13% Fewer Young Workers in AI Jobs — The Quiet Automation

The direct answer to what generative AI is doing to office work in 2026 is this: it is automating specific, well-bounded tasks — first-pass code, routine customer contacts, HR queries, contract review, document drafting, translation and stock imagery — and firms are banking the gains mainly by slowing recruitment, not by mass dismissal ◈ Strong Evidence [3] [43] [44]. Payroll records from ADP show that workers aged 22 to 25 in the occupations most exposed to AI suffered a 13% relative decline in employment after late 2022, while older workers in the same jobs held steady or grew ◈ Strong Evidence [3]. Across the whole economy, by contrast, the Yale Budget Lab and Brookings found no discernible disruption 33 months after ChatGPT was released [86], and a study of 25,000 Danish workers found no significant effect on earnings or hours in any of 11 exposed occupations [4]. This report examines the work itself, function by function, and the gap between task-level gains and company-level results. The conclusion it reaches is that the most consequential labour-market effect of AI so far is a narrowing of the entry rail into professional work.

The scale of corporate intent is not in doubt. Challenger, Gray and Christmas, which tracks announced US job cuts, recorded AI as the most-cited reason for layoffs in 2026, with 120,136 cuts attributed to it by the end of September — about 21% of all announced cuts ✓ Established [1]. The share peaked at a record 40% in May before falling to 9% in September [1]. Amazon has removed about 30,000 corporate roles since October 2025, roughly 10% of its corporate staff, after its chief executive told employees that AI would reduce the corporate workforce over the next few years [45] [46]. Salesforce cut its support organisation from about 9,000 people to about 5,000 as AI agents took over roughly half of customer conversations [17]. Block cut about 4,000 of its more than 10,000 staff in February 2026 and said most companies would follow within a year [81].

Set against those announcements, the hard labour-market data look strikingly calm. The New York Fed’s regional surveys in August 2025 found that 40% of service firms were using AI, up from 25% a year earlier, but only 1% had laid anyone off because of it ✓ Established [43]. A year later the same bank concluded that AI-driven layoffs remained very rare and that firms preferred retraining to firing, describing the short-run effects as more evolutionary than disruptive [44]. US nonfarm productivity rose 2.1% in 2025 and 2.2% in the year to the second quarter of 2026, with hours worked almost flat [41]. That pattern — more output, barely more labour — is consistent with an automation story, but productivity statistics cannot by themselves say how much of it AI caused ⚖ Contested [41] [42].

The reconciliation lies in what employers actually do with the time AI saves. Most do not fire the person whose tasks shrink; they stop backfilling, freeze graduate intake and let attrition do the work ◈ Strong Evidence [15] [19] [44]. Klarna’s customer-service headcount fell from about 5,000 to about 3,000 largely through attrition before the company reversed course [15]. IBM paused hiring for back-office roles in 2023 rather than cutting them [19]. KPMG cut its UK graduate scheme by 29% between 2023 and 2024 [48]. None of these moves appears in unemployment data as a dismissal. Each appears only as a job that did not come into existence — which is why the automation of white-collar work has been so quiet, and why it is so easy to understate.

The task-level evidence explains why the effect is concentrated where it is. Controlled studies show the largest productivity gains go to novices and low performers: 34% for the least experienced customer-support agents [14], 43% for the weakest consultants in a Boston Consulting Group experiment [34], and the biggest gains among junior developers in field trials of GitHub Copilot [8]. AI, in other words, is best at exactly the foundational work — research, first drafts, basic analysis, routine correspondence — that junior employees have traditionally been hired to do and learn from ◈ Strong Evidence [51]. When that work can be done by a tool supervised by a senior employee, the economic case for the junior hire weakens, even as the tool makes the junior hire more productive.

This report is organised around that tension. Sections two to four examine the functions where automation is furthest advanced — software, back-office operations, law and creative work — and what the controlled evidence says about each. Section five sets the productivity data against headcount, and against the history of earlier office technologies. Section six follows the effect to the entry level. Section seven compares how the United States, the United Kingdom, Japan, Germany, India and Denmark are absorbing the change and how regulators are responding. Section eight weighs the strongest arguments on each side of the displacement debate and sets out what the evidence, taken together, supports.

