Jobs are growing, and some searches are taking longer

The index tracks long-term unemployment among managers and professionals: people whose last job was in management, business, financial or professional occupations and who have been unemployed 27 weeks or more, as a share of that labor force. It stood at 162 in August 2026 (2019 = 100), close to its June reading of 170, the highest since 2021. The index describes one part of the market, the people searching longest; Chapter 02 covers hiring and openings.

Two cautions shape how to read the index. First, single months are noisy: the margin of error on a single month's reading is roughly ±24 points, so August 2026 (162) is not statistically different from August 2025 (128) on its own. The eight-month comparison below is. Second, the 2019 base was the strongest labor market of the past decade. The index ran between 131 and 201 in 2015–2016, so today's level is high relative to the late 2010s, not unprecedented.

Does the result depend on seasonal adjustment?

No. Whatever adjustment method is used, the index is 53–62% above its 2019 level and about a quarter higher than a year earlier. Our fixed seasonal factors give the highest current reading; the Census Bureau's standard X-13ARIMA-SEATS program gives the lowest.

Companion measure: people who lost their jobs

The index counts everyone long-term unemployed whose last job was white-collar, including people who quit or are returning to work. Limiting it to people who lost a job (laid off, let go or whose temporary job ended) gives a measure closer to the everyday meaning of displacement. It has risen faster than the overall measure: up 76% from 2019, against 59% for all white-collar long-term unemployment. Job losers make up just over half of the white-collar long-term unemployed in both years.

Beyond unemployment: how are employed professionals doing?

Unemployment counts only people without jobs. Workers could also be affected if they were moved from full-time jobs into part-time hours, second jobs or self-employment. We checked the same survey for employed managers and professionals.

None of the three measures rose. Professionals who keep their jobs are not visibly being pushed into part-time hours, second jobs or self-employment; multiple jobholding fell. Hours are below 2019 but unchanged from last year. On current data, the slower part of the white-collar market shows up in how long people who leave jobs take to find the next one, not inside existing jobs.

Employers added 29,000 jobs in September, below the 84,000 economists expected (per Axios). July was revised to a loss of 10,000 and August to +133,000, a combined downward revision of 60,000. Unemployment ticked up to 4.2% as more people joined the labor force. Over the first nine months of 2026 the economy added 612,000 jobs, versus 232,000 in the same months of 2025 on current data. Payroll gains of this size mean something different than they used to: the labor force is growing slowly (net immigration is lower), and Federal Reserve estimates of the "breakeven" job growth needed to keep unemployment steady have fallen from about 155,000 a month in 2023–24 to 85,000 in 2025, and for 2026 range from under 10,000 a month (a Federal Reserve Board low-immigration scenario) to 87,000 (St. Louis Fed, using CBO assumptions). September's gain is weak, but not far below breakeven.

The main constraint for job seekers is not a wave of firings: layoffs and discharges are near 1% of jobs. It is a steady but slow-moving market. Few people quit, there is about one opening per unemployed person, and people who switch jobs earn 1.4 points more in raises than those who stay, so job seekers who are out of work compete with employed candidates for a limited flow of openings.

Each job listing now yields fewer hires

In the early 2000s employers made more hires each year than they had job openings, about 1.2 to 1.5 hires per opening. By 2019 that had fallen to 0.82, and in January–August 2026 it was 0.73: about seven hires for every ten openings. In professional and business services, the largest white-collar sector, it fell from 0.90 in 2019 to 0.79. Combined with the rise in applications per opening (Chapter 08), each listing now produces fewer hires and takes far more applicants per hire.

Information and financial activities are where white-collar payrolls fell over the year. Professional and business services, the largest white-collar sector, grew by 123,000, though computer systems design lost 32,000. Federal employment is down 327,000 since December 2024.

Employers announced 43,281 job cuts in September, the fewest for a September since 2022, according to Challenger, Gray & Christmas. Year to date, announced cuts total 573,195, down 39% from 2025. Most of that decline reflects the 2025 federal government cut announcements not recurring: excluding government, cuts are down 15%. Hiring plans were the lowest for any September since 2011.

Technology accounts for 165,925 announced cuts this year, 29% of the total and up 54% from 2025. Notable September announcements include Uber (about 3,300, citing fewer management layers), Workday (about 500, restructuring) and Microsoft (about 500, Xbox).

What companies say about AI versus what they file

Employers cited AI as the reason for 120,136 announced cuts this year, about 21% of the total and the most-cited reason, per Challenger. Formal filings cite it less often. We searched every 2026 SEC Form 8-K that references Item 2.05, the item public companies use to report material restructuring costs, and read every AI passage in full.

AI is more prominent in how layoffs are announced (21% of announced cuts) than in formal restructuring filings (about 9% of filings), but both measures show AI as a real part of how companies describe restructuring in 2026. The comparison has three limits. First, the units differ, workers versus filings, so the two percentages are not directly comparable. Second, Form 8-K Item 2.05 requires companies to disclose a restructuring decision and its expected costs, not its reasons, so a filing that is silent about AI does not show AI played no role. Third, Challenger codes the reason employers give publicly, which may reflect how companies want a cut to be understood. Neither source is wrong; they measure different things, and the gap between them is worth tracking.

Long searches rose across most occupations, which points to a broad slowdown rather than one sector. Against that backdrop, computer and mathematical occupations stand out: 41% of their unemployed have been out 27 weeks or more, up from 28%, the largest rise of any occupation; increases for business and financial operations and for education also exceed their margins of error.

Immigration: does it explain long searches?

Immigration fell sharply in 2025–2026, including new H-1B and student visas. Fewer new foreign professionals could ease competition for U.S. job seekers, or push work offshore. We used the survey's citizenship question to test whether immigration explains this year's rise in white-collar long searches.

The rise is concentrated among U.S.-born workers, where it clears the margin of error. Naturalized citizens show no change, the noncitizen change is within its margin, and the noncitizen share of the white-collar labor force did not grow from 2025 to 2026.

• H-1B and students · USCIS; IIE

  A $100,000 fee on many new H-1B petitions took effect in September 2025 (few employers paid it, and it is being challenged in court), and the first wage-weighted H-1B lottery cut registrations 38.5% in March 2026. New international students fell 17% in fall 2025. Net international migration fell from 2.7 million (year to mid-2024) to a projected 321,000 (year to July 2026, Census).

• Research on H-1B and U.S. workers

  Findings are mixed: firms that won extra H-1B visas hired fewer other workers (Doran, Gelber and Isen, 2022); a 2004 cap cut did not raise hiring of U.S.-born workers (Mayda and others, 2018); rejected petitions shifted hiring abroad (Glennon); and H-1B-driven STEM growth raised U.S.-born graduates' wages (Peri, Shih and Sparber, 2015).

In this test, the rise in long searches is concentrated among U.S.-born workers and the noncitizen share of the white-collar labor force did not change, so changes in immigration do not account for most of the rise in these data. The test does not measure indirect channels such as offshoring. The clearest link to our data is through payroll growth: with the labor force growing slowly, smaller monthly payroll gains correspond to steady unemployment (Chapter 02).

Do occupations more exposed to AI fare worse?

We tested whether occupations where AI can do more of the work show worse labor-market outcomes. For each of 22 occupation groups we used a published measure of AI capability: the share of an occupation's tasks that a large language model could perform at least 50% faster with equivalent quality, rated task by task (Eloundou, Manning, Mishkin and Rock, Science, 2024), weighted by May 2022 employment. We then compared outcomes before ChatGPT's release (January–August 2022, and 2019) with January–August 2026.

Two results hold and one does not. First, in occupations where AI can do more of the work, young people make up a shrinking share of employment, whether measured from 2019 or from 2022; in the least-exposed occupations their share rose. Second, since ChatGPT's release, long-term unemployment has risen much more in high-exposure occupations than in low-exposure ones, a gap beyond the margin of error. But measured from 2019 there is no gap: high-exposure occupations did unusually well during 2020–2022 (remote work and a tech hiring boom), so part of the post-2022 rise is a return to normal. These data cannot separate that normalization from an AI effect, and employment in exposed occupations has kept growing.

The pattern is the one the rest of this report describes: the clearest AI-linked signal is at the entry gate, in fewer young hires, not in falling employment. The long-search gap is consistent with AI but not proof of it. This is a cross-section of 22 broad groups; it is a first test, not a causal estimate, and we will repeat it each quarter with detailed occupations and a second, usage-based exposure measure.

What independent research says about AI, functions and seniority

Our tables show where long searches are concentrated; they do not show what caused them. Research that tracks AI exposure directly points the same way on two points: early-career workers and some tech occupations are most affected, and most of the effect is on hiring rather than firing.

• Seniority · Stanford Digital Economy Lab (Aug 2026)

  Employment of 22–25-year-olds in the most AI-exposed jobs is 19% below comparable peers, up from 15% a year earlier. Older workers show no comparable gap; declines are concentrated where AI automates tasks rather than assists.

