AI job exposure by occupation in 2026: full rankings across 800 US roles

Computer programmers: 74.5% observed AI task coverage. Customer service representatives: 70.1%. On March 5, 2026, Anthropic published the first AI job exposure rankings built from real usage data across 800 US occupations, and the picture looks different from what three years of headlines predicted.

The research introduces a measure called "Observed Exposure": the share of job tasks AI is actually performing in professional workflows, not what it could theoretically handle. It is the first time a major AI lab has combined its own real-world usage data with government occupational databases to produce an exposure estimate grounded in what's happening, not what's possible.

This article covers the full findings: the top-ten list, the complete breakdown by occupational category, the demographic data, what the BLS employment projections say, and what's changed in the April 2026 data updates. It also covers the one signal that should get the most attention from workers under 30.

Why previous AI exposure estimates have been off

Most existing AI exposure estimates deserve skepticism, and understanding where they fall short is the only way to read the new data correctly.

The most widely cited academic framework, from Tyna Eloundou (OpenAI), evaluated tasks across US occupations using a scoring system: 1 if an LLM could speed the task up by at least 2x on its own, 0.5 if it required additional software built on top of the model, and 0 if neither applied. By that measure, 94% of tasks in Computer and Math occupations were marked as theoretically AI-feasible.

Anthropic's observed coverage for the same category: 33-36%, depending on the measurement period.

That 58-60 percentage point gap is not an error in the older framework. It measures something real: the distance between what's technically possible and what's actually happening in professional workflows. The gap exists because of model limitations for specific tasks, legal and compliance hurdles, software integration requirements, human verification steps, and adoption lag.

Eloundou's framework marked "Authorize drug refills and provide prescription information to pharmacies" as fully exposed. Anthropic has not observed Claude performing that task in practice, even though the theoretical assessment is correct. This pattern: the capability-deployment gap. This pattern applies across many high-exposure occupations and is why theoretical estimates systematically overstate near-term risk.

The historical track record gives further reason for caution. A prominent effort in the 2000s identified roughly a quarter of US jobs as vulnerable to offshoring. A decade later, most of those jobs had maintained healthy employment growth. Technology does disrupt labor markets. The track record of forecasts is poor not because the direction is usually wrong, but because the timing and mechanism almost always are.

How Anthropic built the "Observed Exposure" metric

Researchers Maxim Massenkoff and Peter McCrory combined three datasets:

The O*NET database enumerates specific tasks for roughly 800 US occupations, with approximate time weights for each task, how much of the job does this represent?

Eloundou's beta scores flag which tasks are theoretically LLM-feasible: 1 if directly feasible by an LLM alone, 0.5 if it requires an LLM plus additional software tools, and 0 otherwise.

Anthropic's Economic Index tracks actual Claude usage from professional contexts, what share of work-related Claude interactions map onto each O*NET task. This is the ingredient that distinguishes the paper from every prior framework.

The final score combines these inputs. A task is counted as "covered" if it's theoretically feasible (beta > 0) and appears in work-related Claude usage with sufficient frequency. Fully automated implementations receive full weight. Augmentative use, where Claude assists but a human does the substantive work, receives half weight. Task-level scores are averaged to the occupation level, weighted by time spent.

The result is an observed exposure score between 0 and 1 for each of the 800 occupations. The data is publicly available on HuggingFace at the Anthropic Economic Index dataset.

One finding worth noting: 97% of tasks observed in Anthropic's usage data fall into the categories rated as theoretically feasible by Eloundou. Users are, in practice, using AI for tasks it was designed to handle. The gap is in deployment breadth, not in whether the capability exists.

The 10 jobs with the highest AI exposure right now

Source: Massenkoff & McCrory, "Labor market impacts of AI: A new measure and early evidence," Anthropic, March 5, 2026.

A few notes on what's not on this list. Software developers, distinct from computer programmers in the BLS classification, show high theoretical exposure but don't crack the top ten. This likely reflects the more complex, collaborative, and judgment-intensive nature of software development compared to programming tasks that follow well-defined specifications. Lawyers and accountants appear in middle-range exposure territory, consistent with the pattern that professional services requiring contextual judgment diffuse more slowly than information processing and data tasks.

At the other end of the distribution, 30% of US workers score zero observed exposure: cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers. For these workers, the current conversation about AI displacement is largely abstract.

AI exposure by occupational category: theoretical vs observed

The radar chart in the original paper shows both theoretical capability and observed usage across the eight major occupational categories. This is the data the chart represents, in table form:

Exact figures from Anthropic Economic Index (March 2026 Learning Curves report) for confirmed columns. Approximate figures (~) derived from Anthropic paper Figure 2 radar chart. "pp" = percentage points.