Software development is where generative AI has moved fastest from novelty to default. The DORA 2025 survey of technology professionals found AI adoption among developers at 90%, up from 76% a year earlier, with a median of two hours of use a day ✓ Established [11]. Stack Overflow’s 2025 survey of 49,009 developers found 84% using or planning to use AI tools [12]. At the largest firms the claims are larger still. Satya Nadella said in April 2025 that 20% to 30% of the code in Microsoft’s repositories was written by AI [6]. Sundar Pichai said in April 2026 that about 75% of new code at Google was AI-generated and then reviewed by engineers, against roughly 25% in October 2024 and 50% in the autumn of 2025 [5]. These are executive statements about undefined metrics, produced by companies that sell AI coding tools, and should be read as such ⚖ Contested [5] [6].

The controlled evidence is more modest but still substantial. The first widely cited experiment, in 2023, gave developers a single bounded task — writing an HTTP server in JavaScript — and found that those using GitHub Copilot finished 55.8% faster [7]. A larger and more realistic study, published in Management Science, randomised access to Copilot across 4,867 developers at Microsoft, Accenture and an anonymous Fortune 100 company. It found a 26.08% increase in completed tasks per week, 13.55% more commits and 38.38% more builds ✓ Established [8]. Less experienced developers adopted the tool more readily and gained more from it [8]. Pull requests measure throughput rather than value, but these are the strongest causal estimates available for professional software work.

The strongest counter-evidence comes from METR, a non-profit research group. In a randomised trial published in July 2025, 16 experienced open-source developers worked on 246 real issues in large, mature codebases they knew well. When allowed to use AI tools, they took 19% longer to finish ✓ Established [9]. Before the study they had forecast a 24% speed-up; afterwards they still believed AI had made them 20% faster [9]. A follow-up published in February 2026, using late-2025 tools and 57 developers, pointed to a likely speed-up of 18% for returning participants and 4% for new ones — but METR itself called the result only very weak evidence, because developers increasingly refused to take part without AI and skipped tasks where it helped most [10].

Taken together, the studies point to a consistent pattern: AI accelerates bounded, well-specified work in unfamiliar territory and helps less experienced developers most, while offering less — or even negative — value to experts working in complex systems they already understand ◈ Strong Evidence [8] [9] [10]. Quality is the unresolved question. DORA found that, even as more than 80% of respondents reported higher productivity, AI adoption remained negatively associated with software delivery stability [11]. Stack Overflow found trust falling as use rose: 46% of developers said they distrusted the accuracy of AI output in 2025, up from 31% in 2024, against 33% who trusted it, and 66% named answers that are almost right, but not quite, as their main frustration [12].

What is happening to software employment is more nuanced than the code-share figures suggest. The Stanford AI Index 2026 reports that employment of software developers aged 22 to 25 fell by nearly 20% from 2024 ✓ Established [58]. Yet Indeed data show US software-development postings rising about 15% from February 2025, while total postings fell about 7% — with 71% of the increase in senior roles [13]. Demand is shifting rather than collapsing: employers want experienced engineers who can specify, review and integrate AI-generated code, and they want fewer juniors to write it from scratch ◈ Strong Evidence [13] [3].

This is the first and clearest instance of the pattern that recurs throughout this report. The tasks being automated are the tasks through which junior developers used to learn — boilerplate, tests, small features, migrations. Pichai described one complex code migration, done by agents and engineers together, as finished six times faster than engineers alone could have managed a year earlier [5]. Work of that kind once required a team of junior staff. The senior engineers who now supervise the agents learned their craft doing exactly that work, which raises a question the industry has not yet answered: where the next generation of senior engineers will come from ◈ Strong Evidence [3] [58].

Software also shows how far executive claims can outrun measured reality. Code share is not productivity: a line generated by AI and rewritten by a human counts in the first metric but adds nothing to the second, and none of the companies publishing code-share figures has defined how it measures them ⚖ Contested [5] [6]. The defensible conclusion from the controlled evidence is that AI has raised developer throughput by something like 10% to 30% on typical professional tasks, with large variation and unresolved questions about quality and long-term maintainability ◈ Strong Evidence [8] [10] [11]. That is a large gain for a single technology. It is not, on present evidence, the end of the software profession.

