same window. But 73 percent of the gainer-family growth is in jobs paying under $45,000 per year — home health aides, food service workers, warehouse hands, personal care — with materially weaker employer health and retirement benefits than the families that shrank. Aggregate headcount is up. Compositional value is sharply down.
The Definition That Changes the Number
A permanently lost job is peak headcount in a functional family minus current headcount, regardless of whether the firm later hired anyone. If a company fires three thousand software engineers and hires fifty AI engineers to do the same work, the loss is two thousand nine hundred fifty — not three thousand. AI engineer hires count as headcount inside the software family, not as new jobs. A laid-off factory worker who becomes a Walmart greeter does not erase the factory loss. The factory function is still hollowed.
The Argument
There are two stories you can tell about American employment between 2015 and 2025, and both are true. The first story is a triumph: total nonfarm payroll employment rose from approximately 141.6 million in January 2015 to roughly 158.5 million in early 2026 — a gain of seventeen million jobs across a window that contained the COVID economic shock, two presidential transitions, the release of GPT-4 and its successors, and the most aggressive interest-rate tightening cycle in forty years. The U.S. labor market absorbed those shocks, returned to record-low unemployment, and remained the strongest demand-side labor market in the developed world. That is the story that shows up in headline employment statistics, in stump speeches, and in the optimistic strand of think-tank commentary about the resilience of the American economy.
The second story is what those headline numbers are hiding. Beneath the aggregate seventeen-million-job gain, specific kinds of work have disappeared at a scale and pace that the gross employment figure cannot capture, because the gross figure treats a laid-off underwriter taking a delivery-driver job as a wash. By the definition used in this analysis — peak headcount in a functional family minus current headcount, regardless of subsequent rehires under new titles, with cross-family migration explicitly disqualifying as a refill — the United States has permanently lost on the order of five to six million jobs in clerical, administrative, and production families, and has stood up another three to four million equivalent positions offshore through corporate captive centers and BPO arrangements between 2015 and 2025.
Those permanent losses have been more than offset by gross job creation elsewhere in the economy, which is why the headline employment numbers look healthy. But the gross creation is overwhelmingly concentrated in low-wage service roles — home health aides, food preparation workers, warehouse hands, personal care providers — where median compensation sits between thirty-four and forty-two thousand dollars annually and employer-sponsored benefits are materially weaker than in the lost families. The U.S. economy is not running out of jobs. It is running out of the kinds of jobs that allowed a high school education to sustain a middle-class life, and the families absorbing that displacement are the ones the U.S. social model was designed around.
The purpose of this analysis is to produce a defensible map of which functions disappeared, which functions absorbed the displaced workers, what the wage gradient between those buckets looks like, and what the compositional shift implies for any social policy framework that conditions provision on employment in functions the economy is no longer producing. The numbers do not on their own argue for any particular policy response — they make the policy question unavoidable.
The Methodology — Headcount by Functional Family
The standard public dataset for occupational employment in the United States is the Bureau of Labor Statistics Occupational Employment and Wage Statistics (OEWS) series, released annually in May with prior-year data. OEWS reports total U.S. employment by Standard Occupational Classification (SOC) code, aggregated across all firms. It does not report firm-level headcount, but because the user's definition aggregates to occupation-level net change when summed across firms, OEWS is the natural implementation of the definition for public analysis. A laid-off factory worker in Toledo offsets a hired factory worker in Bentonville at the occupation level; both events net to zero in OEWS. This is the conservative implementation of the headcount-by-family rule — more permissive than a strict firm-level definition would be, but it is what public data can support.
Five interpretive rules govern this analysis:
First, the family is broader than the SOC code. If a firm fires three thousand software engineers and hires fifty "AI engineers" or "ML engineers" to do the same work, the AI hires count as headcount inside the Computer Occupations family (BLS 15-1200), not as new jobs in a separate AI category. BLS classifies machine learning and AI specialists inside 15-1252 and 15-1257 — inside the existing Computer Occupations family — which means the user's "AI engineers count as software refills" rule maps cleanly onto BLS data without further adjustment.