The core insight is the same across every category: observed deployment is substantially lower than theoretical capability, and the gap ranges from roughly 48 to 74 percentage points. No occupational category is close to deploying AI at the level that theoretical models predict is possible.

The Computer and Mathematical category is where actual deployment is highest in absolute terms (35.8%), yet even here, only about one-third of theoretically feasible tasks are actually being handled by AI in professional workflows. Legal occupations show a particularly wide gap: nearly all legal tasks are theoretically LLM-feasible, but observed professional usage remains low, reflecting verification requirements, liability constraints, and compliance workflows.

Who is actually most exposed: the demographics

This is where the data gets counterintuitive.

Most early AI risk discourse focused on administrative, clerical, and lower-wage work. The Anthropic data tells a different story about where AI is actually being deployed.

Using Current Population Survey data from August to October 2022, the three months before ChatGPT's release, establishing a clean baseline, the researchers compared workers in the top quartile of observed exposure to the 30% with zero exposure.

The most exposed group earns 47% more on average. Their median hourly wage is $32.69, versus $22.23 for the zero-exposure group. They are 16 percentage points more likely to be female: 54.4% versus 38.8%. They are nearly twice as likely to hold a graduate degree: 17.4% versus 4.5%. They are almost twice as likely to be Asian. And they are 10 percentage points more likely to be married.

This demographic profile matters for two reasons. First, it challenges the assumption that AI displacement disproportionately threatens economically precarious workers. The jobs where AI is actually being deployed most heavily are better-paid, more educated, more knowledge-intensive positions. Second, it shifts the risk calculus. Higher-paid workers have more resources to adapt, but they also have further to fall. A financial analyst who sees meaningful portions of their billable work automated faces a different negotiation dynamic than a data entry specialist whose tasks are automated wholesale.

The 47% wage premium in high-exposure roles means knowing your current market rate matters more in this environment, not less. PayScope benchmarks your specific profile against active job postings, not surveys, so you can see where you actually stand. For a detailed look at how AI exposure connects to salary trajectories, see our analysis of AI job exposure and what it means for your compensation.

The gender pattern is also worth watching. If AI is concentrating deployment in roles that skew female at higher pay grades, market research, medical records, financial analysis, the downstream effects on compensation equity deserve close attention in future updates.

What the BLS employment projections show

The Bureau of Labor Statistics publishes employment projections through 2034. When Massenkoff and McCrory compared BLS projections against their observed exposure measure, a consistent pattern emerged.

For every 10 percentage point increase in observed coverage, BLS's growth projection for that occupation drops by 0.6 percentage points. The relationship is modest, with an R-squared of 0.027, but statistically meaningful at the occupation level, weighted by employment size.

The same relationship does not hold for the Eloundou theoretical exposure measure alone. Only the observed exposure metric, which incorporates actual usage data, correlates with independent BLS projections. This is the paper's clearest validation that measuring real-world deployment, not theoretical capability, produces forecasts that align with what labor market analysts are independently projecting.

Customer service representatives are projected to decline by roughly 5% through 2034. Cashiers show a steeper projected decline. Software developers, despite substantial theoretical AI exposure, are projected to grow around 11%, productivity gains appear to be expanding total demand, not redistributing it. Lawyers project modest positive growth of around 5%. Electricians, at minimal AI exposure, project around 11% growth.

The occupations with the highest observed exposure are not uniformly contracting. But they are, on average, expected to grow less over the next decade than their lower-exposure counterparts.

What the unemployment numbers actually show

No systematic increase in unemployment has appeared in the most AI-exposed occupations since ChatGPT launched in late 2022.

The researchers tracked unemployment rates for workers in the top quartile of observed exposure and compared them to workers with zero exposure, using a difference-in-differences framework from 2016 through 2025. The gap between these two groups' unemployment rates has remained flat. The pooled post-ChatGPT estimate is +0.0020 with a standard error of 0.0019, statistically indistinguishable from zero.

To understand what this framework can detect: the authors estimate that a 1 percentage point differential increase in unemployment would be visible in their data. A scenario where the top 10% most-covered workers were all laid off simultaneously would push aggregate unemployment from 4% to 13%, and would obviously show up. A "Great Recession for white-collar workers" scenario, where unemployment in the top exposure quartile doubled from 3% to 6%, would also be detectable. Neither has materialized.

This doesn't mean nothing is happening. It means the effects, so far, are either too small to detect at the unemployment level or are appearing through different channels, like hiring slowdowns and entry-level compression.

What's changed since March 2026

Updated April 9, 2026

Since the original paper was published, three developments have added important context.

