The impact of AI on white-collar jobs: what the evidence shows in 2026

Exposure, experiments and job data answer different questions

One headline says AI threatens 40% of jobs, the next says there is no sign of AI job losses. Both can be accurate, because they measure different things.

The US Bureau of Labor Statistics states the first distinction plainly on its AI exposure categories page: "Exposure does not imply job loss, productivity gains, automation probability, or wage effects." A highly exposed occupation might shrink, grow, or simply change what its people do all day. Exposure scores also mix automation (AI doing a task instead of a person) with augmentation (AI helping a person do it better), and the BLS notes that its categories do not tell the two apart.

What AI exposure studies measure

Exposure studies break each occupation into its tasks, ask whether a large language model (LLM) could do each one or cut the time it takes without losing quality, and add the tasks up into a score per occupation: its occupational exposure.

One exposure measure that both the BLS and the Yale Budget Lab use comes from GPTs are GPTs by Tyna Eloundou, Sam Manning, Pamela Mishkin and Daniel Rock, published in Science in June 2024. Human experts and GPT-4 rated occupation-specific tasks from the US O*NET database on whether an LLM could cut the time needed by at least half. In the 2023 working paper, the authors estimated that around 80% of US workers could have at least 10% of their tasks affected, and about 19% could see at least half of their tasks affected. Higher-income jobs were more exposed, not less. An LLM on its own could do about 15% of all US worker tasks significantly faster at the same quality; with software built on top of LLMs, the share rises to between 47% and 56%. Most of the potential sits in tools wired into the work, not in a chat window.

The International Monetary Fund applied the idea worldwide in a January 2024 staff discussion note. Its headline figure, in Kristalina Georgieva's summary, is that almost 40% of global employment is exposed to AI: about 60% of jobs in advanced economies, 40% in emerging markets and 26% in low-income countries. In advanced economies, roughly half of the exposed jobs may benefit through higher productivity; in the other half, AI may take over key tasks, which could lower wages and hiring. The note also finds women and college-educated workers more exposed, and older workers potentially less able to adapt.

The International Labour Organization's refined global index, published on May 20, 2025, rebuilt the ILO's 2023 scores with a survey of 1,640 workers and expert review. It finds one worker in four worldwide in an occupation with some exposure to generative AI, but only 3.3% of global employment in the highest-exposure group. Exposure rises with income, from 11% of employment in low-income countries to 34% in high-income ones, and it is higher for women: 4.7% of women's employment sits in the highest group, against 2.4% of men's. The update also lowered some 2023 scores: tasks such as taking meeting notes or scheduling appointments turned out to still need substantial human effort. Because most occupations are made of tasks that still need human input, the ILO concludes that transformation of jobs, not replacement, is the likeliest outcome.

Which white-collar jobs are most exposed to AI

Clerical work tops the indexes. Of the 13 occupations in the ILO's highest-exposure group, most are clerical, and the 2025 update pushed some digitized professional jobs higher too:

Official projections turn exposure into forecasts. The BLS Employment Projections for 2025 to 2035, released on August 27, 2026, expect office and administrative support to be the fastest-declining major occupational group, down 4.0% or 752,100 jobs, the largest loss of any group, as automation tools, including AI-powered ones, spread through workflows. The BLS also expects generative AI to limit demand for some jobs in arts, design, entertainment, sports and media, and e-commerce plus AI tools in the sales process to keep cutting sales jobs, down 1.4%. At the other end, data scientists (up 34.6%) and computer and information research scientists (up 21.8%) are among the ten fastest-growing occupations.

Earlier BLS case studies, in the Monthly Labor Review of February 2025, show the reasoning occupation by occupation for the 2023 to 2033 projections. AI is expected to hit hardest where generative AI can most easily replicate an occupation's core tasks, such as customer service representatives (down 5.0%) and medical transcriptionists (down 4.7%). Credit analysts are projected to decline 3.9%, and paralegals and legal assistants to see lower demand. Lawyers are still projected to grow 5.2%: they have to review what an LLM drafts, clients may prefer a person for advice, and demand for legal services stays strong.

Exposure is also not the same as use. The Yale Budget Lab points out that software developers, in the top fifth of exposure, adopted generative AI very quickly, while adoption in clerical work lagged well behind despite a similar level of exposure.

To see where your own occupation sits, the BLS publishes a downloadable table that puts every occupation it projects in one of four relative categories (low, moderate, high and very high AI exposure), and the ILO page has an interactive chart by occupation. Read both as relative signals: the BLS says a "very high" category does not necessarily mean employment will decline, and a "low" one does not mean the occupation is safe from change.

What controlled experiments found

Experiments measure what happens when real workers get AI for real tasks, against a comparison group that does not. Three are the most useful for office work.