Customer support was the first white-collar function where generative AI was studied at scale in a real workplace, and the results remain among the most cited. Erik Brynjolfsson, Danielle Li and Lindsey Raymond tracked 5,179 support agents at a software firm as an AI assistant was rolled out in stages. Productivity, measured as issues resolved per hour, rose 14% on average and 34% for novice and low-skilled agents, with minimal effect for the most experienced ✓ Established [14]. Customer sentiment improved and staff turnover fell [14]. The tool worked by spreading the practices of the best agents to everyone else. In that study it augmented workers rather than replacing them — but the same capability, deployed directly to customers, becomes a substitute.

That is what Klarna tried. In February 2024 the Swedish payments company announced that its AI assistant was handling about 75% of customer chats in more than 35 languages, doing the equivalent work of 700 full-time agents and cutting resolution times from 11 minutes to two [15]. Its headcount fell from about 5,000 to about 3,000, mostly through attrition and a hiring freeze [15]. By May 2025 the company had changed direction, piloting a new model of remote human agents. Chief executive Sebastian Siemiatkowski told Bloomberg that cost had been too dominant a factor and that the result was lower quality ✓ Established [16].

Salesforce offers the opposite case: a company that cut deeply and says it is satisfied. Marc Benioff said in early September 2025 that the company had reduced its support headcount from about 9,000 to about 5,000 because its Agentforce agents were handling about half of all customer conversations, with a supervisory system handing complex cases to humans ✓ Established [17]. Months earlier he had said that AI was doing 30% to 50% of the work at Salesforce [17]. IBM’s path was slower and more revealing. In May 2023 Arvind Krishna said about 30% of 26,000 non-customer-facing roles — about 7,800 jobs — could be replaced by AI over five years, and paused hiring for them [19]. By May 2025 the confirmed number of HR roles replaced by AI agents was about 200, while IBM’s AskHR system was resolving 94% of routine HR queries and total employment had risen [18].

Insurance and financial operations are following the same path. Allianz Partners, which employs more than 22,000 people, about 14,000 of them in phone enquiries and claims, confirmed in July 2026 that it would cut 1,500 to 1,800 jobs in Europe — 7% to 8% of its staff — citing AI ✓ Established [20]. Munich Re’s Ergo is separately cutting about 1,000 German posts, partly because of automation [20]. JPMorgan told investors in May 2025 that AI would allow operations headcount to fall by 10% over five years; one senior executive said she would take the over on that projection [85] [47]. In the third quarter of 2025 the bank’s profit rose 12% while its headcount rose only 1% [47].

Accounting shows the augmenting side of back-office automation. A Stanford study by Jung Ho Choi and Chuan Xie surveyed 277 accountants and analysed more than 200,000 transactions at 79 small firms. Accountants using AI supported 55% more clients, moved 8.5% of their weekly time to higher-value work and closed the monthly books 7.5 days earlier ◈ Strong Evidence [21]. They still corrected the tools’ errors [21]. The gains here went to output, not to headcount reduction — at least so far. Whether the 55% more clients per accountant eventually means fewer accountants depends on whether demand for accounting services rises to match.

The MIT report that became famous in August 2025 for finding that 95% of corporate generative-AI pilots delivered no measurable profit-and-loss return contained a less-quoted detail that matters here: the best returns it found were in back-office automation, where firms cut spending on outsourced processing and external agencies ⚖ Contested [22]. The methodology of that report has been widely criticised for its small interview sample and narrow definition of success [22]. But the direction of the finding is consistent with other evidence. The money that firms save through back-office automation often comes first from contractors, outsourcers and agencies — workers who do not appear on the firm’s own payroll and whose losses are invisible in its headcount.

The lesson of the back office is that substitution is real but rarely clean. Where AI handles a well-defined, high-volume interaction — a password reset, a refund status, a routine HR policy question — it can replace human contact at scale ◈ Strong Evidence [17] [18]. Where interactions are emotionally loaded, ambiguous or high-stakes, quality falls and customers notice, as Klarna discovered [16]. Companies are converging on a hybrid: automated handling of routine contacts, human escalation for the rest and steady headcount reduction through attrition. For workers, that hybrid means fewer posts at the bottom of the support ladder and a heavier, more difficult case mix for those who remain.