Goldman Sachs, April 2026: 16,000 US jobs per month

Goldman Sachs published analysis in April 2026 estimating that AI is now cutting approximately 16,000 US jobs per month in the United States, with Gen Z workers taking a disproportionate share of that impact. The mechanism aligns with what Massenkoff and McCrory found: the displacement is not showing up as mass layoffs of existing workers, but as a collapse in entry-level hiring in high-exposure fields. Where the Anthropic paper found a 14% drop in job-finding rates for workers aged 22-25, Goldman Sachs is now attaching a specific monthly number to what that compression looks like in aggregate.

For a full breakdown of what the April 2026 displacement data means for workers entering AI-exposed fields, see our analysis: How many jobs is AI replacing in 2026?

Anthropic Learning Curves report, March 2026

Anthropic's March 2026 Economic Index report, released after the original paper, adds February 2026 usage data. Computer and Mathematical occupations now account for 35% of all professional conversations with Claude, up from earlier periods. Experienced Claude users (6+ months of use) achieve a 10% higher success rate in their interactions than newcomers, and that gap is widening.

Two workflow categories showed at least 2x growth between November 2025 and February 2026: business sales and outreach automation and automated trading and market operations. Both involve directive, low-human-oversight workflows, exactly the kind of use that registers as "automated" rather than "augmentative" in the exposure scoring framework.

Fortune interview with Peter McCrory, April 7, 2026

McCrory, one of the paper's co-authors, gave a follow-up interview to Fortune in early April confirming that the paper's authors continue to see no mass unemployment signal in the data, while noting that the entry-level hiring compression continues to develop. His view: the "Great Recession for white-collar workers" scenario remains possible in theory, but is not what the current data shows.

The young worker signal: the one number that should get your attention

The unemployment findings are reassuring. The hiring data tells a more complicated story for one specific group.

Brynjolfsson, Chandar, and Chen (2025) found a 6-16% fall in employment specifically among workers aged 22 to 25 in AI-exposed occupations, attributed primarily to a slowdown in new hiring rather than increased separations. People weren't being laid off. They just weren't being hired.

The Anthropic team ran a parallel analysis tracking the monthly rate at which young workers (22-25) start new jobs in high-exposure versus low-exposure occupations. Since ChatGPT's release, the job-finding rate for young workers entering exposed occupations has fallen by roughly half a percentage point per month. The averaged post-ChatGPT estimate: a 14% drop in job-finding rates for this age group versus 2022 baseline levels. The same effect is not visible for workers over 25.

The mechanism: young workers entering AI-exposed fields often start by handling the tasks now most amenable to automation, data processing, customer inquiry handling, code generation for well-defined specifications, document drafting. If employers are using AI tools to cover those entry-level task loads, they have less need to hire the 23-year-old to do them.

This shows up not as unemployment; many of these young workers are staying in school, taking different jobs, or remaining at existing employers, but as fewer on-ramps into the field. The starting positions that earlier generations used to build skills in finance, programming, and market research are, tentatively, contracting.

For a detailed look at the April 2026 displacement data and what entry-level workers should do with this information, see How many jobs is AI replacing in 2026?

What this means if your job is on the list

A 74% observed exposure score for computer programmers doesn't mean 74% of programmer jobs will disappear this year. It means AI tools are measurably touching nearly three-quarters of the tasks in that role in professional settings today. Some of that is augmentative, faster code completion, better test generation, AI-assisted debugging. Some of it is automated, code generation pipelines that complete entire implementation tasks with minimal human direction.

The distinction matters because augmentative and automated use have different implications for compensation. When AI makes a programmer 40% faster at routine tasks, the value of that programmer's judgment, architectural thinking, and domain knowledge increases. When AI automates the routine tasks entirely, the supply dynamics for "routine programming work" shift, and compensation for that tier of the role faces pressure.

For workers in high-exposure occupations, the practical question is not whether their job disappears but whether the skills they're selling today are the ones that hold or gain value as the automated layer expands. Judgment, domain expertise, client-facing interpretation, and the ability to catch what the automated layer gets wrong are the durable components.

The question of how AI exposure is actually affecting salaries for specific roles, including whether high-exposure workers are seeing wage premiums for AI skills or wage compression from automation, is covered in detail in our analysis of how AI job exposure affects your compensation.

Knowing which of your skills sit above or below the market median, specifically which are commanding a premium versus being increasingly commoditized: this is a different kind of analysis than a job title salary lookup. Market value data now needs to account for which skills within a role are seeing demand growth and which are being absorbed into automated workflows. That specificity is what makes it possible to use data to negotiate a better salary rather than anchoring on a job title average that no longer tells the full story.