Customer support: 14% more issues resolved per hour

Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the staggered rollout of a generative AI assistant to 5,179 customer support agents (NBER working paper, 2023, published in the Quarterly Journal of Economics in 2025). Productivity, measured as issues resolved per hour, rose 14% on average. Novice and low-skilled agents improved by 34%, while experienced and highly skilled agents saw minimal change. The authors' reading is that the assistant spread the practices of the best agents to newer ones, helping them move down the experience curve. Customer sentiment and employee retention improved too.

Professional writing: 40% less time, 18% better quality

Shakked Noy and Whitney Zhang gave 453 college-educated professionals incentivized writing tasks drawn from their own occupations, and gave half of them ChatGPT (Science, July 2023). The average time taken fell by 40% and output quality rose by 18%. Inequality between workers decreased. Those who had used ChatGPT in the experiment were twice as likely to report using it in their real job two weeks later.

Consultants and the jagged frontier

The experiment that matters most for managers ran with 758 consultants at Boston Consulting Group, about 7% of the firm's individual contributors (Harvard Business School AI Institute). On tasks inside what the researchers called the frontier of AI capabilities, covering creativity, analysis, writing and persuasion, consultants using GPT-4 completed over 12% more tasks, worked over 25% faster and produced work rated over 40% higher in quality.

On a business problem designed to fall just outside that frontier, the result reversed. According to MIT Sloan's account of the study, performance for consultants using AI dropped by an average of 19 percentage points: the model's analysis was wrong, and many followed its recommendation. The researchers also found that it was not obvious to these highly skilled workers which of their everyday tasks AI could handle. That is the jagged technological frontier: AI excels at some tasks and falls short at others that look just as hard, even inside one workflow.

Across the three, the gains are largest for the least experienced: 34% for novice support agents, and 43% against 17% for the lower and upper halves of the consultants by skill. A productivity gain is still not a job count. As the Stanford Digital Economy Lab researchers note, higher productivity can mean less or more employment, depending on how much more of the output people want once it gets cheaper; the BLS makes the same argument for software developers.

What the labour market data shows so far

The Yale Budget Lab first asked the broad question in October 2025: has the mix of occupations in the US workforce changed faster since ChatGPT's release in November 2022 than during earlier technology shifts? After 33 months it found no discernible disruption. The occupational mix was changing only about 1 percentage point faster than it did after the internet spread from 1996, a trend that had started before generative AI, and the shares of workers in low, medium and high exposure groups held steady at about 29%, 46% and 18%.

The lab now runs a monthly AI labour market tracker. Its September 15, 2026 update, using August 2026 Current Population Survey data, still finds no clear evidence of AI-related disruption: occupational churn, AI exposure among the unemployed and usage measures are flat, within historical ranges or on pre-AI trends. A May 2026 analysis compared AI-exposed occupations with a synthetic comparison group built from unexposed ones and found no statistically or economically significant effect on their employment or wages. Unemployment in exposed occupations had risen roughly half a percentage point more than in the comparison group, a difference that was not statistically significant as of early 2026. The authors warn that the survey is too small to measure narrow groups, such as recent graduates, reliably.

Canada's data tell a similar story. Statistics Canada found that from November 2022 to December 2025, employment generally grew whether or not occupations were exposed to AI. Younger and less-educated workers saw weaker job growth, and in coding-intensive jobs such as software engineers and web designers, the gains went to workers aged 30 to 49 while the number of coders under 30 stagnated. The share of Canadian businesses using AI to produce goods or services doubled from 6% to 12% between the 2023 to 2024 and 2024 to 2025 periods, while the share of those businesses that cut staff because of AI stayed at about 6%. Statistics Canada is careful to say it cannot tell whether recent trends reflect AI, post-pandemic adjustment, rapid demographic shifts or trade tensions with the United States.

Why early-career workers are hit first

The clearest signal comes from payroll records. "Canaries in the Coal Mine?", by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab, uses ADP payroll data covering millions of US workers through June 2026. Its August 2026 revision finds no evidence of widespread job displacement. But employment of workers aged 22 to 25 in AI-exposed occupations stands 19% below where it would be had it kept pace with that of their less-exposed peers, and experienced workers in the same occupations show no comparable gap. The gap has widened since the authors first reported it in August 2025. It comes mainly from fewer young people being hired, not more being let go, and it is concentrated in occupations where AI use mostly substitutes for human tasks. Where AI mostly complements workers, employment is flat or rising. Base pay has not adjusted; headcount has.