Law was an early adopter. Allen and Overy, which has since merged with the US firm Shearman and Sterling, rolled out Harvey to more than 3,500 lawyers in 43 offices in February 2023, after a beta in which lawyers asked about 40,000 questions [24]. By March 2026 Harvey had more than 1,000 customers, including a majority of the 100 largest US law firms, and about 100,000 lawyers using it, and had raised 200 million dollars at an 11 billion dollar valuation ✓ Established [23]. Thomson Reuters’ 2025 survey of 2,275 legal, tax and accounting professionals found they expected AI to save five hours a week, or about 240 hours a year [28]. The tasks being automated are specific: first-draft memos, contract review, due-diligence summaries, case-law research and document comparison.

The controlled evidence supports large speed gains and more qualified quality gains. The first randomised trial of AI in legal analysis, by Jonathan Choi, Amy Monahan and Daniel Schwarcz, found that GPT-4 produced only slight and inconsistent improvements in quality but large and consistent reductions in time, with the lowest-skilled participants gaining most in quality ✓ Established [26]. A 2025 follow-up tested two newer tools on six realistic legal tasks. Productivity rose 50% to 130% on five of the six; both tools improved quality; the retrieval-based legal tool did not raise hallucination rates, while the general reasoning model produced deeper analysis but introduced fabricated content [25]. Which tool a firm buys, and how its lawyers use it, matters as much as whether it uses AI at all.

The cost of getting it wrong is now measurable. A database maintained by Damien Charlotin of HEC Paris had recorded about 1,600 court decisions worldwide by June 2026 involving hallucinated AI content — fabricated cases, false quotations or invented citations — and was growing by five to six a day ✓ Established [27]. More than 1,000 of those cases were in the United States [27]. Each is a case where time saved in drafting was lost, with interest, in sanctions, reputational damage and judicial rebuke. The Bureau of Labor Statistics now names AI explicitly as a factor limiting demand for paralegals and legal assistants, whose research and document-preparation work overlaps most with what the tools do, and projects their employment to be roughly flat to 2034 [29].

Creative and language work offers the cleanest evidence of substitution, because online freelance platforms record every job and every payment. Xiang Hui, Oren Reshef and Luofeng Zhou found that after ChatGPT’s release, freelance writers on a large platform saw a 2% fall in monthly jobs and a 5.2% fall in monthly earnings; after image generators arrived, image freelancers saw jobs fall 3.7% and earnings 9.4% ✓ Established [30]. Contrary to the common assumption that AI hurts only low-skill workers, top-rated and more experienced freelancers were hit harder [30]. A second study, by Ozge Demirci, Jonas Hannane and Xinrong Zhu, found a 21% fall in posts for automation-prone writing and coding jobs within eight months, with writing jobs down more than 30% and image-creation posts down 17% [31].

Translation and illustration show the same pressure from the workers’ side. A 2024 survey of 787 members of the UK Society of Authors found that 36% of translators and 26% of illustrators had already lost work to generative AI, and that more than 43% of translators reported lower income ◈ Strong Evidence [32]. That is a self-selected member survey, not a random sample, and should be read as a strong signal rather than a precise rate [32]. In April 2025 Duolingo’s chief executive told staff that the company would gradually stop using contractors for work that AI could handle, including content creation and translation, and would add headcount only where teams could show their work could not be automated [33]. After a public backlash he said that AI would not replace the company’s full-time employees [33] — a distinction that, once again, places the first losses outside the payroll.

The BCG study, conducted with researchers from Harvard, Wharton and MIT, remains the most detailed experiment on AI in high-end knowledge work, and its central finding has held up across functions: AI compresses the skill distribution ◈ Strong Evidence [34] [14] [26]. The weakest consultants gained 43%, the least experienced support agents 34% and the lowest-skilled law students most in quality. Ethan Mollick, one of its authors, observed that when the AI is very good, humans have no reason to work hard and pay attention [34]. That observation is the link between the speed gains and the hallucination cases. As the tools improve, the human check that catches their errors becomes both more important and less practised.

For law, writing and translation, the evidence supports three conclusions. First, AI has already reduced demand at the most substitutable edge of these markets — platform freelance work, routine translation, stock imagery and junior research ◈ Strong Evidence [30] [31] [32]. Second, the remaining work is more complex and, according to Demirci and colleagues, better paid, but there is less of it and competition for it is fiercer [31]. Third, in regulated professions the bottleneck has moved from producing documents to verifying them, and the professionals best placed to verify are those trained in the old way, by doing the drafting that AI now does ◈ Strong Evidence [26] [27]. The professions are being hollowed out at the bottom of the pyramid, not at the top.