• Seniority · Revelio Labs (Aug 2026)

  Compared with non-adopters, firms using AI grew senior roles by 32% but junior roles by only 6%.

• Job postings · Indeed Hiring Lab (Jul 2026)

  Senior postings rose 14.7% over the year to May 2026 while entry-level postings fell 7.5%.

• Job titles · Stanford; Anthropic

  Software developers and customer service representatives show the clearest early-career employment declines. AI usage data show the heaviest task coverage in data entry, software development, medical transcription and database architecture.

• Broad effect · Yale Budget Lab; Federal Reserve (Sep 2026)

  The occupational mix is not yet shifting in ways tied to AI. Fed Governor Barr saw "little evidence of significant displacement so far" but "some indications" of fewer entry-level jobs.

• Middle management · surveys

  41% of employees say their organization cut management layers (Korn Ferry survey, cited by Forbes, May 2026). There are no hard counts yet of manager headcount, so "flattening" remains a trend to watch.

The best independent research reaches a consistent and fairly narrow conclusion. AI is not yet causing broad job losses across the economy. Its clearest measurable effect so far is fewer young workers being hired into AI-exposed jobs. The open questions are about the future and about transitions: how quickly firms reorganize around AI, and how costly the move to the next job is for the people affected. Each item below is labeled as primary research (the authors' own data) or commentary, and estimates of exposure are kept separate from observed job changes.

Observed effects so far

• Stanford Digital Economy Lab · primary research · Aug 12, 2026

  Using ADP payroll records covering 3.5–5 million workers a month, Brynjolfsson, Chandar and Chen find employment of 22–25-year-olds in the most AI-exposed occupations is 19% below where it would be had it kept pace with less-exposed peers. The gap runs through reduced hiring of young workers rather than more separations, and workers 35 and older did not diverge. The authors describe these as descriptive patterns, not causal estimates, and say they "do not see widespread, economy-wide job displacement associated with AI." Their February 2026 follow-up found that part of the timing reflects factors other than AI, including interest rates.

• Harvard (Hosseini and Lichtinger) · primary research, working paper · Oct 2025

  Across 62 million workers at 285,000 firms, junior employment at firms that adopted generative AI fell 7.7% relative to non-adopters after six quarters, while senior employment did not fall. Again, the mechanism was slower hiring, not layoffs.

• Brookings and Yale Budget Lab · primary research · Oct 1, 2025

  Molly Kinder and co-authors found the mix of jobs across high-, medium- and low-AI-exposure occupations "remained remarkably steady" in the 33 months after ChatGPT's release, changing only marginally faster than in the early computer and internet eras. They caution that this method can miss damage concentrated in narrow occupations, and that a widening gap between younger and older graduates is consistent with, but not proof of, an AI effect.

• Oxford (Llanos-Paredes and Frey) · primary research, working paper · 2025

  In one occupation where the technology is mature, translation, each one-point increase in machine-translation use was associated with about 0.7 points slower growth in translator employment, roughly 28,000 translator jobs that would otherwise have been created in 2010–2023.

• Stanford, MIT and Harvard field studies · primary research

  Within a job, AI helps novices most: customer-support agents gained 14% in productivity on average and 34% for the least experienced (Brynjolfsson, Li and Raymond, QJE 2025). In a Procter & Gamble experiment with 776 professionals, one person working with AI matched the output quality of a two-person team without it (Dell'Acqua, Sadun, Lakhani and others, 2025). Read together with the hiring studies, the tension is clear: AI makes junior workers more productive, yet fewer of them are being hired.

Exposure, adaptability and who is at risk

• Brookings (Manning, Aguirre, Muro, Methkupally) · primary research · Jan 21, 2026

  Of 37.1 million workers in the most AI-exposed quarter of jobs, 26.5 million (about 70%) have above-median capacity to adapt, based on savings, transferable skills, local job density and age. About 6.1 million workers, 4.2% of the workforce, are both highly exposed and poorly positioned to adapt; 86% of them are women, concentrated in clerical and administrative roles. For most of NextChapter's audience of experienced professionals, the issue is therefore less whether they can transition than how long and how costly the transition is.

• Opportunity@Work (Byron Auguste) with Brookings · primary research · Apr 2, 2026

  15.6 million workers without four-year degrees ("STARs," Skilled Through Alternative Routes) hold highly AI-exposed jobs, including 11 million in "gateway" jobs that have historically led to better-paid work. Only 51% of the paths from gateway jobs to better-paid jobs avoid high exposure. Opportunity@Work's 2026 State of the Paper Ceiling report says STARs have regained 783,000 good jobs after losing access to 7.4 million in 2000–2020, and that 33 states have dropped degree requirements for many public jobs. These are advocacy-organization figures; partner-network data come from self-selected employers.

• Harvard Business School and Burning Glass Institute (Joseph Fuller and co-authors) · projection · 2025

  "The Expertise Upheaval" projects that generative AI will change the learning curves of about 50 million U.S. jobs, potentially making many entry-level roles obsolete while opening "mastery" roles to less-experienced workers. These are estimates from skills and job-posting data, not observed job losses.

• MIT (Autor and Thompson) · primary research · 2025

  Across 303 occupations from 1980 to 2018, automation that removes the expert parts of a job lowers wages but widens access, while automation that removes routine parts raises wages but narrows who can do the job. Which way AI cuts for a given white-collar role is an empirical question, and it is the right question for mid-career workers to ask about their own jobs.

The economy-wide picture and the policy response

• MIT (Acemoglu) · peer-reviewed model · 2025

  "The Simple Macroeconomics of AI" estimates AI will raise total factor productivity by no more than about 0.66% over ten years, and by less than 0.53% once harder-to-automate tasks are accounted for. This is a calibrated model, not observed data, and it sits at the cautious end of published estimates.

• MIT Stone Center on Inequality and Shaping the Future of Work · commentary · 2026

  Acemoglu, Autor and Johnson argue in a Hamilton Project paper (Feb 2026) that AI's labor-market effects depend on design choices, and that "more than six out of 10 workers in 2018 were employed in occupational specialties that did not yet exist in 1940." The center launched in November 2025.

• Stanford HAI AI Index 2026 · compilation · Apr 2026

  The Index's economy chapter compiles other organizations' data, including that one-third of organizations expect AI to reduce their workforce in the coming year (McKinsey survey) and that measured productivity gains range from 14–15% in customer support to 26% in software.

• New York FutureWorks Commission · policy · 2026

  Governor Hochul launched the commission on March 19, 2026 to advise on workers' economic security and to build ways to track AI's effect on New York jobs in real time. Its co-chairs are Tom Perez, Thasunda Brown Duckett and Molly Kinder, and recommendations are due by the end of December 2026. Separately, New York's layoff-notice system has asked employers since March 2025 whether automation contributed to layoffs; in its first year, none of more than 160 filers said yes (Hunton Andrews Kurth, May 2026). Whether that reflects reality or under-reporting is unknown.

• RAISE US · program · launched Jun 25, 2026

  A new bipartisan nonprofit led by Gina Raimondo and Eric Holcomb, with more than $500 million in commitments from AI companies (including OpenAI's foundation, Anthropic, Microsoft and Amazon) and other employers, to pilot retraining, wage insurance for workers who take lower-paying jobs, and career navigation in four states. It has not yet published outcome data. (It is unrelated to New York's RAISE Act, an AI-safety law with no employment provisions.)

• Markle Foundation · status note

  Markle's skills-based hiring work, the Rework America Alliance, moved to Jobs for the Future in December 2023. We found no 2025–2026 Markle research on AI and jobs; we use Opportunity@Work and Jobs for the Future for current skills-based hiring evidence.

• MIT Martin Trust Center (Paul Cheek) · book · Aug 2026

  In No One Works Here (Wiley, 2026), Paul Cheek argues that firms will be reorganized around AI decision-making, and that hierarchies built around the limits of human communication become "structural debt." It is an argument, not a measurement, but it describes the organizational redesign that this report's Q3 addendum tracks through company filings and services firms.

What this may mean What this adds up to: the measurable effect of AI today is concentrated at the entry point to white-collar careers and runs through hiring, not layoffs. Our own data show the other end of the problem, longer searches for experienced workers who lose jobs, but cannot attribute that to AI. Interest rates, the post-2022 technology-sector adjustment and changes in federal employment are among the other factors that may contribute.

According to BLS, 1.94 million people had been unemployed 27 weeks or more in September, 27.1% of all unemployed, up from 23.6% a year earlier. Most of them have been out far longer than six months: about two in three have been looking for a year or more.