Second, the COVID rule. Between February and April 2020, the U.S. lost approximately 22 million payroll jobs. Most were temporary; the Leisure and Hospitality industry, for example, recovered its pre-COVID employment by 2022. Under the user's rule, COVID-era losses only count as permanent if the family did not recover its peak headcount by 2025. Manufacturing industry employment (MANEMP) recovered to within 0.1 million jobs of its December 2019 peak by late 2024 — almost all of the COVID-era manufacturing loss is excluded from the permanent-loss tally on that basis. Office and Administrative Support, on the other hand, did not recover; the family sits roughly four million jobs below its 2015 window peak, and all of that gap counts.
Third, cross-family migration does not count as a refill. A laid-off clerical worker who finds a job as a home health aide is still a permanent loss in the Office and Administrative Support family. The Healthcare Support gain is recorded in a separate column. This is the rule that prevents the analysis from being mathematically vacuous; without it, the very high U.S. gross hiring rate would mechanically erase any structural decline.
Fourth, offshored functions are a separate bucket called "Filled Offshore." When a U.S. firm closes a five-hundred-seat back-office operation in Tampa and stands up an identically-staffed function in Manila or Hyderabad, the U.S. five hundred jobs are counted as permanently lost; the five hundred Philippine or Indian seats are counted as Filled Offshore. The function still exists globally — but it does not exist in the United States. This distinction is important because the policy responses differ: automation losses require workforce transition policy, while offshoring losses also implicate trade, tax, and labor-cost competitive policy.
Fifth, the window is strict — January 2015 through the most recent available data, which is the OEWS May 2024 release with cross-checks against the preliminary OEWS May 2025 release and Current Employment Statistics (CES) industry-level data through April 2026.
The measurement caveats are real and worth flagging up front. OEWS data is occupation-level aggregated; it does not resolve same-family same-occupation churn between firms. BLS does not ask firms why headcount declined, so automation versus AI versus offshoring versus consolidation must be inferred from external industry research. Gig and 1099 work — Uber, DoorDash, Instacart — is partially undercounted in payroll surveys. And the OEWS May 2025 release (published by BLS in 2026) is the most current snapshot available; 2025 data for the most recent quarters is partial. Where this matters, the analysis labels estimates explicitly.
The White-Collar Hollowing — Four Million Office Jobs
The largest single permanent-loss bucket in the U.S. economy over the window is also the cleanest case for the user's definition. Office and Administrative Support occupations — SOC major group 43-0000, which includes secretaries and administrative assistants, general office clerks, bookkeepers, tellers, data entry workers, switchboard operators, file clerks, and mail clerks — peaked around twenty-two million U.S. workers in 2014–2015. By May 2024, the family sat at approximately eighteen million. The May 2020 COVID trough was around nineteen million — meaning the family is now operating below its COVID trough four years on.
Office and Administrative Support (SOC 43-0000) — Total U.S. Employment by Year (millions of workers)
| year | employment_millions |
|---|---|
| May 2013 | 21.5 |
| May 2015 | 22 |
| May 2017 | 21.3 |
| May 2019 | 20.3 |
| May 2020 (COVID) | 19 |
| May 2022 | 18.9 |
| May 2023 | 18.5 |
| May 2024 | 18 |
The four-million headcount loss in 43-0000 over the decade is concentrated in a small number of subroles. Secretaries and administrative assistants fell from about 4.1 million in 2014 to 3.5 million in 2024 — a 600,000 net decline. General office clerks dropped from 2.8 million to roughly 2.2 million — another 600,000. Data entry keyers, switchboard operators, file clerks, and mail clerks each contracted by thirty to sixty percent over the decade; switchboard operators are now effectively an extinct occupation in the U.S. labor market, with the function fully absorbed by enterprise telephony software and remote-receptionist services.
The causes layer on top of each other in descending weight. The first and largest is front-office digitization that began well before the AI wave: booking, billing, expense management, payroll, and procurement all migrated from clerical-staffed processes to self-service SaaS over the decade. The second is customer-service offshoring to Philippine and Indian BPO operators, which by industry estimate accounts for roughly 1.4 million Philippine seats serving U.S. clients alone. The third — and most recent — is generative AI, which Goldman Sachs' March 2023 analysis modeled as having the highest task- automation exposure of any major U.S. occupational family at forty-six percent. The Goldman number is a projection of exposure, not an observation of displacement, but the directional decline in 43-0000 hiring since the ChatGPT commercial launch in late 2022 is consistent with an early-stage acceleration of a trend that was already running.