Frequently asked questions

Does high AI exposure mean I'm likely to lose my job? Based on current data, no, not as a near-term outcome. The Anthropic study found no systematic increase in unemployment for workers in the most exposed occupations since ChatGPT launched in late 2022. The effects appearing so far are slower employment growth in exposed fields and reduced entry-level hiring for workers aged 22 to 25. Mass displacement of existing workers in these roles has not materialized in the data.

Which jobs have the highest AI exposure right now? Computer programmers top the list at 74.5% observed task coverage, followed by customer service representatives at 70.1% and data entry keyers at 67.1%. The full top ten includes medical record specialists (66.7%), market research analysts (64.8%), wholesale and manufacturing sales representatives (62.8%), financial and investment analysts (57.2%), software QA analysts (51.9%), information security analysts (48.6%), and computer user support specialists (46.8%). These figures come from Anthropic's March 2026 study using real Claude usage data combined with BLS O*NET occupational task data.

Are lower-wage workers most at risk from AI? Not according to this measure. Workers in the most exposed occupations earn 47% more on average ($32.69/hr versus $22.23/hr) and are substantially more likely to hold graduate degrees (17.4% versus 4.5%) compared to the least exposed group. The jobs seeing the heaviest actual AI deployment tend to be knowledge-work and white-collar roles, not the lower-wage service positions that earlier frameworks flagged as most vulnerable.

What makes this exposure measure different from older ones? Most prior frameworks measured theoretical capability: could an LLM speed up this task at all? Anthropic's measure looks at actual usage, whether AI tools are being deployed in professional settings for specific tasks, weighted toward automated rather than augmentative use. The result is consistently much lower than theoretical estimates. Computer and Math occupations, for example, show 94.3% theoretical capability exposure but only 35.8% actual task coverage in practice.

What is the "theoretical capability and observed exposure" chart? The chart referenced in the paper (Figure 2) is a radar chart with eight axes, one per major occupational category. Each axis shows two data points: theoretical capability (how much of the category's work an LLM could theoretically handle) and observed exposure (how much is actually happening in professional Claude usage). The gap between the two lines on each axis represents the capability-deployment gap. The table in this article renders that same data in tabular form for each category.

What's the most important signal to watch right now? The hiring slowdown for workers aged 22-25 in high-exposure occupations is the clearest early signal. A 14% drop in job-finding rates for young workers entering exposed fields since ChatGPT's launch, with no equivalent effect for workers over 25, suggests employers may be handling entry-level task loads differently. Goldman Sachs puts the aggregate monthly job-cut number at approximately 16,000 as of April 2026, concentrated in these same high-exposure fields. This doesn't show up as mass unemployment, but as fewer starting positions.

How is AI exposure connected to salary? Workers in the most exposed occupations currently earn 47% more than the least exposed group, counterintuitively, high AI exposure is correlated with higher pay, not lower. But this reflects where in the labor market AI is actually being deployed, not a protective effect of AI exposure itself. The emerging pattern is wage bifurcation: workers who develop AI skills in these roles are commanding premiums, while workers performing the specific tasks that AI is automating face compensation pressure. For a full analysis, see our breakdown of how AI job exposure affects your compensation.

What should workers in high-exposure fields do with this information? Knowing your occupation's exposure score is the starting point. The next question is which of your specific skills are below, at, or above market median for your role, and which are positioned in the parts of the job that AI is automating most quickly versus least quickly. PayScope benchmarks your specific profile against live job postings so you can see where your market value sits today. Salary negotiation and career decisions both work better with that level of specificity than with job-title averages.

What jobs have zero AI exposure? Roughly 30% of US workers show zero observed AI exposure in Anthropic's data. This includes cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers, and similar roles where AI tools have not shown up in professional workflows in any measurable way. These are predominantly roles requiring physical presence, manual dexterity, or real-time human interaction in unpredictable environments.

Will AI replace my specific job? The data doesn't support "replace" as the near-term framing for most occupations. The more accurate framing is: AI tools are handling an increasing share of specific tasks within your role, which changes the skill premium within that role and the volume of new hiring into entry-level positions. Whether that adds up to job elimination depends on the specific occupation, the pace of adoption, and how employers respond, which varies widely across industries and company types.

Before you negotiate a raise or make a career move in an AI-exposed field, find out where your specific skills stand in the current market. Upload your resume to PayScope; it's free, takes under ten minutes, and pulls market rates from active job postings rather than salary surveys. Once you know your market rate by skill and location, decisions about negotiating, upskilling, or switching roles become data problems rather than guesses.

This article is based on "Labor market impacts of AI: A new measure and early evidence" by Maxim Massenkoff and Peter McCrory, published March 5, 2026. The full paper is available at anthropic.com/research/labor-market-impacts.