Age

Age shapes how long searches last more than whether people lose jobs. Among unemployed white-collar workers aged 55–64, 39% have been looking 27 weeks or more, up from 27% a year earlier, a rise larger than its 90% margin of error. Younger white-collar workers have somewhat higher unemployment rates but shorter observed spells so far, while older workers are much more likely to have been unemployed 27 weeks or longer; these tables do not measure how often people lose jobs or how fast they are rehired. Across all workers, unemployment for ages 25–34 rose to 4.8% from 4.3% (beyond margin of error), consistent with research showing fewer entry-level openings.

• Résumé gaps · Socio-Economic Review (2026)

  A meta-analysis of 28 résumé experiments finds gaps under six months do not hurt callbacks; penalties begin around 12 months and are larger for people who appear to have stopped looking.

• Scarring · Goldman Sachs (Apr 2026)

  Workers displaced by technology search about a month longer and see roughly 10 points less earnings growth over a decade; vocational training within three years of job loss is linked to better outcomes.

The degree-holding core: education and experience

This report defines white-collar work by occupation (management, business, financial and professional jobs), not by degree; about 70% of that workforce holds a bachelor's degree or more. Because most of the jobs NextChapter's readers seek require a degree, we also track workers aged 22 and older by education, and within degree holders by years of potential experience (age minus years of schooling minus six; the survey does not record actual experience).

Since 2022, essentially all net employment growth among adults 22 and older went to people with a bachelor's degree or more (+4.5 million; employment of those without a degree was flat). Yet long-term unemployment rose more for degree holders than for others, and in the past year it rose fastest for master's holders. By experience the pattern is U-shaped: new entrants and the most experienced degree holders saw the largest rises since 2022, while mid-career workers (15–24 years) saw the smallest and the fastest employment growth. Workers with 25 or more years of experience who are unemployed are the most likely to have been looking six months or more (35%).

• Degree requirements in postings · Indeed Hiring Lab

  19.3% of U.S. job postings required a bachelor's degree or higher in November 2025, against 20.4% in 2019; about half of postings name no education requirement. Indeed changed its method between releases, so the figures are not one continuous series. Lightcast finds degree requirements down about 14% since 2019 in jobs that traditionally required a four-year degree, but more common in postings that require AI skills (Sep 2026).

• Experience requirements · Indeed; ZipRecruiter

  Postings asking for years of experience fell from about 40% (2022) to 30% (2024), but in tech the share asking for five or more years rose from 37% to 42% (Q2 2022 to Q2 2025). By May 2026, senior postings were up 14.7% over the year and entry-level postings down 7.5%; in software development 69% of postings were senior and under 5% entry-level. 31% of hiring managers say AI has raised experience requirements for their entry-level roles (ZipRecruiter survey, Jun 2026).

• No official openings data by degree · BLS

  The Labor Department's job-openings survey has no education or experience fields. Occupations whose typical entry requirement is a bachelor's degree or higher hold 30% of employment and are projected to grow faster than average through 2035 (BLS Employment Projections, Sep 2026).

Demand for degree holders is still growing, but the people in the degree-holding core who lose jobs are taking longer to land, especially at the two ends of a career. Employers are asking for more seniority in the jobs they post, which narrows the entry path for new graduates, while long-tenured candidates face the longest searches. Mid-career professionals are, for now, the best-positioned group.

Workers 55 and older are less likely to lose a job than younger workers, but when they do, they are much less likely to get back to an equivalent one. Participation for ages 55+ fell to 37.1% in January–August 2026 from 38.1% a year earlier, but almost all of that drop reflects a shift toward older ages within the 55+ population, from aging and from the Census Bureau's January 2026 population revisions. Within each age and sex group, participation was essentially unchanged (see the correction below).

What happens after a layoff

The Bureau of Labor Statistics' Displaced Worker Survey (January 2026, released August 27, 2026) counted 3.3 million long-tenured workers displaced in 2023–2025. By January 2026, 72.9% of those aged 25–54 were re-employed, against 57.3% of those 55–64 and 38.6% of those 65 and older. Across ages, only about 49% of workers who moved from one full-time job to another were earning as much as before, down from about 62% in the 2024 survey. For an experienced professional, the risk is less the layoff itself than the long road back to comparable pay.

• Age bias · AARP (2026)

  64% of workers 50+ have seen or experienced age discrimination at work and 22% felt pushed out. In AARP/NORC polling through summer 2026, two-thirds of adults 50+ think finding a new job would be hard, and 35% of those name age discrimination as the main barrier.

• Employers · SHRM (Oct 2025)

  93% of organizations have no formal program to recruit older workers, even though 88% of HR professionals rate older workers' performance as equal to or better than others'.

• Money pressure · AARP/NORC (2026)

  39% of working adults 50+ say covering basic expenses is their main reason for working, and nearly a third expect to retire later than planned, most often because of living costs.

• AI use by age · AARP

  30% of adults 50+ use AI, up from 18% a year earlier (AARP, Dec 2025). The most common bias older workers report is the assumption that they lack tech skills (33%).

• Litigation · Mobley v. Workday

  The nationwide age-discrimination case over AI screening of applicants 40 and older remains in discovery; in June 2026 the court denied most of Workday's motion to dismiss, according to law-firm summaries.

What this may mean Reading the data: the headline fall in 55+ participation is mostly arithmetic, as the 55+ population ages. The clearer signals for older professionals are long-term unemployment (37% of unemployed white-collar workers 50+ have been looking 27 weeks or more) and lower re-employment after a layoff.

Applications per job have roughly doubled since 2022, increasing competition for each opening; separate data from Ashby suggest the chance of moving from application to interview has also fallen. Candidates use AI to apply faster, employers use AI to screen faster, and trust on both sides has fallen.

• How many applications it takes

  Among Huntr users, two-thirds of offers came within 50 applications. Interview rates fall as volume rises: 9.3% per application for people sending 11–20, 2.6% for people sending 100 or more. In that sample, seekers with 20+ years of experience had the highest interview rate (9.2%).

• Recruiters are overloaded

  Greenhouse's North American data show a paradox: from 2022 to 2025, annual applications per recruiter tripled (1,610 to 4,890) as recruiters per organization fell from 10.4 to 4.6, recruiters made more than twice as many hires a month (2.2 to 4.9), and yet time to fill rose from 43.6 to 59.7 days.

• AI on both sides

  In Greenhouse's November 2025 survey (U.S. subset: 1,200 job seekers, 665 hiring professionals), 70% of hiring managers said AI leads to faster, better decisions, but only 8% of job seekers called it fair. 41% of U.S. job seekers said they had put hidden instructions in résumés to get past AI filters. Only 21% of recruiters were very confident their AI does not reject qualified candidates.

• Fake jobs, fake candidates

  69% of job seekers said they had seen fake job postings (Greenhouse). Gartner forecasts that one in four candidate profiles worldwide will be fake by 2028; that is a prediction, not a measurement.

• Employer adoption · SHRM (Apr 2026)

  39% of organizations use AI in HR, rising to 60% of those with 5,000+ employees; recruiting is the most common use.

• Law and litigation

  Illinois began regulating AI in hiring decisions on January 1, 2026. California's rules on automated decision systems (October 2025) hold employers responsible for their vendors' tools. Colorado replaced its AI Act with a narrower law effective January 2027. In Mobley v. Workday, a nationwide age-discrimination case over AI screening of applicants 40 and older, about 14,000 people reportedly opted in; the case is in discovery. The EEOC withdrew its AI hiring guidance in January 2025.

• The AI-vs-AI loop · research

  When language models screen résumés, they favor résumés written by the same model 67–82% of the time; in simulations, candidates using the screener's own model were 23–60% more likely to be shortlisted (Xu, Li and Jiang, University of Maryland, NUS and Ohio State, 2025). On a large freelancing platform, the hiring premium for customized applications disappeared after AI writing tools arrived, so cover letters stopped signaling effort (Galdin and Silbert, Dartmouth and Princeton, 2025). In a 70,000-applicant field experiment at a call-center firm, AI-led interviews raised offers 12% and job starts 18%, with humans making final decisions (Jabarian and Henkel, 2025; the firm sponsored the study).

• Volume · LinkedIn; Ashby

  LinkedIn reported about 11,000 applications per minute in mid-2025, up 45% from a year earlier. Referrals are about 1% of applications in Ashby's data but reach interview at about 40%, versus about 3% for inbound applicants.

• A myth to drop

  The widely repeated claim that "75% of résumés are never seen by a human" traces to a 2012 marketing claim by a defunct company with no published method. Surveys suggest automatic rejection before human review is far lower, mostly through knockout questions.

AI skills now carry a measurable pay premium, and demand for them has spread well beyond tech. Inside companies, career ladders have slowed: promotions and internal moves are down since 2022 on every payroll and HR platform we track. Colleges are adding AI programs at every career stage, from first-year requirements to five-day executive courses, but there is little evidence yet on what they are worth. Public retraining is covered in Chapter 13.