Beyond 43-0000, white-collar permanent losses extend into related families with different cause profiles. Inside sales and telemarketing roles (BLS 41-3000 series and 41-9041) dropped from roughly 2.4 million to 1.9 million — a 500,000 family loss driven by CRM automation, marketing-automation tooling, outbound-call regulation, and offshoring. Telemarketers specifically fell from about 218,000 to under 100,000 — a 55 percent contraction in a single decade.
Financial back-office subroles — bank tellers, bookkeeping and accounting clerks, loan interviewers, bill and account collectors — collectively lost about 470,000 net jobs over the window. Tellers alone declined from 502,000 to about 315,000, a clean automation story (ATMs, mobile deposit, RPA, rules- based credit decisioning) where the same banks employing fewer tellers are hiring in compliance, risk management, and technology functions — different families, different wages, different educational requirements.
Legal support work is in transition rather than decline at the family level. Paralegals and legal assistants grew by about ninety thousand over the decade, while legal secretaries lost about sixty thousand and court clerks / title examiners lost another fifteen thousand. The net family change is roughly flat but the role-mix shift is significant: document-automation software is absorbing legal-secretary work and being deployed by paralegals with broader scopes of responsibility. Goldman Sachs flagged legal as the second-most AI-exposed family (44 percent task-automatable). The next 24 to 48 months will be the test of whether that exposure converts to family-level contraction.
Newspaper publishing employment at the industry level fell from 178,000 in 2015 to 74,000 in 2024 — a 100,000-job permanent loss in a single industry, not COVID-related, continuing a secular trend Pew Research has tracked since 2008. Journalism occupations narrowly defined (news analysts, reporters, editors) declined more modestly because the surviving roles consolidated into fewer but larger operations.
The Computer Occupations family (15-1200) — which the user's definition treats as a single functional bucket absorbing AI engineer, ML engineer, and prompt engineer hires as same-family backfill — is the case that most sharply illustrates how the definition shapes the number. Family-level employment went from approximately 3.95 million in 2015 to roughly 5.0 million in 2024. Net headcount change in the family: positive one million jobs. That is the honest read under the user's rule, because the AI/ML hires at firms cutting "traditional" software developer roles are counted as refills. But the Layoffs.fyi cumulative tally for the tech sector — 632,000 cumulative layoffs from 2022 through 2025 — captures a level of gross within-family churn that the family-aggregate number completely obscures. Most laid-off workers were rehired within the family, often at different firms, often at lower comp, often with explicit requirements to demonstrate AI-tool fluency. The family is not hollowing in headcount yet. It is hollowing in hiring composition, and that is a real but different phenomenon worth treating as its own section below.
The Production Floor — Manufacturing Recovered, the Jobs Didn't
The mismatch between manufacturing as an industry and production as an occupational family is the analytically cleanest example of why the user's definition matters. The Federal Reserve's MANEMP series — total U.S. manufacturing industry employment — bottomed at 11.39 million in April 2020 during the COVID shock and has recovered to roughly 12.76 million as of December 2024, within one hundred thousand jobs of its December 2019 peak of 12.84 million. The headline industry number is essentially fully recovered. The CHIPS Act and Inflation Reduction Act semiconductor and EV-battery construction, combined with broader reshoring momentum, drove the recovery.
U.S. Manufacturing Industry Employment (CES MANEMP) — Millions of Workers, January 2015 Through April 2026
| date | manempm |
|---|---|
| Jan 2015 | 12.3 |
| Jan 2017 | 12.4 |
| Dec 2019 | 12.84 |
| Apr 2020 | 11.39 |
| Dec 2021 | 12.59 |
| Dec 2023 | 12.97 |
| Dec 2024 | 12.76 |
| Apr 2026 | 12.7 |
But manufacturing-industry employment includes engineers, technicians, supply- chain managers, sales staff, IT staff, finance staff, and administrators working inside manufacturing firms. The narrower question — production occupations (SOC 51-0000), the people running the line — tells a different story. Production occupations sat at approximately 9.3 million in 2015 and 9.2 million in 2019; they bottomed at roughly 8.0 million during COVID and have since partially recovered to 8.6 million as of May 2024. Net permanent loss in production occupations specifically: approximately seven hundred thousand jobs.