The authors' proposed mechanism is that AI substitutes for codified knowledge, the formal, documented kind that education and written procedures teach, and complements tacit knowledge that comes from practice and experience. Occupations that rely heavily on codified knowledge show slower entry-level employment growth; those that rely on tacit knowledge show faster growth for mid-career and senior workers. The authors call their results early descriptive indicators, not causal estimates: the pattern weakens when they control for education, shows some divergence before generative AI arrived, and is stronger in the ADP sample than in national surveys.

The Yale and Stanford findings are compatible. Economy-wide employment has not moved, while a narrow group, young people entering exposed occupations, has. The Yale authors themselves note that if AI's effects are limited to a narrow slice of the workforce, other datasets are better suited to find them.

Tasks, not jobs: how to judge your own exposure

Exposure indexes and projections describe occupations. Your risk depends on how your own week is spent. Work through it the way the researchers do:

1. List your tasks for a typical week, with rough hours for each. This is the unit every exposure study uses.
2. Mark each task by what AI does with it today: does it well, speeds you up, or cannot do. Test on your real work rather than guessing; the consultant study showed that skilled workers could not tell in advance which tasks AI would get right.
3. Look up your occupation in the BLS exposure table or the ILO index, as a relative signal of where your field is heading.
4. Move your time toward the tasks that stay human: reviewing AI output, decisions you are accountable for, knowledge built from experience, and work with clients and colleagues.
5. Upskill in the tools of your own field, and learn how they fail. The person who uses AI well and catches its mistakes is in a better position than the one who avoids it.

What employers should do about AI and office work

The evidence points to redesigning work task by task, not to cutting roles on the strength of an exposure score.

1. Map the work by task before changing headcount. Measure a baseline for time, quality and error rates, then compare after AI is introduced. The experiments above found gains on specific tasks, not across whole jobs.
2. Find your own frontier. Test models on real cases with known answers, including the awkward ones. The researchers behind the consultant study recommend an onboarding phase in which staff learn where AI works well and where it does not.
3. Keep a human in the loop wherever a mistake is costly. Decide who reviews and signs off AI output, and make that person accountable.
4. Retrain the people you have. In Statistics Canada's survey of the second quarter of 2026, 44.4% of businesses using AI had changed their training or staffing practices because of it, and 32.0% had given existing employees AI-related training. Among businesses with 100 or more employees that used AI, 68.1% had trained existing staff. Reskilling matters most for clerical and administrative staff, whose occupational group the BLS expects to decline fastest.
5. Keep hiring and training juniors. If AI now does the codified tasks junior staff used to learn on, redesign entry-level roles instead of eliminating them, or the pipeline to your future senior staff dries up. The support study suggests AI can shorten the learning curve for new workers, which is an argument for hiring them with AI, not replacing them with it.
6. Do not let AI make decisions about people on its own. Hiring, promotion and credit decisions need human review and a check for bias; see fairness in AI decision-making.

If you are working out where AI fits in one of your own workflows, our AI and automation service starts with one real workflow, reviewed with the people who run it, before anything is built. It uses a language model only where a step needs judgment, such as reading a document or drafting a reply, and ordinary code everywhere else, and nothing changes in your systems until a rule or a person approves it. The review ends with a fixed quote, or a written recommendation not to build.

Where human judgment stays required

Across the studies, the same kinds of work keep a person in charge:

• Tasks outside the frontier, where consultants who followed the model did worse than those who worked alone.
• Sign-offs that law or professional rules assign to a person. The BLS notes that lawyers must review LLM drafts because AI output may contain errors or biases, and that regulations still require professional engineers to approve work done with emerging technologies.
• Knowledge built through experience, where Stanford's data show employment growing for experienced workers.
• Advice and relationships, where clients may still prefer a person, one reason the BLS expects lawyers to be less affected than paralegals.
• Decisions about people, where accountability cannot be delegated to a model.

How to read the next headline about AI and jobs

A few questions sort most claims:

• Exposure, experiment or employment data? A share of jobs "exposed to AI" is a measure of potential. Only employment data can show jobs lost.
• Causal or descriptive? Most labour market findings so far, including Stanford's, are descriptive. They show a pattern, not proof that AI caused it.
• Which workers? An aggregate figure can hide a narrow group, such as new graduates, and a narrow group's figure can overstate the whole.
• Layoffs or hiring? In June 2026 the Yale Budget Lab examined a sharp rise in Information-sector layoffs: hiring was rising too, and unemployment insurance claims and survey data showed no distress in the sector. One data series is not a trend.
• Whose data? Usage data from one AI vendor reflects that vendor's users: Yale found one assistant's conversations dominated by computer and mathematical occupations.

The honest summary, as of September 2026: AI is reshaping office work task by task, entry-level hiring in exposed occupations is the first measurable casualty, and the wider job losses that exposure studies make people fear have not shown up in the statistics. The trackers above are updated regularly; check their latest release before relying on any single number.