How AI Is Quietly Changing White-Collar Jobs Before Anyone Is Ready

AI is changing white-collar work mainly by removing the entry-level tasks that once trained junior staff, not by firing experienced professionals en masse.

Payroll data shows employment for workers aged 22 to 25 in the most AI-exposed occupations has fallen about 13 percent relative to peers, while aggregate unemployment has barely moved.

That mismatch explains why the debate feels so unsettled. The loudest forecast came from Dario Amodei, chief executive of Anthropic, who told Axios that AI could wipe out “half of all entry-level white-collar jobs” and spike unemployment to 10 to 20 percent within one to five years.

Nothing close to that has appeared in national statistics. Yet the evidence that something real is happening at the bottom of the career ladder has grown steadily, and it points to a disruption that is structural, uneven, and easy to miss in headline numbers.

What the Payroll Data Shows

The strongest evidence comes from a Stanford Digital Economy Lab paper by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, titled “Canaries in the Coal Mine?”

The authors drew on millions of anonymized ADP payroll records and compared workers by age and by how exposed their occupation is to generative AI. Software developers and customer service representatives sat at the exposed end. Young workers there lost ground while older colleagues in the same jobs held steady.

The most useful finding concerns the type of AI use. In occupations where AI automates work, young workers saw declining employment, while in occupations where AI augments work, employment grew for workers of all ages.

The adjustment also ran through headcount, not pay. Adjustments occur primarily via employment rather than compensation, which means affected workers are not seeing wage cuts. They are simply not being hired.

Erik Brynjolfsson offered a compact explanation of why juniors are exposed first: “It appears what younger workers know overlaps with what LLMs can replace.” Early-career employees tend to hold codified knowledge, the kind learned from textbooks and coursework.

Experienced employees hold tacit knowledge, the judgment accumulated from watching deals collapse and audits fail. Current models are far better at the first kind.

A companion tracker built with ADP Research now follows the pattern monthly. Workers aged 22 to 25 in highly exposed roles show employment contracting at 3.8 percent annually, against 2 percent growth for the same age group in the least exposed occupations.

The Layoff Numbers and the AI-Washing Problem

Corporate announcements tell a louder story. Challenger, Gray & Christmas has tracked AI as a stated layoff reason since 2023, and the totals have climbed sharply.

AI was cited in 38,579 of 97,006 announced U.S. job cuts in one recent month, 40 percent of the total and the highest figure the firm has recorded. Through June, AI had been cited in 101,743 cuts, nearly double the 54,836 attributed to it in all of 2025. Technology employers account for most of them.

Andy Challenger, the firm’s chief revenue officer, has been careful about what the figures mean. “AI isn’t yet the jobpocalypse some predicted,” he said, and he has noted that naming AI in a layoff announcement can win over investors while pushing current and prospective employees away.

That incentive creates a measurement problem. Announced cuts record what a company says, not what caused the decision.

Martha Gimbel, executive director of the Yale Budget Lab, has pointed out that AI offers executives a convenient explanation for downsizing driven by tariffs, immigration policy, pandemic-era overhiring and interest rates.

Analysts call the practice “AI-washing.” A reader looking at layoff trackers should treat them as a record of corporate narrative first and a record of displacement second.

Challenger’s own comment on budgets sharpens the point: “Regardless of whether individual jobs are being replaced by AI, the money for those roles is.” Even when a role is not literally automated, capital is being redirected from payroll to AI spending, and the effect on hiring is the same.

Why Yale and Stanford Can Both Be Right

The skeptical case is serious. Yale’s Budget Lab, working with Brookings, examined the occupational mix of the U.S. workforce and the length of unemployment spells for AI-exposed workers.

Its conclusion was that the broader labor market had not experienced a discernible disruption in the 33 months after ChatGPT’s release.

About 18 percent of workers hold jobs in the highest exposure category, a share that has stayed flat since early 2023. In later reporting, the lab described the picture as one that largely reflects “stability, not major disruption” at an economy-wide level.

The two findings are compatible because they measure different things. Yale looks at the whole economy. Stanford looks at one cohort inside a handful of occupations. A shift that touches a few percent of the workforce, concentrated among 22-to-25-year-olds, can be invisible in national aggregates and still be severe for the people affected.

Anthropic’s own labor research lands in the same middle position. Its authors found little evidence that AI has increased unemployment in the most exposed occupations, but a roughly 14 percent drop in the job-finding rate for workers aged 22 to 25 relative to 2022, a result they described as only marginally statistically significant.

The study also quantified how far deployment lags capability: in computer and math occupations, theoretical AI coverage reaches 94 percent of tasks while observed usage covers only 33 percent. That 61-point gap is the real story. It suggests the current effects reflect early, partial adoption, and the exposure that remains has not yet been realized.

A fair account must also include the strongest counterargument. A New York Fed study attributes much of the rise in young-graduate joblessness to remote work, not AI. It found that unemployment among recent college graduates rose from 3.6 percent in March 2019 to 5.6 percent in March 2026, with remote work driving 64 percent of the increase.

The reasoning is plausible: employers hesitate to onboard inexperienced workers into distributed teams where informal training is harder. The most defensible reading is that several forces are compressing entry-level hiring at once, and AI is one of them whose weight is still being estimated.

Which White-Collar Roles Change First

Exposure follows a recognizable profile: screen-based, language-heavy, repeatable work with outputs that can be checked against documentation.

Computer programmers top Anthropic’s list at 75 percent task coverage, while about 30 percent of workers, such as cooks and mechanics, have no measurable exposure. Bloomberg’s analysis, cited by the World Economic Forum, estimated that AI could replace more than 50 percent of the tasks performed by market research analysts (53 percent) and sales representatives (67 percent).