• Pay premium · Lightcast (Jul 2025)

  U.S. postings that list AI skills offer 28% higher pay, about $18,000 a year; 51% are outside IT and computer science. Demand for AI skills grew 66% a year in HR, 50% in marketing and 40% in finance.

• Pay premium · PwC (Jun 2026, global)

  Workers with AI skills earn a 62% wage premium on average, up from 57%. PwC measures worker wages and Lightcast posted salaries, so the two are not directly comparable.

• Fastest-growing skill · McKinsey (Nov 2025)

  Demand for "AI fluency" in postings grew almost sevenfold in two years.

• Rising skills · LinkedIn; WEF

  LinkedIn's 2026 U.S. list of fastest-growing skills (as reported by EdTech Innovation Hub) is led by AI implementation, workflow automation and AI business strategy. The World Economic Forum expects about 39% of core skills to change by 2030.

The mid-career and senior skills gap

The skills gap for experienced professionals is less about technical knowledge than about hands-on use of AI and the training to build it. Employers increasingly screen for it; few are paying to close it.

• Usage by role · Gallup (May 2026)

  Workers whose managers support AI are 1.7 times as likely to use it frequently.

• Training gap · BCG (Jun 2026, 14 countries)

  88% of workers think they will need major new skills within five years; only 36% say they get enough training.

• Hiring screens · Patriot Software (Jul 2026, small businesses)

  40% of small employers rejected a candidate in the past year for weak AI skills: 65% of Gen Z employers and 26% of Gen X employers. Experienced candidates are increasingly interviewed by younger managers who expect AI fluency.

• Entry-level spillover · NACE (2026)

  13.3% of entry-level postings now ask for AI skills, nearly three times the fall 2025 level, a sign of how quickly the expectation is spreading.

• Who gets trained · WEF; Urban Institute

  Of every 100 workers, 59 will need training by 2030 and 11 are unlikely to receive it (WEF). 92% of workers 50+ want to learn new skills, but only 10% have taken AI training in their field (Urban Institute).

• Does retraining pay? · Workcred; Chicago Fed

  Older trainees in federal workforce programs mostly choose short credentials, and industry certifications do best, but most still earn less than before their job loss. Earlier research found a year of community college raised older workers' earnings 8–10%. There is little recent rigorous evidence on mid-career AI retraining.

Career ladders: fewer promotions, fewer internal moves

No federal survey measures promotions, so we rely on payroll and HR-software companies, each describing its own customers. Their definitions differ, but they agree on direction: since 2022, fewer people are being promoted, fewer jobs are filled from inside, and manager roles are being thinned. Slower ladders matter for people out of work too: each promotion usually opens a seat below it.

Two channels of advancement have narrowed at once: moving up inside a company and moving out for a big raise. Workers who switch jobs still earn more than those who stay (Chapter 04), but double-digit raises from switching are less common than in 2022. For an experienced professional who loses a job, a market with fewer internal promotions also means fewer vacancies created by promotion chains.

Employer training: two surveys, two directions

Whether employers are spending more to retrain their workers depends on which survey you read. The two main industry surveys disagree on both spending and hours. The one consistent signal is that AI's share of training is rising from a very small base.

• Career conversations · LinkedIn (2025)

  Only 15% of workers say their manager helped them build a career plan in the past six months, down 5 points from 2024 (Workplace Learning Report 2025; we found no 2026 edition).

• Tuition benefits · SHRM; Mercer

  Nearly half of employers offer tuition reimbursement (SHRM, 2025). The 2025 federal tax law made tax-free employer student-loan repayment permanent, with the $5,250 annual limit indexed to inflation after 2026.

How colleges are responding to AI, by career stage

Colleges are adding AI at every stage of a career. Undergraduates face new AI requirements, working professionals have low-cost online master's degrees, MBA programs are making AI compulsory, and executive programs charge up to $18,500 for a week. What is missing is evidence on outcomes: we found no rigorous placement or wage data for graduates of the new AI degrees.

• Students are moving from computer science to AI · NSC; CRA; MIT

  Computer science enrollment fell in fall 2025 at every award level, by 3.6% to 14.0% (National Student Clearinghouse), and new computer science bachelor's enrollments fell 12.9% (CRA Taulbee Survey). AI majors are growing: MIT's AI and Decision Making major went from 37 students in 2022 to 372 in 2025.

• AI companies on campus

  The California State University system gave ChatGPT Edu to more than 460,000 students (OpenAI, 2025). Anthropic's Claude for Education runs at Northeastern, the London School of Economics and Syracuse. Google committed $1 billion over three years to AI education at U.S. colleges and nonprofits; Microsoft's Elevate program, $4 billion over five years.

• Older learners · edX (Nov 2025)

  72% of Gen X and Baby Boomer adults without formal AI education say free courses would most motivate them to learn. We found no federal data on enrollment in AI programs by age.

• Does it pay off? · Oxford Internet Institute (Jan 2026)

  In a hiring experiment with about 1,700 recruiters in the U.S., U.K. and Germany, listing AI skills raised the chance of an interview invitation by about 8–15 percentage points and partly offset disadvantages of age and education. The study measures interview invitations, not jobs or pay.

• Skeptics

  "Is it glitz or is it substance?" asked Andrew Armacost, president of the University of North Dakota (Boston Globe, Jun 2026). The Burning Glass Institute calls AI credentials "a wild, wild west." Employers rank strategic thinking and problem-solving above AI literacy in MBA hires (GMAC), and an unverified industry index finds 42% of 60 business schools require no AI coursework.

For an experienced professional, the most accessible options are also the cheapest: free alumni courses and online master's degrees under about $16,000. The most expensive option, executive education, has the least outcome evidence. Until colleges publish placement and earnings data for AI programs, the safest reading is that an AI credential helps get an interview, not that it gets a job.

The cost side of the degree: price, completion, payback and debt

If the bachelor's degree is the entry ticket to most of the jobs this report tracks, its cost, the odds of finishing and the debt it leaves matter, especially for someone who is out of work.

• Debt and unemployment · data gap

  No public data source reports student debt or delinquency by employment status or by how long someone has been unemployed. The Federal Reserve's household survey splits payment trouble by income only. We therefore cannot say how much debt the long-term unemployed carry; we can say that payment trouble is concentrated among lower earners, which is where long searches push people.

• What happens to payments when you lose a job

  Federal borrowers can currently defer payments for unemployment (generally up to three years, with interest still building on unsubsidized loans). For loans made on or after July 1, 2027, unemployment and economic-hardship deferment are eliminated and forbearance is capped at 9 months in any 24; a borrower with no income on the new Repayment Assistance Plan still owes $10 a month. The Education Department paused wage garnishment and Treasury offsets for defaulted loans in January 2026, and no restart had been announced by September 30.

• Graduate borrowing caps · July 1, 2026

  Grad PLUS loans ended for new borrowers. Graduate students can now borrow $20,500 a year ($100,000 total) and professional students $50,000 a year ($200,000 total); Parent PLUS is capped at $20,000 a year per child. The new Repayment Assistance Plan opened July 1, 2026, and the SAVE plan ended, with 7.5 million enrollees told to choose a new plan. For mid-career professionals considering a master's to change fields, these caps mean more out-of-pocket or private borrowing.

For most graduates a bachelor's degree still pays, but the margin depends on finishing, on the program and on not being out of work long. A long search hits degree holders twice: lost earnings and loan payments that continue unless a deferment is in place, and the deferment safety net is narrowing for new loans from 2027.

Alumni programs

• Alumni view · NACM (2024)

  In a survey of 9,000+ alumni at 34 institutions, 49% rated their school highly on career preparation and 23% on its ongoing investment in their careers.

• Adult learners

  First-time college students aged 25+ fell 15.5% from fall 2024 to fall 2025, as reported by Inside Higher Ed.

AI is changing which jobs grow, not yet how many there are. Postings that mention AI doubled in a year, to 6.7% of all U.S. postings. The government's new ten-year projections expect total employment to grow 3.5%, led by health care, energy and technology occupations, with declines concentrated in clerical, customer-service and routine programming work.

Source: BLS Employment Projections 2025–35 (August 27, 2026). Total employment is projected to grow 3.5%, to 176.2 million. BLS expects AI-powered automation to reduce office and administrative support employment by 4.0% (−752,100 jobs) and cites AI in projected losses in sales.

• AI demand · Indeed Hiring Lab

  Postings that mention AI reached 6.74% of U.S. postings on August 31, 2026, up from 3.44% a year earlier and 1.70% in 2019 (NextChapter calculation from Indeed's published tracker data).

• Rising roles · LinkedIn (Jan 2026)

  AI engineers and AI consultants top LinkedIn's 2026 U.S. Jobs on the Rise; data annotators (#4), independent consultants (#7) and founders (#9) also made the list, as reported by Allwork.Space.