The composition of the recovery matters. Welders, industrial-machinery mechanics, and process-control technicians have grown — these are the roles required to build and maintain new fab and battery facilities. Team assemblers (51-2092), the most archetypal "assembly line" occupation, declined from 1.10 million to 0.90 million over the decade. Sewing machine operators fell from 145,000 to 95,000. Inspectors, testers, and sorters lost 60,000 net positions. Print binding, prepress, and finishing workers each declined by thirty to fifty percent — these are functions where computer- vision quality control and direct-to-substrate digital printing have absorbed the work.
Acemoglu and Restrepo's research on industrial robots (NBER Working Paper 23285 and the Journal of Political Economy 2020 follow-on) estimated that robot deployment displaced between 360,000 and 750,000 U.S. jobs cumulatively through 2017, concentrated in manufacturing local labor markets. Their methodology was an empirical attribution exercise — comparing labor outcomes in geographic areas with more versus less robot exposure — and the upper end of that range maps closely to the 700,000 production-occupation permanent loss measured directly in OEWS. The convergence between two independent estimation approaches gives confidence in the magnitude. Robots took the assembly seats. The reshoring boom is bringing back the technician and engineer seats. The line worker on a 1985-era job description is the seat that is structurally gone.
This is the pattern that gets missed when commentators say "manufacturing has recovered." Manufacturing as a sector of the economy has recovered. The specific kind of work that allowed a high school graduate to support a family on a single income in 1985 has not, and will not, recover. The factories are back; the assembly-line jobs are not, and the difference is the technician education premium that the recovery has installed in their place.
Filled Offshore — Three to Four Million Equivalents
The hardest bucket to quantify is also the one that most cleanly answers the question "where did the jobs go." No single public dataset directly counts "U.S. function eliminated and identical function stood up offshore," so the estimate triangulates across three sources.
The Bureau of Economic Analysis tracks employment by majority-owned foreign affiliates of U.S. parent companies through its Activities of U.S. Multinational Enterprises program. Worldwide foreign-affiliate employment by U.S. MNEs grew from approximately 13.8 million in 2015 to roughly 16.0 million in 2023 (the most recent year with full data). That 2.2-million-job expansion is not all offshored U.S. work — much is in-country sales and distribution serving local foreign markets — but the share involving back- office, IT, customer-service, and engineering captives serving the U.S. parent operations is large and growing.
The Indian IT-BPM industry, tracked by industry association NASSCOM, grew from approximately 3.7 million workers in fiscal 2015 to about 5.6 million in fiscal 2024. Industry analysts attribute roughly 60 to 70 percent of that work to U.S. clients. The U.S.-attributable seat growth over the decade is approximately 1.2 to 1.4 million seats. The Philippine BPO sector grew from 1.05 million to 1.4 million workers over the same period, with about 70 percent serving U.S. clients — an additional roughly 250,000 U.S.-facing seats. Mexico nearshore (Guadalajara, Monterrey, Mexico City) and other secondary destinations contribute another 150,000 to 250,000 U.S.-attributable seats.
The Economic Policy Institute's analysis of U.S.-China trade deficit impacts quantified 3.7 million U.S. jobs lost to that single trade relationship between 2001 and 2018. For the 2015 to 2024 sub-window specifically — which captures the marginal additional loss after the bulk of pre-2015 trade- driven manufacturing offshoring had already occurred — EPI estimates 600,000 to 1 million additional U.S. manufacturing jobs displaced by ongoing trade deficits with China and Mexico. This number is partially counter-flowed by reshoring: the Reshoring Initiative reports approximately 1.3 million cumulative announced reshoring and FDI manufacturing jobs since 2010, with 244,000 announced in 2024 alone. Announced is not filled — typical build-out lag is two to five years — but the counter-flow is real and bears on the net number.
Filled Offshore — Composite Estimate of U.S. Jobs Filled by Offshore Functions, 2015-2025 (millions, midpoint of estimate range)
| Name | Value |
|---|---|
| IT & Software offshore (India + Mexico + Philippines + captive centers) | 1.8 |
| Customer Service / BPO U.S.-facing | 1.2 |
| Manufacturing offshore net of reshoring | 0.55 |
The composite estimate — about 3 to 4 million U.S. equivalent jobs filled offshore over the decade — is roughly the same magnitude as the entire white-collar permanent-loss bucket inside the U.S. The two phenomena are not independent. The same firms automating customer service domestically are also operating Manila and Bangalore captives. The same firms cutting U.S. back-office headcount are standing up global capability centers. Filled Offshore is best understood as a parallel hollowing — same firms, same functions, different geography.