Every earlier wave of office automation produced the same early argument about whether it would destroy clerical work, and the record is instructive. When LEO I took over the J. Lyons payroll in 1954, it replaced a function performed by rooms of clerks [78]. When spreadsheets arrived, the United States lost about 400,000 bookkeeping and accounting clerks over the next three and a half decades — but gained about 600,000 accountants and auditors, because cheaper analysis created demand for more of it ✓ Established [76]. Automated teller machines grew from about 60,000 in 1985 to 352,000 by 2002, while the number of bank tellers rose slightly, from about 500,000 to 527,000, because each branch needed fewer staff and banks opened more branches [77]. In each case the task disappeared before the occupation did, and the occupation changed shape rather than vanishing.

Generative AI is already producing measurable time savings at the level of individual workers. A survey by economists at the St. Louis Fed found that users of generative AI saved 5.4% of their working hours, implying a saving of about 1.1% of all hours worked across the economy ◈ Strong Evidence [37]. Anthropic’s analysis of millions of conversations with its Claude models estimated that tasks requiring a university degree were sped up about 12 times, and cut its success-adjusted estimate of the potential boost to labour productivity growth to 1.0 to 1.2 percentage points a year [38]. Usage patterns are shifting toward automation. In enterprise traffic through Anthropic’s programming interface, most interactions now delegate whole tasks rather than collaborating on them [38].

The firm-level returns are much thinner. McKinsey’s 2025 State of AI survey found that only 39% of organisations reported any enterprise-level effect on earnings before interest and taxes from AI, and most of those put it below 5% ✓ Established [39]. IBM’s 2025 survey of 2,000 chief executives found that only 25% of AI initiatives had delivered their expected return, and only 16% had been scaled across the enterprise [40]. Census Bureau data show business AI use at 17% to 20% of US firms between December 2025 and May 2026, though 37% of firms with 250 or more employees use it [35]. Adoption depends heavily on how it is measured: a Federal Reserve note found that 18% of firms use AI but 78% of workers are employed at firms that do ◈ Strong Evidence [36].

The Danish result is the most important single piece of evidence on what AI is doing to work, because it links individual use to administrative records rather than self-reports. Its message is not that AI does nothing. It is that the time saved is being absorbed — into new tasks, into AI oversight and integration, into more output at the same wage — rather than converted into either higher pay or fewer hours ◈ Strong Evidence [4]. Firms, not workers, are capturing the gain, and they are capturing it mainly through what they do not spend. That brings the analysis back to headcount. Large employers have spent 2025 and 2026 explicitly managing their white-collar workforces down or flat while output grows:

• Amazon: about 30,000 corporate roles cut in October 2025 and January 2026, about 10% of corporate staff, after a memo saying AI would shrink the corporate workforce [45] [46]
• Salesforce: customer-support headcount reduced from about 9,000 to about 5,000, with AI agents handling about half of conversations [17]
• Block: about 4,000 of more than 10,000 staff cut in February 2026, with AI tools cited directly [81]
• JPMorgan: operations headcount guided down 10% over five years, while third-quarter 2025 profit rose 12% on 1% headcount growth [85] [47]
• KPMG UK: graduate intake cut 29%, from 1,399 to 942, between 2023 and 2024 [48]

Several of these firms say AI is not the main reason. Amazon framed its cuts as removing layers and bureaucracy [46]. Accounting firms cite a slower economy and offshoring alongside AI [48]. The Yale Budget Lab warned in February 2026 that some companies may be AI-washing layoffs that have deeper causes, such as post-pandemic over-hiring and the reallocation of budgets to AI capital spending ⚖ Contested [42]. Andy Challenger, whose firm compiles the layoff data, put the underlying logic plainly in May 2026: regardless of whether individual jobs are being replaced by AI, the money for those roles is [75]. Budgets that once paid for headcount are being redirected to computing capacity, and that reallocation is real whether or not any specific worker was replaced by a specific tool.