• Programmers vs. developers

  BLS projects computer programmers to shrink while software developers grow: the work of writing routine code is being automated, while designing and integrating systems is not. Fortune, citing a Washington Post analysis of BLS survey data, reports programmer employment at its lowest since 1980.

• Skill demands · PwC (Jun 2026, global)

  AI-specialist roles are growing 69% a year against 9% for all jobs, and entry-level roles in AI-exposed fields are 7 times more likely to ask for senior-level skills.

• Global view · WEF (2025)

  Employers expect 170 million jobs created and 92 million displaced worldwide by 2030; cashiers, administrative assistants and graphic designers are among the fastest-declining roles.

Americans are filing to start businesses at near-record rates, and AI is making it cheaper to launch one. But the share of workers who are self-employed has not risen, so the boom in filings has not yet become a broad shift of workers between jobs into running their own firms.

• AI lowers the cost of starting · Gusto (May 2026)

  60% of founders who started businesses in 2025 used AI to launch, up from 21% in 2023. Financial stability (51%) now outranks autonomy (46%) as the main reason for starting.

• Smaller founding teams · Carta; Stripe

  Solo founders were 36.3% of new startups in the first half of 2025, up from 23.7% in 2019 (Carta). 42% of companies formed through Stripe Atlas in 2025 were AI-focused, and 20% had a paying customer within 30 days.

• Business AI adoption · Census Bureau

  Between 17% and 20% of businesses reported using AI from December 2025 to May 2026, but under 20% of firms with four or fewer employees, against 37% of firms with 250 or more.

• Independent work · MBO Partners (2025)

  72.9 million Americans did independent work, 27.6 million of them full time, and 74% of independents used generative AI.

• Funding · SBA

  The SBA's 7(a) and 504 programs lent a record $44.8 billion in fiscal 2025, including $5.6 billion to startups.

For professionals between jobs: consulting, fractional and founder roles are growing options, and AI tools lower the cost of trying one. The data so far show more people filing to start businesses, not more people earning a living from self-employment, so a bridge business works best alongside a search, not as an assumed replacement for salary. No rigorous 2025–2026 study yet tracks how many laid-off professionals start businesses.

Over the year to August 2026, job growth was fastest in South Carolina, New Mexico and Louisiana (+1.6% each), and Texas added the most jobs (+159,400). The District of Columbia was the only place with a statistically significant decline (−27,000, −3.6%), driven by government job losses.

For white-collar work the split is sharp. Professional and business services jobs grew about 3% in Texas and North Carolina but fell in Virginia (−1.8%), DC (−2.7%), California and Massachusetts. Financial-activities jobs fell in California, Washington, Massachusetts, Virginia and Maryland. Unemployment rates are highest in DC (5.7%), California, Connecticut and Oregon (5.1% each) and Michigan (5.0%), and lowest in South Dakota (2.0%) and North Dakota (2.2%). Connecticut (+1.0 point) and Oklahoma (+0.9) had the largest over-the-year increases; New Jersey (−1.2) and Ohio (−1.1, a record low of 3.3%) improved most.

In the 12 months to August 2026, 39.3% of unemployment insurance claimants exhausted their benefits. Claimants drew benefits for 15.8 weeks on average, at $493 a week. These are completed benefit spells; the 24.8-week average in Exhibit 7 is for searches still in progress, so the two measure different things. Together they suggest many long-term job seekers are past the end of their benefits.

• Uneven access

  State recipiency rates ranged from 8% to 55% of the unemployed in 2025 (Minneapolis Fed). Florida and North Carolina cap benefits at 12 weeks. Virginia raised its maximum weekly benefit to $478 and Iowa to $790 in July 2026.

• Severance

  Some states delay or offset benefits during paid severance; executives should check their state's rules before filing.

• Paid family leave

  Delaware and Minnesota began paying benefits January 1, 2026, and Maine on May 1; Minnesota received about 100,000 applications in six months. Maryland's start is delayed to 2028. Returnship programs continue at large employers, but we found no reliable 2026 count.

• AI layoff disclosure

  California's SB 951 takes effect January 1, 2027. A federal bill, the AI-Related Job Impacts Clarity Act, would require large employers to report AI-related layoffs to the Labor Department.

Public retraining: what WIOA offers a laid-off professional

The main federal retraining program for laid-off workers is the Dislocated Worker program under the Workforce Innovation and Opportunity Act (WIOA). In program year 2024 it served about 187,000 people nationwide. Funding is level, about one participant in five received training, and earnings after leaving are at middle-wage levels. Its scale is small next to the roughly 546,000 white-collar workers searching 27 weeks or more, though not every long-term unemployed worker is eligible for it or would use it.

• Funding status and proposals

  The fiscal 2026 spending law (Feb 3, 2026) held WIOA formula grants level, and a stopgap signed September 2 (H.R. 6500) extends those levels through December 11, 2026. Three proposals are pending: the President's fiscal 2027 budget proposes combining 12 programs into a $3.4 billion "Make America Skilled Again" block grant; the House fiscal 2027 Labor-HHS bill, approved in committee in June, would reduce WIOA's main formula grants by about 62% (Youth eliminated, Adult nearly eliminated, Dislocated Worker largely unchanged); and a reauthorization bill, H.R. 8210 (A Stronger Workforce for America Act), passed committee 19–14 in April. We found no enactment of these proposals as of September 30.

• AI-specific guidance · Labor Department

  Guidance issued in August 2025 (TEGL 03-25) lets states spend WIOA funds on AI literacy and training, and the department published an AI Literacy Framework in February 2026. In June 2026, TEGL 15-25 offered about $50 million in Rapid Reskill Employment Recovery National Dislocated Worker Grants, which name "the broad and accelerating economic transformation driven by AI" as the qualifying event; it points grantees toward manufacturing, aerospace, energy, construction and semiconductors more than office work. We found no award announcements by September 30.

• Workforce Pell

  Short-term Pell grants for 8–15-week programs began July 1, 2026, and bachelor's degree holders can qualify. By September 28, 21 programs in six states had federal approval (Opportunity Data), mainly in trades and health care such as EMT and medical-assistant training, not white-collar retraining.

• Trade Adjustment Assistance

  The federal program for workers displaced by trade has accepted no new petitions since its authority lapsed on July 1, 2022. No federal program is specifically designated for workers affected by technology change.

• Does training work? · Mathematica (2019); sector studies

  The national evaluation of WIOA's predecessor found staff assistance raised earnings by about $7,100 over 30 months; the effect of training was inconclusive. Sector programs built with employers (Year Up, Per Scholas, Project QUEST) show large, lasting gains, but mainly for young and low-income adults. We found no rigorous evaluation of retraining for displaced managers and professionals.

• Review of trials · Anthropic (Aug 2026)

  A review of 56 U.S. randomized trials finds training raises employment 2–3 points and earnings about $1,000 a year, at roughly $13,000 per participant; sector programs built with employers do several times better.

• State partnerships · California (Jul 2026)

  "AI-Ready California" offers free AI-literacy micro-credentials through San Diego State, including to unemployment-insurance claimants, and the state now tracks jobless claims in AI-exposed occupations monthly.

WIOA's measured outcomes describe a participant group whose median earnings after leaving are about $40,000 a year, and training spending averages about $7,300 per trainee. We found no outcome data for public programs among experienced managers and professionals specifically. For this group, the supports that appear in practice are severance, employer-paid outplacement, alumni programs, self-funded courses and the programs described in Chapter 13A.

Federal, state and local governments are adding to how people move between jobs: reemployment services for people who file for unemployment insurance, funding tied to apprenticeship and training results, short-term Pell aid, and public data and tools. This chapter describes a selection. For each we give who runs it, what it does, its scale and what outcome evidence exists. The evidence is uneven: one program has a large randomized trial, one has descriptive wage data, and the newest have none yet. Programs are described as sourced; we do not rank them or assess policy.

Design ideas that recur, and what the evidence says about each

• Pair eligibility checks with reemployment help

  RESEA is the best-tested idea in this chapter. The Rhode Island trial found shorter time on benefits and higher wages and reemployment. The trial covered all claimants in one state, so the effect for experienced professionals is not known.

• Release funding at outcome milestones

  The apprenticeship awards pay sponsors on retention and progression, and Workforce Pell requires 70% completion and placement rates. The approach ties payment to results the programs already track. We found no evidence yet on whether it raises outcomes.

• Build the outcome data first

  The Texas grant funds data systems before programs scale, which addresses the gap this report keeps finding: many training programs have no outcome data for the people they serve.

• Publish early-warning data

  California's tracker makes claims in AI-exposed occupations visible monthly, with the breakdowns needed to target help. Whether it changes service decisions has not been measured.

Most of these programs accept outside participation: employers, colleges, nonprofits and private providers serve as apprenticeship sponsors, consortium members, evaluation partners and data contributors. None of the programs above was designed specifically for experienced managers and professionals, and our data cannot say how many white-collar job seekers use them.