The AI Wave Is Just Starting
The defining problem with the public conversation about AI and jobs is the gap between what is observed and what is projected, and the conversation collapses the gap. Observed AI-attributed displacement through end of 2025 is small. Challenger, Gray & Christmas began tracking "artificial intelligence" as a cited reason for U.S. layoffs in May 2023; cumulative AI-cited announcements through end of 2025 total approximately 71,825 jobs, of which 54,836 were announced in 2025 alone. The Layoffs.fyi tracker shows roughly 632,000 cumulative tech-sector layoffs from 2022 through 2025, very few of which were AI-cited in firm statements — most cited "restructuring," "macroeconomic environment," or "right-sizing" — but the temporal pattern since the ChatGPT commercial launch in November 2022 is consistent with generative-AI-driven productivity reorganization.
AI Displacement — Observed Versus Projected, Logarithmic Scale of U.S. Jobs (2022-2030 horizon)
| timeline | jobs |
|---|---|
| AI-cited cuts 2023 | 2000 |
| AI-cited cuts 2024 | 14989 |
| AI-cited cuts 2025 | 54836 |
| Tech-sector layoffs 2022-2025 (Layoffs.fyi) | 632000 |
| McKinsey U.S. shifts by 2030 (projection) | 12000000 |
| Goldman global FTE exposure (projection) | 300000000 |
The projection side of the gap is loud. Goldman Sachs Global Investment Research in March 2023 estimated that generative AI exposed roughly 300 million full-time-equivalent jobs globally to some degree of automation, with U.S. office and admin tasks at 46 percent automation potential and legal at 44 percent. McKinsey Global Institute's July 2023 analysis projected about 12 million U.S. occupational shifts by 2030, with 80 percent concentrated in customer service, food service, production, and office support — and specifically projected demand declines of 830,000 for retail salespersons, 710,000 for administrative assistants, and 630,000 for cashiers by 2030. The OECD Employment Outlook 2023 estimated 27 percent of OECD jobs at high risk of automation. Brookings researchers have repeatedly noted that generative AI's exposure pattern inverts the prior automation wave: better- paid, better-educated, urban knowledge workers are more exposed than manufacturing or service workers, which is the opposite of what robotics did in the 1990–2017 window.
The honest read of the curve so far is that observed displacement is small relative to projected exposure, and the mechanism currently visible in BLS data is not mass layoffs in the families most exposed. It is hiring composition shift — the Computer Occupations family ticking sideways while the mix of roles inside it shifts toward AI/ML/data engineering and away from "general software developer" — and selective non-replacement, where the Office and Administrative Support family declines at one to one and a half percent per year through normal attrition rather than mass layoffs. The McKinsey 12-million-shift scenario over a six-year horizon implies an average pace of two million per year. We are currently observing one to two hundred thousand per year of attributable AI displacement, with another roughly six hundred thousand per year of likely-AI-adjacent within-Computer- Occupations churn. The wave is real. It is small so far. It is mathematically certain to grow on the projected trajectory, and the user's headcount definition will pick up the inflection cleanly when the Computer Occupations family begins to contract — which is the prediction worth making.
I am going to make that specific prediction now and commit to it as a falsifiable check on the analysis. By the May 2027 OEWS release, the Computer Occupations family will be at or below 4.85 million, representing the first net headcount contraction in a family that has grown steadily for three decades. If the number lands at or below 4.85 million, the AI displacement wave will have crossed from "observed only in hiring composition" to "observed in family-aggregate headcount." If it does not, the wave is moving slower than this analysis estimates.
What Replaced Them — Eight Million Sub-$45K Service Jobs
The U.S. economy is not running out of jobs. It is creating them at a healthy clip. The question is which kinds.
The five largest gainer families over the window are Healthcare Support (SOC 31-0000), Transportation and Material Moving (53-0000), Healthcare Practitioners (29-0000), Computer Occupations (15-1200), and Food Preparation and Serving (35-0000). Together those families added approximately 8.4 million net jobs between 2015 and 2024.