The macroeconomic estimates span an order of magnitude, which is itself informative. Daron Acemoglu’s task-based model puts the gain in total factor productivity at no more than 0.66% over ten years [59]. The Penn Wharton Budget Model projects a 1.5% increase in the level of GDP by 2035, with the boost to annual productivity growth peaking at 0.2 percentage points in 2032 [60]. Goldman Sachs, which in 2023 calculated that the equivalent of 300 million full-time jobs worldwide were exposed to automation [56], estimates that 6% to 7% of the US workforce would be displaced if AI were widely adopted, with unemployment temporarily about 0.5 percentage points higher during the transition ⚖ Contested [57]. The spread reflects disagreement about how many tasks AI can perform reliably and how fast firms will reorganise around it — precisely the variables that the task-level studies show are most uncertain.

The most rigorous evidence that AI is already affecting employment comes from payroll data, not surveys. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen used monthly records from ADP, the largest US payroll processor, covering millions of workers. They found that employees aged 22 to 25 in the occupations most exposed to AI — software development and customer service among them — experienced a 13% relative decline in employment after late 2022, after controlling for firm-level shocks ◈ Strong Evidence [3]. Older workers in the same occupations held steady or gained [3]. A revised version of the paper published in August 2026, using data through June 2026, puts the gap at 19% [3].

Two details make the finding more persuasive than a simple correlation. First, the declines are concentrated in occupations where AI is mainly used to automate tasks, while employment in occupations where it is mainly used to augment human work has grown [3]. Second, the adjustment is happening through employment rather than wages — firms are hiring fewer young workers, not paying them less [3]. Brynjolfsson summarised it in August 2025: employment had begun to decline for young workers in highly exposed occupations like coding and call centres, but older workers and those who use AI to augment their work had seen job gains [3]. The authors describe their results as descriptive rather than causal [3].

The pattern is visible well beyond software. The New York Fed’s tracking of recent college graduates aged 22 to 27 shows unemployment at about 5.6% in the second quarter of 2026, against about 4.2% for all workers — an unusually wide inversion — and underemployment at about 42% ✓ Established [50]. The Burning Glass Institute found that postings requiring less than three years’ experience fell steeply after 2022 in AI-exposed fields such as business operations, project management and finance, and that the unemployment advantage of a degree had narrowed to its smallest in 30 years [51]. Its report noted that generative AI excels at precisely the foundational work that junior employees traditionally perform: research, first drafts, basic data analysis and routine communications [51].

In the United Kingdom the evidence is starker in the vacancy data. Adzuna found that entry-level vacancies — graduate roles, apprenticeships, internships and junior positions — fell by 31.89% between November 2022 and mid-2025, and that their share of all postings fell from 28.9% to 25% ◈ Strong Evidence [49]. The four largest accounting firms cut graduate intake sharply: KPMG by 29%, Deloitte by 18%, EY by 11% and PwC by 6%, and Indeed recorded a 44% fall in UK accountancy graduate adverts compared with 2023 [48]. Federal Reserve Governor Michael Barr said in February 2026 that early-career workers in AI-exposed fields were already seeing relative employment declines, and warned that the short run may have long-term consequences for them [79].

The causal case is genuinely contested ⚖ Contested [52] [42]. The Economic Innovation Group argued in January 2026 that the decline in young workers’ employment in AI-exposed occupations is better explained by the sharpest monetary tightening in four decades, which hit interest-rate-sensitive technology and finance hiring hardest [52]. The Burning Glass Institute itself also points to a graduate glut [51]. The Yale Budget Lab found in April 2026 that unemployed and employed workers had the same exposure to AI tasks, which would not be expected if AI were driving people out of work [42]. Post-pandemic over-hiring in technology and the end of cheap capital affected exactly the sectors where AI exposure is highest, which makes the two explanations hard to separate.

The weight of evidence nonetheless supports a narrower claim that survives these objections ◈ Strong Evidence [3] [44] [58]. Within the same occupations and the same firms, young workers are losing ground relative to older ones, and the losses are larger where AI automates tasks than where it augments them [3]. Interest rates do not obviously explain that within-occupation age gap. Employers surveyed by the New York Fed say they expect AI to slow hiring, particularly of college graduates, rather than to drive layoffs [43] [44]. The quiet automation of white-collar work is, at this stage, mostly an automation of entry: the work that used to be handed to new recruits is being handed to software, and the recruits are not being hired.