Public programs add resources at several points in a job change: the first weeks of a claim, training, and the data that show what works. The strongest evidence is for early reemployment services. For training and for experienced professionals, the evidence is thinner, and the new awards are designed to produce it. This chapter will be updated as outcome data appear.

The two groups outside the white-collar core are moving in opposite directions. New college graduates face the weakest entry-level market in years, and their unemployment advantage over peers without degrees has narrowed. Skilled trades are short of workers, wages are rising faster than average, and AI exposure is low.

New college graduates

The degree still pays, but less of a cushion than it used to. Young graduates' unemployment rate is 1.5 points above 2019 while young high-school graduates' rose 0.8 points, so the gap between them has narrowed from 3.0 to 2.2 points. The Cleveland Fed finds young graduates' job-finding rates have fallen to roughly match high-school graduates', the narrowest gap since the late 1970s (as reported by Fox Business).

• Official data · NY Fed; BLS

  Recent graduates aged 22–27 had 5.6% unemployment and 42% underemployment in Q2 2026. Unemployment for all 20–24-year-olds fell to 8.0% in September from 9.2% a year earlier (BLS, seasonally adjusted).

• By major · NY Fed (2024 data)

  Computer engineering (7.8%) and computer science (7.0%) are among the five majors with the highest unemployment, yet they still have low underemployment and median early-career pay near $87,000–$90,000. Criminal justice and performing arts graduates are most often underemployed (66% and 64%). Figures as reported by Research.com.

• Hiring and pay · NACE (2026)

  Employers planned to hire 5.6% more class-of-2026 graduates and 45% rated the market "fair." Projected starting salaries: computer science $81,535 (+6.9%), engineering $81,198, business $68,873.

• Entry-level hiring · Stanford; Anthropic; PwC

  Employment of 22–25-year-olds in the most AI-exposed jobs is about 19% below less-exposed peers (Stanford, Aug 2026). Job starts in AI-exposed occupations for that age group fell about 14% after ChatGPT launched (Anthropic). Entry-level roles in AI-exposed fields are 7 times more likely to ask for senior-level skills (PwC).

• AI readiness · Handshake; NACE; ZipRecruiter

  85% of 2026 seniors use AI, but only 28% say their programs meaningfully built it in, while 13.3% of entry-level postings already ask for AI skills (NACE). Fewer than 30% of 2026 graduates received substantial AI training (ZipRecruiter survey). 62% of seniors are pessimistic about the market, up from 46% two years ago (Handshake).

• What helps · ZipRecruiter; Strada

  82% of 2026 graduates with work experience found jobs, against 41% without. Graduates whose first job requires a degree are 3.5 times as likely to be in a college-level job ten years later (Strada), which is why the first job matters so much.

Blue-collar and skilled trades

• Shortages · ABC; HBI

  Construction needs about 349,000 new workers in 2026 and 456,000 in 2027, mostly to replace retirees; about one in five electricians is over 55 (Associated Builders and Contractors, Jan 2026). The skilled-labor shortage costs homebuilding about $10.8 billion a year (Home Builders Institute).

• Data centers · BLS; IBEW/NECA

  BLS projects electricians to grow 9% over 2025–35, with about 72,700 openings a year, citing AI and data centers. Electrical work is 45–70% of the cost of building a data center, and union apprenticeship applications rose from about 70,000 to 120,000 between 2022 and 2024 (Fortune, citing IBEW and NECA).

• Training pipeline · DOL; National Student Clearinghouse

  Registered apprentices number about 700,000, the first year-over-year dip in a decade (Bloomberg Law, citing DOL data). Enrollment at vocational-focused community colleges rose 2.8% in spring 2026, and undergraduate certificate programs grew fastest (+10.2%) while computer science enrollment fell.

• Automation · IFR (Jun 2026)

  U.S. industrial robot installations rose 11% to 38,000 in 2025, led by food and non-manufacturing industries rather than autos.

• AI exposure · Anthropic (Mar 2026)

  Construction, installation and repair, and transportation occupations show the lowest observed AI exposure; about 30% of workers have none.

What this may mean What it means for professionals: trades are not a realistic pivot for most experienced office workers, but the contrast explains why headline job numbers look steadier than white-collar workers feel. Adjacent roles do exist where office skills meet the building boom: project management, estimating, procurement, safety and operations in construction, energy and data-center firms.

The BLS jobs report is released at 8:30 a.m. Eastern, usually on the first Friday of the month. This report publishes within two business days after it; the White-Collar Index is updated when the Census Bureau posts that month's microdata, usually one to two weeks later.

Use-case watch: from AI deployments to jobs

Company announcements about AI are easy to find; their effect on jobs is not. Each month we pair a few well-documented deployments with a public jobs series for the affected industry, so readers can see whether employment moves the way the announcements imply. Company claims are not measurements, and a jobs series cannot separate AI from other causes. We list failures and reversals alongside successes.

Frontier signals: language and expectations

• "Super Intelligence" in federal usage · Sep 29, 2026

  Executive Order 14434 directs federal agencies to use "Super Intelligence" (SI) in place of "Artificial Intelligence" in their own documents. For now SI carries the same legal definition as AI (15 U.S.C. 9401(3)), and the President's science adviser has 60 days to propose a statutory definition. The order has no employment provisions. A voluntary "White House Accord on Super Intelligence" signed the same day by leaders of major AI companies commits them to safety controls and audits. This report continues to use "AI," the term in its data sources, and notes the federal terminology where relevant. The order changes terminology only.

• Recursive self-improvement (RSI)

  RSI refers to AI systems that design and develop their own successors. Anthropic reported that, as of May 2026, more than 80% of the code merged into its codebase was written by its own model, and has called for the option to slow or pause frontier development if needed; OpenAI was reported in September 2026 to have reached its goal of an automated AI "research intern." METR estimates the length of tasks AI can complete has been doubling roughly every six to seven months, but cautions that this does not mean tasks of that length can be delegated reliably. Skeptics, including Princeton's Sayash Kapoor, find AI agents still weak at original research. We include RSI not as a forecast but because it is the main argument for why AI's labor-market effects could accelerate faster than past technologies'.

The September edition of each quarter (March, June, September, December) carries a quarterly addendum. It uses data that only make sense over a full quarter or a year, such as company filings, business surveys and corporate results, and it asks a question the monthly figures cannot answer: are firms changing how they are organized, and is that showing up in hiring? Figures here are labeled as source-reported or as NextChapter calculations. Several Q3 government releases (GDP, the Employment Cost Index and productivity) come out in late October and early November and will be added in the next edition.

The quarter combined fewer announced layoffs with very little hiring. AI's share of stated layoff reasons fell in Q3 (14%, against 21% for the year), and so did the share of formal restructuring filings citing AI. One quarter is too short to call that a trend. The productivity and labor-share figures are economy-wide and say nothing on their own about AI, but they are the measures to watch if AI begins to raise output faster than employment.

What firms expected versus what happened

Business surveys let us compare what firms said they would do with what they later did. Four patterns stand out. Firms expect sales to grow three to four times faster than employment. Small firms keep saying they plan to hire, but their actual employment is shrinking. AI adoption has run ahead of expectations. AI-related layoffs have run well behind them.

The pattern fits a cautious, low-hire market more than a wave of AI substitution: firms expect to grow without adding much headcount, and only a small minority report that AI has reduced their need for workers so far. A larger minority expect it to in the next few years. That gap between expectation and outcome is exactly what this section will track each quarter.

What company filings say about AI and restructuring

We read every 2026 Form 8-K that references Item 2.05, the item public companies use to report material restructuring costs, and coded each AI mention by what role AI plays, following the evidence levels in Appendix B. A mention in a risk factor is not evidence that AI caused a layoff; a company's statement that its plan is designed around AI is a company's stated rationale (evidence Level 1), not proof of effect.

Two things stand out. First, when companies do cite AI in a restructuring filing, they split between AI changing how work is done (8 companies) and cutting elsewhere to fund AI (6 companies). The second is a reallocation story, not a substitution story. Second, nine in ten restructuring filings say nothing about AI. Filings are required to disclose the decision and its costs, not its reasons, so silence does not show that AI played no part.

Revenue per employee at large white-collar employers

Can large companies grow without adding staff? We compiled revenue and year-end headcount from the annual reports (10-Ks) of the 30 largest U.S.-listed companies by revenue in white-collar-intensive industries (software and internet, banking, insurance and managed care, financial services and professional services), comparing each company's most recent fiscal year with the year three years earlier. Twenty-seven reported comparable headcount in both years.