Where the New U.S. Jobs Came From, 2015–2024 — Net Headcount Added by Functional Family (millions)
| family | jobs_added_millions |
|---|---|
| Healthcare Practitioners (29-0000) | 1.3 |
| Computer Occupations (15-1200) | 1.05 |
| Healthcare Support (31-0000) | 2.7 |
| Transportation & Material Moving (53-0000) | 2.5 |
| Food Prep & Serving (35-0000) | 0.4 |
| Personal Care (39-0000) | 0.5 |
The single largest gaining occupation in the entire U.S. economy is Home Health and Personal Care Aides, which grew from approximately 1.8 million workers in 2014 to about 4.3 million in 2024. It is the largest occupation in America, period. The Bureau of Labor Statistics' 2024–2034 projection expects it to add another 765,800 annual openings through retirement and turnover, growing 17 percent further over the next decade. The median annual wage for the occupation is under $34,000.
Hand laborers and material movers — the warehouse hands inside Amazon, FedEx, UPS, and retail distribution centers — grew from about 5.2 million in 2015 to roughly 7.0 million in 2024, a 1.8 million-job expansion. Light-truck and delivery drivers (the Amazon Flex, DoorDash, and Instacart layer of the gig economy plus FedEx Ground and UPS) added approximately 400,000 jobs. Heavy and tractor-trailer truck drivers — the freight backbone — added 250,000 net positions over the decade.
The compositional finding is brutal in its simplicity. The high-wage gainer families — Healthcare Practitioners and Computer Occupations — added a combined 2.35 million jobs over the decade. The low-wage gainer families — Healthcare Support, Food Preparation, Personal Care, and most material- moving subroles — added approximately 6.1 million jobs over the same window. For every job created paying above $80,000 per year, the U.S. created roughly three jobs paying under $42,000 per year, and the wage gap between the two tiers is widening, not narrowing.
For comparison, the families that lost jobs over the same window were mostly in the $46,000 to $112,000 median-wage band. Office and Administrative Support median wage in May 2024 was $46,320. Production median wage was $45,960. The lost-family median wage is approximately $46,000. The low-wage gainer-family median wage is approximately $36,000. The decade-long compositional shift is a $10,000 median wage gap, widening to roughly $14,000 when weighted by composition. Compounded over a twenty-year working life, the difference is six figures of lifetime earnings before benefit differentials are counted — and the benefit gap is larger than the wage gap.
Employer-sponsored health insurance coverage in the lost families ranged roughly 65 to 85 percent of workers. In the low-wage gainer families, coverage ranged 30 to 55 percent, with home health aides under 40 percent and food service workers under 30 percent. 401(k) match availability in the lost families ran 50 to 70 percent of workers; in the gainer families it ran under 20 percent for the lowest-wage tiers. The wage gap is the surface phenomenon. The benefit gap is the larger structural fact, and it is the one that determines whether displaced workers can absorb medical, retirement, and unemployment risk on their own.
The Compositional Shift Is the Story
Putting the numbers together produces a picture that the standard "unemployment rate is low" framing cannot capture. Aggregate U.S. employment is healthy. Aggregate headcount is up roughly 12 percent over the decade. Prime-age (25 to 54) male labor force participation has actually paused its multi-decade secular decline and recovered slightly to about 89.2 percent in early 2026, well above the 88.3 percent of 2015 — the labor force is more engaged, not less. By every aggregate measure, the U.S. labor market is working.
But the composition of that working market has shifted approximately 8 million jobs from middle-wage clerical and production work to low-wage service work, while losing another 3 to 4 million U.S. functions to offshore operators. The replacement income for the median lost worker is roughly 22 percent lower than the lost income. The replacement benefits are substantially weaker. The replacement work is overwhelmingly part-time or gig-structured, undercounted in BLS payroll surveys, and concentrated in sectors with the lowest unionization rates in the developed world.