Global institutions agree on the scale of exposure and disagree on what it means. The International Monetary Fund estimated in January 2024 that almost 40% of global employment, and about 60% in advanced economies, is exposed to AI, and warned that AI would likely worsen inequality in most scenarios [54]. The ILO’s 2025 index put exposure at 25% of global employment, with clerical work the most exposed occupational group, and found a sharp gender gap: in high-income countries, 9.6% of women’s employment falls in the highest-exposure category against 3.5% of men’s ✓ Established [53]. The World Economic Forum’s 2025 employer survey projected 170 million jobs created and 92 million displaced by 2030, with 40% of employers expecting to reduce their workforce where AI can automate tasks [55]. Exposure measures what AI could touch, not what it will replace, and the ILO itself stresses transformation over substitution [53].

The United States is where the displacement debate is loudest and the data richest. Corporate announcements, the Challenger layoff counts and the ADP payroll findings all point toward AI-linked restructuring of white-collar work ◈ Strong Evidence [1] [3]. Yet the Yale Budget Lab finds no economy-wide disruption [42], and the New York Fed finds very few AI-driven layoffs [44]. Federal Reserve officials have begun to treat the issue as a policy question. Governor Christopher Waller said in October 2025 that AI seemed to be moving so fast that job losses might come before the new jobs did [80]. The United Kingdom shows the entry-level effect most clearly in vacancy data, as section six showed [49] [48]. IPPR, a think tank, warned in 2024 that up to 8 million UK jobs could be lost in a worst-case second wave of adoption, with back-office, entry-level and part-time roles most exposed first [61].

Japan frames the same technology in opposite terms. The Recruit Works Institute projects a shortfall of more than 11 million workers by 2040 [66], so AI is presented as a remedy for labour scarcity rather than a threat to jobs. Mizuho Financial Group has said AI will absorb work equivalent to about 5,000 administrative roles over a decade, roughly a third of its clerical staff in Japan, with employees redeployed to sales and customer-facing work rather than dismissed ◈ Strong Evidence [64]. The country’s AI Promotion Act, passed by the Diet on 28 May 2025, is a soft-law framework with no penalties and no employment-specific obligations [65]. Redeployment is real, but over a decade it may still mean fewer clerical hires — the same quiet mechanism operating under a different name [64].

Germany and Denmark provide the most careful administrative evidence. The Institute for Employment Research, the research arm of Germany’s Federal Employment Agency, estimates that 38% of employees work in occupations where more than 70% of tasks are technically automatable, with the largest increase among highly qualified experts and IT services [67]. Its 2025 study found no significant AI-driven employment loss so far [67]. Denmark’s registry data, discussed in section five, show the same null result for earnings and hours [4]. In India the picture is split. Tata Consultancy Services announced in July 2025 that it would cut about 12,000 jobs, about 2% of its 613,000 staff, mainly in middle and senior management [62]. Infosys, by contrast, kept hiring about 20,000 graduates a year, trained in AI before deployment, and Wipro targeted about 10,000 ◈ Strong Evidence [63].

Regulation has so far focused on how AI is used to manage workers rather than on whether it replaces them. The EU AI Act classes AI used for recruitment, promotion, termination, task allocation and monitoring as high-risk, and its AI-literacy duty has applied since 2 February 2025 ✓ Established [68]. But in July 2026 the EU’s Digital Omnibus on AI deferred the high-risk obligations, including those covering employment, from August 2026 to 2 December 2027 [69]. Colorado’s comprehensive AI Act was pushed back from February to June 2026 [71]. New York added a box to its layoff-notice form in 2025 asking employers whether automation or AI was a reason for cuts; in the first year, none of more than 160 filers ticked it for AI [70]. US senators Mark Warner and Josh Hawley have proposed mandatory quarterly reporting of AI-related layoffs, hires and retraining to the Department of Labor [73].

Politically, the debate has polarised faster than the evidence. A report by the Democratic minority staff of the Senate HELP Committee in October 2025 claimed AI and automation could replace nearly 100 million US jobs within ten years, but its estimates were generated partly by prompting ChatGPT and were not peer reviewed ⚖ Contested [72]. At the other end, many technology executives insist that new jobs will outnumber those lost. Dario Amodei of Anthropic warned in May 2025 that AI could eliminate half of entry-level white-collar jobs and push unemployment to 10% to 20% within one to five years [82]. Nvidia’s Jensen Huang said he disagreed with almost everything Amodei said on the subject [83], while Ford’s Jim Farley said AI would replace literally half of all white-collar workers in the United States [84].