The pattern is striking, but it is not evidence that AI replaced workers. Under the evidence levels in Appendix B, this is at most Level 2 for any company that also attributes gains to AI. Higher interest rates lifted bank revenues; health insurers raised premiums; Meta, Alphabet and Oracle were correcting pandemic-era overhiring; and acquisitions change both lines. What it does show is that the largest white-collar employers are generating far more revenue per worker than three years ago without adding people, which is consistent with the low hiring rates in Chapter 02 and the "growth without hiring" expectations above. We will update this sample annually and add operating income per employee and payroll as a share of revenue where companies disclose them.

Is AI showing up in productivity? Total factor productivity

Output per hour can rise for three reasons: more equipment per worker (capital deepening), a more skilled workforce, or genuinely better ways of producing, which economists measure as total factor productivity (TFP). Only the last reflects new technology such as AI. The San Francisco Fed publishes a quarterly TFP series that also adjusts for how intensively existing workers and equipment are being used (Fernald, 2014).

Underlying productivity shows no AI boost yet. TFP adjusted for utilization grew 0.3% in 2025 and −0.3% over the four quarters to mid-2026, after a pickup to 1.5–1.75% in 2023–24 that has since faded. Output per hour is still strong (2.5% in 2025), but the main reason is that output kept rising while hours worked barely grew (0.2% in 2025), which raises equipment per hour. Capital services themselves are not growing faster than before the AI investment boom (2.8% in 2025 versus 3.0% in 2019). This is the economy-wide version of the "growth without hiring" pattern among the large employers above.

• Apollo (Torsten Slok) · commentary · Oct 5, 2026 (published after the reference period)

  Apollo's chief economist reaches the same headline conclusion from the same San Francisco Fed series: utilization-adjusted TFP is slightly below zero with no acceleration, while output per hour runs near 2.5%, which he calls "the signature of capital deepening, not of a technology shock"; the AI productivity payoff "remains a forecast rather than an observation." Our check of the underlying data agrees on the headline, with two qualifications: utilization-adjusted TFP did accelerate in 2023–24 before fading, and the capital-deepening contribution reflects weak hours growth more than faster growth in capital services. We include this commentary because it interprets data from the reference period.

• History

  Electricity and information technology each took a decade or more to show up in aggregate productivity (David, 1990; Brynjolfsson, Rock and Syverson, 2021). A flat reading now is consistent with both "AI is overhyped" and "the payoff is still coming."

Illustrative scenario: productivity gains and the cost of long searches

The labor-market services economy

The question: as job searches get longer and screening becomes automated, are workers and employers spending more on help navigating the labor market, such as recruiting, staffing, executive search, outplacement, coaching and job-search tools? The evidence says total spending is not rising, but its mix is shifting.

Staffing remains in a cyclical slump, and freelance marketplaces are shrinking, possibly because AI substitutes for some gig work. Growth is in executive search, permanent placement and paid job-board pricing. Application volume is the clearest sign of a navigation problem: Ashby, a recruiting-software company, reports applications per hire have roughly tripled since 2021 to more than 300 (vendor data). Two cautions apply. Most evidence on outplacement and coaching demand comes from the vendors themselves, and LinkedIn reports that applications per applicant fell 24% over the year to March 2026. We found no independent data on the size of the outplacement or career-coaching markets.

Remaking the organization: forward-deployed engineers and AI services firms

A new kind of company is being built specifically to redesign other companies' workflows around AI. This matters for white-collar workers because it moves AI from tools individuals choose to use toward the organization-wide redesign described in Appendix B's adoption stages, which is the stage most likely to change headcount.

• AI labs' services ventures · 2026

  Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs, with other investors, formed an enterprise AI services company (announced May 4, 2026; named Ode with Anthropic in July) that places engineers inside client companies. OpenAI formed a deployment company backed by TPG and other investors, with McKinsey, Bain and Capgemini as consulting partners, and acquired the London consultancy Tomoro. The engineers in these roles are called "forward-deployed engineers"; Fortune reported postings for the title rose more than tenfold in January–August 2026 from a year earlier.

• AI roll-ups of services firms · 2025–2026

  Thrive Holdings, backed by Thrive Capital with an OpenAI stake, has acquired more than 50 accounting firms and about 20 IT services firms (as reported by The Next Web) and automates their work with AI; it raised $2 billion in August 2026. Long Lake, backed by General Catalyst, has made about 40 acquisitions and agreed in May 2026 to buy Amex Global Business Travel. These models target exactly the back-office and professional-services work that employs many white-collar workers.

• Consultancies · 2025–2026

  Accenture took an $865 million restructuring charge in September 2025; CEO Julie Sweet said the firm was "exiting people so we can get more of the skills in we need." It had about 814,000 people at the end of its fiscal year on August 31, 2026. Bloomberg reported McKinsey plans to cut about 10% of non-client-facing roles. PwC cut 5,600 staff in the year to June 2025 and dropped its pledge to add 100,000 people. Deloitte U.S. replaced its analyst-to-manager titles with new job families in June 2026. Consulting pyramids, traditionally built on large junior cohorts, are the clearest test of the junior-compression hypothesis in the Future Research Agenda.

Annual feature: industry productivity versus long job searches

If AI-driven productivity gains were pushing white-collar workers into long searches, the industries with the fastest productivity growth might show the largest rise in long searches. We compare BLS total factor productivity growth by sector for 2019–2024, the latest available, with our measure of long searches among unemployed white-collar workers by the industry of their last job.

There is no clear pattern. The sectors with the fastest productivity growth (management of companies, professional and technical services, information) show rises in long searches, but none is larger than its margin of error. The only industry with a rise beyond the margin, education and health services, had productivity growth near zero. Newer detail for 2025 points in different directions within single sectors: software publishers raised labor productivity 12.9% while cutting hours 0.7%, and commercial banking 3.2% while cutting hours 0.3%, but engineering services fell 0.2%. Productivity rose in 15 of 30 service industries BLS covers (Aug 26, 2026) and in 42 states and the District of Columbia (May 28, 2026).

At the level of whole sectors, productivity growth and long searches are not yet moving together. That is not evidence against an AI effect: five-year sector averages are too coarse to detect effects concentrated in particular occupations, firms or the most recent year, and long searches depend on hiring across every industry, not only the one a worker left. We will repeat this comparison each third quarter, adding 2025 sector data and a state-level comparison.

Reemployment after job loss: the biennial check

The BLS Displaced Worker Survey, run every two years in January, measures what happens after a layoff. Of 3.3 million long-tenured workers displaced in 2023–2025, 66.1% were reemployed by January 2026 and about 49% of those earned as much as or more than before, so roughly half took a pay cut. After the 2009–2011 layoffs, only 56% were reemployed and about a third took cuts of 20% or more. Today's displaced workers are faring better than after the Great Recession, but reemployment is not recovery: half of those who found work earn less than they did. Reemployment rates for workers 55–64 are lower (see Chapter 07).

Every major wave of automation has eventually created more work than it destroyed. Every wave has also imposed large, long-lasting costs on the specific workers and places whose tasks lost value. History is a guide to the shape of an adjustment, not a forecast of AI's effects.

New work does appear, but not for the same people

About 60% of U.S. employment in 2018 was in job titles that did not exist in 1940 (Autor, Chin, Salomons and Seegmiller, Quarterly Journal of Economics, 2024). The same research finds that the job-destroying effect of automation was more than twice as strong after 1980 as in 1940–1980, while new work increasingly appeared in high-paid professional and low-paid service jobs rather than in the middle. Earlier waves of computerization hollowed out routine clerical and production work and polarized employment toward the top and bottom (Autor, Levy and Murnane, 2003; Goos, Manning and Salomons, 2014).

The costs of displacement are large and long-lasting

• Earnings after a mass layoff

  Long-tenured workers lose about 25% of annual earnings for many years after a mass layoff (Jacobson, LaLonde and Sullivan, 1993). Over 20 years, the loss averages 1.4 years of prior earnings when the layoff happens while unemployment is under 6%, and 2.8 years when it is above 8% (Davis and von Wachter, 2011). When you lose a job matters almost as much as whether you lose it.

• Industrial robots

  Each additional robot per thousand workers reduced the local employment-to-population ratio by about 0.2 points and wages by about 0.42% in 1990–2007 (Acemoglu and Restrepo, Journal of Political Economy, 2020).

• The China trade shock

  Import competition cost an estimated 2.0–2.4 million U.S. jobs in 1999–2011, including spillovers (Acemoglu, Autor, Dorn, Hanson and Price, 2016). Most of the adjustment in hard-hit areas came through people leaving work rather than moving away, and the damage persisted long after the shock (Autor, Dorn and Hanson, 2021).

• Engels' pause

  During Britain's early industrial revolution, output per worker rose 46% in 1780–1840 while real wages rose 12%. Wages caught up only in the following decades (Allen, 2009). Carl Benedikt Frey's The Technology Trap (2019) argues that such lags are the norm when technology replaces rather than complements workers.