Median Annual Wage — Lost Families (43-0000 + 51-0000 average) Versus Gainer Low-Wage Families (31-0000 + 35-0000 + 39-0000 average), 2015–2024
| year | lost_family_median_wage | gainer_low_wage_family_median |
|---|---|---|
| 2015 | 42000 | 31000 |
| 2017 | 43400 | 31900 |
| 2019 | 44200 | 33100 |
| 2021 | 45100 | 34000 |
| 2023 | 45800 | 35400 |
| 2024 | 46140 | 36000 |
The displaced worker outcomes data published by BLS in its biennial Displaced Worker Supplement is the human dimension of these numbers. The most recent release (January 2024, covering 2021–2023 displacements) found that 65.7 percent of long-tenured displaced workers were reemployed at the time of the survey, of whom 62 percent earned wages equal to or greater than their pre-displacement wage. About 35 percent of long-tenured displaced workers were not reemployed six or more months after displacement. Of those who were reemployed, roughly 38 percent took a pay cut. The user's definition tracks this human dimension directly: a factory worker reemployed as a warehouse hand is "reemployed" in the BLS survey but is a permanent loss in the user's functional family definition. Both measurements are correct; they answer different questions.
For the Computer Occupations family specifically, displaced-worker outcomes in the 2022–2023 layoff cohort were unusually strong by historical standards — high reemployment rates, often at equal or higher wages within six months — which is consistent with the family-aggregate barely moving in OEWS. Software engineers who lost their jobs at Meta or Google in late 2022 mostly landed at other firms inside the family. That is exactly what the user's rule predicts: within-family reemployment is registered as a refill; the gross churn does not show up in the permanent-loss number.
What This Means — The Mathematics of the Post-Work Question
The mathematics of the U.S. labor market over the next decade are not a question of whether the economy will have jobs. It will. They are a question of whether the jobs it has will continue to be the kind that support the social architecture the United States built between 1945 and 1985 — single-earner middle-class life, employer-sponsored health insurance, defined-benefit or matched-contribution retirement, employer-tied access to disability and unemployment risk-sharing, employer-paid contribution to local-government tax base, and the broader sense that ordinary work entitles a person to ordinary security.
The data above suggests that architecture is being dismantled occupation by occupation. Not by any single policy decision, but by the cumulative effect of decisions made by individual firms responding to individual incentives — to automate clerical work, to offshore back-office functions, to deploy robots in assembly, to consolidate procurement through SaaS, to commission AI for first-draft content production, to route customer service through foundation models. Each individual decision is rational and most are productivity-improving on the firm's books. The aggregate effect is the gradual disappearance of the kind of work the U.S. social model assumes exists.
The choice the United States faces is not really between "work-based provision" and "universal-basic provision." It is between maintaining a work-based provision framework while the work that historically anchored it is being structurally removed, or building a different framework that provides for what people need — health care, retirement, disability coverage, basic income security — through mechanisms that do not require ordinary employment in disappearing functions to be the eligibility test. The first path requires denying or minimizing the trajectory the numbers above describe. The second path requires accepting it and engineering alternatives — universal basic income, universal health care, expanded disability insurance, public-option retirement, geographic and occupational mobility support, or some combination.
This piece does not advocate a specific policy. It produces the measurement that makes the question unavoidable. Approximately five million white-collar jobs gone. Approximately seven hundred thousand production jobs gone. Approximately three to four million U.S. functions filled offshore. AI displacement just starting and projected at twelve million U.S. shifts by 2030. Net new jobs concentrated in low-wage service work at roughly three to one against high-wage knowledge work. The numbers do not in themselves argue for universal basic income or any other specific intervention. They argue that the framework that conditions human provision on employment in functions the economy is no longer producing is mathematically running out of room.
The question facing the United States is whether to recognize that constraint and design around it now, or to discover it later as a social crisis. The data is the same in both cases. The cost of waiting is the difference.
Further Reading
This analysis builds on prior CrashBytes work tracking specific occupation displacement timelines. The executive administrative assistants displacement piece covers the AI-specific overlay on the 43-0000 hollowing detailed above. The medical coders HAR analysis shows the same compositional pattern inside the healthcare back-office ecosystem. The insurance underwriting barbell piece walks through how a single white-collar industry is splitting into a high-skill specialty tier and an automated middle, the most readable case study of the larger compositional shift. The Cloudflare 1,100-layoff analysis documents one firm's explicit AI-driven workforce reduction. The Computer Occupations family decline prediction referenced earlier in this piece is the falsifiable check on whether the AI wave crosses into family-aggregate contraction by mid-2027. And this week's news digest covers the broader context of AI-labor news from May 18–22.