The most effective protection so far has come not from legislation but from bargaining. The 2023 Writers Guild of America contract, which ended a strike lasting from 2 May to 24 September, bars AI from writing or rewriting literary material and prevents studios from requiring writers to use it ✓ Established [74]. That model works where workers are organised and their output is easy to define. Most office work meets neither condition. The policy gap that emerges from the country comparison is therefore specific: no major jurisdiction measures the hiring that AI prevents, and almost none of the existing rules would capture it if they did ◈ Strong Evidence [70] [73]. Governments are regulating the visible layoff while the invisible non-hire passes unrecorded.

Each column contains strong evidence, and the disagreement narrows once it is clear what each measures. The aggregate studies — Yale’s occupational-mix analysis, Denmark’s registry data, Germany’s administrative records — ask whether AI has changed total employment, earnings or hours across whole occupations ◈ Strong Evidence [42] [4] [67]. It has not, yet. The payroll and vacancy studies ask whether the flow of new entrants into AI-exposed work has changed ◈ Strong Evidence [3] [49] [50]. It has. A labour market can show stable totals while its entry flows shrink, because incumbents keep their jobs and attrition is slow. The stock of white-collar employment is calm; the inflow is not.

The corporate layoff data deserve particular scepticism, in both directions ⚖ Contested [1] [42] [70]. Challenger counts what companies say in announcements, not what caused the cuts, and the AI share of cited reasons fell from 40% in May 2026 to 9% in September [1]. The same firm reported in August 2026 that announced hiring was up 25% year on year [2]. The New York filing data suggest that companies cite AI in press releases, where it signals efficiency to investors, but not in legal filings, where the attribution carries consequences [70]. Yet the opposite error is also possible. When AI absorbs the work of a job that is never posted, no announcement, filing or unemployment claim records it. The quiet effect is by definition under-counted, and the loud one is probably over-counted.

The function-by-function evidence points to a clear ranking of exposure ◈ Strong Evidence [17] [8] [31] [27]. Customer support and routine back-office processing face the most direct substitution, because the tasks are high-volume, well-defined and easy to monitor. Software development faces large throughput gains and a sharp shift in demand from junior to senior staff. Freelance writing, translation and stock imagery face substitution at the low end of already precarious markets. Law and accounting face the largest speed gains with the most expensive failure modes, which keeps the human reviewer in place for now. Across all of them, the less experienced worker gains most from the tool and is also the most easily replaced by it [14] [34].

That paradox is the central finding of this report. AI is, in controlled studies, an equaliser: it lifts the weakest performers most and narrows skill gaps across support, consulting, law and software ✓ Established [14] [34] [26] [8]. In the labour market, it is functioning as a barrier: it makes the junior hire less necessary at exactly the moment it would make that hire more productive [3] [51]. Which of those effects dominates depends on choices employers make about who uses the tool. If firms give AI to juniors and keep hiring them, the equalising effect wins. If firms give AI to seniors and stop hiring juniors, the barrier effect wins — and the payroll data suggest that, so far, the second choice is more common.

Three things would change this assessment. The first would be clear evidence of displacement in aggregate data — rising unemployment among experienced workers in exposed occupations, not just slower entry. The Yale Budget Lab and New York Fed series are the ones to watch [42] [44]. The second would be evidence that AI-generated productivity is raising wages or creating new occupations at scale, as spreadsheets did for accountants [76]. The Danish data, which show earnings gains in only 3% to 7% of cases, do not yet show that [4]. The third would be a reversal in entry-level hiring as firms discover, as Klarna did in customer service, that cutting too deep degrades quality [16].

For policymakers, the evidence points to a measurement problem before a protection problem. No jurisdiction systematically records the jobs that AI prevents, and voluntary disclosure has produced nothing [70]. Mandatory reporting of AI-linked hiring changes, of the kind proposed by Warner and Hawley, would at least make the quiet effect visible [73]. For employers, the evidence suggests that the cheapest short-term strategy — use AI to avoid junior hires — carries a long-term cost in lost training pipelines that no single firm will bear alone. For workers entering professional careers in 2026, the evidence is blunt: the tasks that used to be how people learned the job are the tasks AI now does best, and the path into white-collar work runs increasingly through the ability to supervise, verify and correct it ◈ Strong Evidence [3] [34] [51].