• Bank tellers

  ATMs cut tellers per branch from 20 to 13 in 1988–2004, but banks opened 43% more branches, so teller jobs held up (Bessen). Two decades later, mobile banking did what ATMs did not: BLS now projects teller employment to fall 13% by 2035. The same technology can complement a job for years before substituting for it.

Predictions have often been wrong in both directions

Frey and Osborne's 2013 estimate that "about 47 per cent of total US employment is at risk" of computerization was widely read as a forecast of job loss, which the authors did not intend. Task-based estimates put the share closer to 9% (OECD, 2016), and the predicted wave did not arrive in the following decade. Exposure estimates for AI today, such as MIT's "Iceberg Index" finding 11.7% of U.S. wage value technically exposed, measure what AI could do, not what employers will do. Those are different questions.

What this may mean Lesson for this report: the history most relevant to white-collar workers today is not whether total employment survives (it has every time) but the duration and cost of transitions for the people whose tasks lose value. That is what our long-term unemployment measures track.

Remote work solves geography but not occupational transition. It has leveled off at about a fifth of all workers and a third of managers and professionals. Fully remote jobs are scarce and draw a heavily outsized share of applicants, and the evidence suggests they are also more exposed to hiring abroad. Each of these links is a hypothesis to test, not an established cause of longer searches.

• Competition for remote roles

  When 9% of postings draw 37% of applications, a remote search is a far more crowded search. That is consistent with, but does not prove, longer searches for professionals who restrict themselves to remote work.

• Offshoring of remote-friendly work · Revelio Labs

  From 2019 to 2024, U.S. firms' overseas headcount grew 32% against 16.7% at home, and roles suited to remote work grew 42% faster abroad, while roles that are not remote-friendly grew at similar rates in both. This is vendor analysis of online profiles, not official data. Richard Baldwin's "globotics" thesis predicts exactly this combination of remote work and overseas "telemigrants."

• Return-to-office mandates

  The Flex Index found 34% of U.S. companies required full-time office attendance in Q3 2025, with required office days up 12% since early 2024 but actual attendance up only 1–3%.

What this may mean What we can and cannot say: remote work widens the pool of competitors for each remote job, at home and abroad. We have no evidence yet linking remote work to longer white-collar unemployment spells. The CPS telework questions make that testable in future editions.

This report's central finding is that job cuts are down and jobs are growing, while searches are longer for people who leave white-collar jobs. Announced job cuts are down, AI appears in only 9% of formal restructuring filings, and professionals who keep their jobs are not being pushed into part-time or contract work. Yet professionals who lose jobs stay out far longer, and each job opening yields fewer hires than in 2019 or the early 2000s. One way to read that combination is that three clocks are running at different speeds. This is a framework for reading the data, not a finding, and it builds on older ideas: Paul David's account of the electric dynamo, Erik Brynjolfsson's "productivity J-curve" and Agrawal, Gans and Goldfarb's "between times."

Three speeds

Capability is moving fastest on the measures we have. METR estimates that the length of software task frontier models can complete half the time has doubled every four to seven months, from about 5 minutes for GPT-4 in 2023 to at least 16 hours in spring 2026; METR itself calls estimates above 16 hours unreliable, and the length models complete 80% of the time is only one to three hours. On OpenAI's GDPval, a test of well-specified professional tasks in 44 occupations, the share of tasks where models matched or beat experts rose from under half at its September 2025 launch to about 85% by April 2026, as reported by OpenAI. These are vendor-run, self-contained tasks, not jobs.

Organizational rewiring is slower. About 18–20% of U.S. firms report using AI in any business function, and roughly 95% of those report no change in headcount attributable to it (Census Business Trends and Outlook Survey, late 2025 to early 2026). Workers report more use than firms: 39% used generative AI for work in the previous week, and AI saved about 2.2% of all hours worked in the second quarter of 2026, up from 1.6% in late 2024 (St. Louis Fed Real-Time Population Survey). Use is spreading faster than redesign. Aggregate data agree: total factor productivity adjusted for utilization, the closest measure of technology-driven productivity, was −0.3% over the year to mid-2026 (San Francisco Fed).

The labor market is adjusting on a third timescale, and mostly through hiring rather than firing. Employment of 22–25-year-olds in the most AI-exposed jobs is 19% below comparable peers (Stanford Digital Economy Lab, August 2026). The Richmond Fed finds AI-related structural change "an important but not dominant factor" in the decline in job-finding since mid-2023. Our own data add longer searches for white-collar workers who lose jobs and no sign yet of displacement hidden inside jobs.

Task versus system

The constraint is that an AI system can perform a task without an organization being able to remove the human role around it. Roles bundle tasks with exception handling, accountability, legal responsibility, escalation, relationships and coordination, and they sit on data and legacy systems that were not built for AI. The relevant question is not whether AI can do the task, but whether the organization can redesign the whole workflow around it without losing value. History suggests that takes time: factory electrification passed half of U.S. horsepower around 1920 but raised productivity only after plants were redesigned, decades after the first power stations (David, 1990), and computers paid off mainly at firms that also reorganized work (Bresnahan, Brynjolfsson and Hitt, 2002). AI is delivered as software and may move faster; no study has yet estimated the lag from demonstrated capability to measurable job effects.

A postulate: the conversational threshold

In a 2025 test, judges who chatted for five minutes with a person and with GPT-4.5 picked the AI as the human 73% of the time; the model was prompted to adopt a humanlike persona (Jones and Bergen, UC San Diego). We put forward a hypothesis, not a finding: once machines pass this kind of conversational test, the first jobs to change are those whose core output is routine, remote conversation, such as customer service by chat and phone, ahead of other office work. If the hypothesis is right, those occupations should show shrinking employment, fewer young entrants and longer searches first, while conversation-heavy jobs that require physical presence should not.

The first evidence fits the postulate but does not prove it: employment of customer service representatives fell about 8% in a year, beyond the margin of error, and young workers' share kept edging down, while receptionists, whose work is conversational but in person, held steady. Offshoring, self-service and the slow economy also cut these jobs, and Gartner's surveys show most service leaders are keeping, and expanding, human roles for complex cases. We will track this monthly; the postulate is wrong if remote-conversation occupations do not diverge from comparable in-person ones over the next year.

Customer service: where the human role persists

Customer service shows the constraint clearly. In Gartner's survey of 321 service leaders (published April 2026), 85% were expanding human agents' responsibilities, even as 31% had made or planned AI-driven frontline layoffs and 63% were shrinking headcount through attrition. In a separate Gartner survey, 87% of customers said a company using generative AI for service must also offer a human agent. Klarna is the best-known case. In February 2024 it said its AI assistant handled two-thirds of customer chats, the work of about 700 agents; its November 2025 investor presentation shows company headcount falling from about 5,500 in 2022 to about 2,900, largely through a hiring freeze and attrition. In May 2025 its chief executive said cost had been "a too predominant evaluation factor" and that the result was "lower quality," and the company began piloting on-demand human agents. The case shows that automating a task is not the same as redesigning a service system. It does not show that humans "won," and there is no independent measure of Klarna's service quality.

Direction, not only speed

Daron Acemoglu, David Autor and Simon Johnson argue that the effect on workers depends on what kind of AI firms build and buy. In "Building Pro-Worker Artificial Intelligence" (NBER, February 2026) they distinguish labor-augmenting, capital-augmenting, automating, expertise-leveling and new-task-creating technologies, and write that "only the last category is unambiguously pro-worker." Acemoglu's own macroeconomic estimate is cautious, about 0.7% higher total factor productivity over ten years; other economists estimate gains roughly ten times larger per year (Aghion and Bunel, 2024). The rewiring view is compatible with both camps: the disagreement is about how fast, and in which direction, organizations rebuild work around AI.

How to read current data in this light: low counts of AI-attributed layoffs do not mean AI has no labor-market effect, and fast benchmark gains do not mean job losses are imminent. If the framework is right, effects show up first where rewiring is cheapest: positions that are not refilled, junior hiring, standardized workflows, and firms that grow without adding staff. They show up last in layoff announcements. The indicators to watch are hiring rates, search durations, hours and output per worker, not headlines about AI capability.

Some of the most important questions about AI and white-collar work cannot yet be measured rigorously. Rather than estimate them to make the report feel complete, we list them here with the data and validation each requires. No number appears for any of these measures until it meets its publication threshold.

Framework: four stages of AI adoption (a hypothesis, not a finding)

We organize this agenda around a hypothesis: AI's effect on jobs may depend less on whether workers use AI and more on whether companies redesign work around it. Stage 1, individual augmentation: AI helps a worker do the job. Stage 2, task substitution: AI performs part of the job. Stage 3, workflow automation: AI performs an entire workflow. Stage 4, organizational redesign: the company changes its structure. Most survey evidence today describes stages 1 and 2. The forward-deployed engineering firms and AI roll-ups described in the Q3 addendum are explicitly selling stages 3 and 4.