Two details run against common assumptions. First, the most exposed workers are not the lowest paid. Workers in the most exposed occupations are more likely to be older, female, highly educated and higher-paid than those in roles where AI plays little or no part.

Second, credentials no longer buy the protection they once did. The New York Fed’s outcomes-by-major data shows recent computer engineering graduates at a 7.8 percent unemployment rate in 2024, up from 2.3 percent in 2022. A degree in the field AI is learning fastest is now a riskier bet than it was a few years ago.

The roles feeling pressure first include junior software developers, tier-one customer support, sales development, paralegal and document review work, junior financial analysis, and entry-level content production. In each, the traditional first-year work consisted of exactly the codified, high-volume tasks that models handle well.

The Apprenticeship Problem

The deepest risk is not unemployment. It is the loss of the training function that entry-level work performed.

A first-year analyst who reconciles accounts, a junior associate who reviews contracts, and a support agent who triages tickets are all doing real work, and they are also learning how the business fails.

Remove that work and the productivity gain shows up immediately on this quarter’s budget, while the cost surfaces in five to ten years as a shortage of mid-level people who understand the underlying craft.

This is an inference from the structure of the problem, not an observed outcome, and the evidence so far cannot confirm it. Still, the Stanford authors flagged the same concern, and a cohort that misses its first rung may not simply catch up.

If entry-level AI-exposed work traditionally builds the experience-based skills that protect mid-career workers, and that work is disappearing, the cohort may not catch up as it ages.

The New York Fed’s data on underemployment adds weight: recent graduate unemployment remained elevated at about 5.7 percent in the first quarter of its latest release, with underemployment at 41.5 percent.

Common Mistakes Employers and Workers Make

Treating Exposure as Replacement

Exposure measures the share of tasks a model can assist with, not the probability a job disappears. Jobs are bundles. When automation removes the most routine tasks, the remaining work often becomes more demanding and the headcount requirement changes unevenly.

The occupations hit hardest are those where AI substitutes for the whole bundle, which is why the automation-versus-augmentation distinction matters more than any exposure score.

Assuming the Cost Savings Are Permanent

Klarna remains the instructive case. In 2024 the company said its assistant did the work of 700 customer service agents. Its chief executive, Sebastian Siemiatkowski, later told Bloomberg that cost had been “a too predominant evaluation factor” and that service quality suffered.

The company resumed hiring human agents. The lesson is narrower than the headline suggests: Klarna moved to a hybrid model, with AI handling routine cases and humans handling nuance, complaints and edge cases. Companies that automate entire functions on cost grounds alone risk paying twice.

Believing Seniority Is a Shield

Senior staff are safer today because they hold tacit knowledge and carry accountability. That protection weakens if the pipeline feeding senior roles dries up, and it weakens faster in firms that use AI to flatten management layers. Experience protects only while it is being renewed.

Confusing Tool Fluency With Job Security

Knowing how to prompt a chatbot is becoming a baseline skill, not a differentiator. The durable advantage lies in judgment the model cannot supply: verifying output, owning the consequences of an error, and understanding client or regulatory context.

A Practical Framework: The Three-Question Task Test

Workers and managers can assess any role without relying on a vendor’s exposure score. Break the job into its recurring tasks and put three questions to each.

The first question is whether the task’s output is codified. If a competent stranger could check the result against documentation, a template or a rulebook, a model can probably produce it. Summarizing a filing, drafting a standard response and cleaning a dataset all qualify.

The second is what an error costs and who verifies it. Tasks with cheap errors and easy checks automate quickly. Tasks where a mistake triggers legal, financial or safety consequences keep a human accountable, and that accountability is the part of the job that persists.

The third is whether the task builds judgment for the role above it. This is the question most organizations skip. A task that scores high on automation but also serves as training should be redesigned instead of deleted, for example by having a junior employee review and correct AI output under a senior’s supervision.

Tasks scoring high on the first question and low on the second are the ones to expect AI to absorb. Tasks scoring high on the third deserve deliberate protection.

What Employers and Workers Should Do

For employers, the practical step is to redesign entry-level roles instead of simply freezing them. That means pairing junior hires with AI tools and explicit review responsibilities, tracking the internal promotion pipeline as a metric next to cost per hire, and budgeting for reskilling before restructuring.

The World Economic Forum’s Future of Jobs Report found that 40 percent of employers expect to reduce their workforce where AI can automate tasks, yet the same forecast anticipates large numbers of new roles.

The transition cost lands on whoever is mid-transition, and firms that fund upskilling programs and outplacement support retain institutional knowledge that competitors lose.

For workers, the priorities are to move toward the augmentation side of the divide, to build verification and domain expertise that makes AI output trustworthy, and to seek roles with direct client or stakeholder accountability.

Employer-sponsored AI skills certification and structured reskilling courses carry the most value when tied to a specific workflow, not a generic credential. Worker anxiety is already widespread: Mercer’s Global Talent Trends report found 40 percent of employees fear losing their jobs to AI, up from 28 percent in 2024. The more useful response is a task-level audit of one’s own role, not a general retreat.

What to Watch Next

Four data sources will settle most of the open questions. Challenger’s monthly report shows how companies describe their own cuts. The New York Fed posts recent-graduate unemployment and underemployment data quarterly.

The Stanford and ADP tracker shows whether the young-worker gap is widening or closing. And Yale’s Budget Lab has committed to regular updates on occupational mix, the measure that would register a true economy-wide shift.

Martha Gimbel has argued that the real test comes in a downturn, when firms facing pressure on margins have the strongest incentive to adopt automation at scale. A softening economy would make the current hiring pattern much harder to attribute to anything but AI.

Until then, the evidence supports a precise claim and not a dramatic one: the technology has not yet reduced the number of white-collar jobs, but it is already narrowing the way into them.

What People Ask