What Anthropic's AI Jobs Report Actually Says

Why this report is different from everything else

Most AI job risk research answers the question: "what could AI theoretically do?" Researchers read job descriptions, break them into tasks, and ask whether current AI systems could perform each task. That is useful - but it tells you about AI's ceiling, not what is actually happening on the ground right now.

Anthropic's approach is different because they have something no academic researcher has: actual data on how Claude is being used at work. They can see which occupational tasks are genuinely being handled by AI in real workplaces, at scale, today. That is what "observed exposure" measures. It is not a prediction. It is a measurement.

Think of it this way. A car can theoretically travel at 200 miles per hour. But if you measure how fast people actually drive on their morning commute, you get a different number. The difference between those two numbers is the gap. The Anthropic report measures that gap for AI and work - for approximately 800 occupations.

The single most important finding: the gap

The report's central finding is that AI is doing far less than it theoretically could in most workplaces. The gap between capability and actual deployment is large - and it varies enormously by occupation.

Here is the core data in the report for the three occupation groups with enough data to measure precisely:

Source: Massenkoff, M. and McCrory, E. (2026), Labor Market Impacts, Anthropic Research, March 5, 2026.

The gap is not random. The smaller the gap, the more disruption is already happening. An 8-point gap for customer service clerks means the deployment has nearly caught up with the theoretical maximum - the people in those roles are competing with AI that is already doing most of what they do. A 61-point gap for ICT professionals means the technology exists but deployment has barely started for most of what those workers do.

The top jobs in Anthropic's data: who Claude is actually replacing

The report identifies the occupations with the highest "coverage" - meaning the largest share of their tasks are actively being handled by Claude in real workplaces. The top three with exact numbers are:

• Computer Programmers: 75% task coverage. Three quarters of what a computer programmer does is actively being performed by Claude across real workplaces right now.
• Customer Service Representatives: second-highest ranked. The exact percentage is not published, but the report confirms they follow directly after programmers.
• Data Entry Keyers: 67% task coverage. Two thirds of data entry work is being handled by AI in practice.

The coverage figure is important to understand correctly. It does not mean "75% of computer programmers are being replaced." It means 75% of the tasks that make up a programmer's job are the kinds of tasks where Claude is actively being used in real workplaces. A programmer at a company that has not adopted AI tools yet is untouched. A programmer at a company that has deployed AI coding assistants is already working in a different job than they were two years ago.

The 30% who have nothing to worry about (yet)

This is one of the most important findings in the report that almost no coverage mentioned. About 30% of US workers are in occupations with zero observed AI exposure. Zero. Claude is not doing any of their tasks in any measurable way.

The report lists examples: cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers, dressing room attendants. These are roles defined by physical presence, manual dexterity in unpredictable environments, and real-time reaction to messy real-world conditions. AI can generate a recipe, but it cannot stand in a kitchen and cook to order during a Friday dinner rush. AI can describe lifesaving technique, but it cannot watch a pool and respond in seconds to a struggling swimmer.

For these workers, the Anthropic data is essentially reassuring in the near term. The technology that would threaten these roles - general-purpose physical robots with reliable real-world dexterity - does not exist at scale in 2026. This is not a prediction that it never will exist. It is a measurement that it does not exist now.

Has AI caused unemployment? What the data actually says

This is the question most people want answered, and the report gives a careful, honest answer: no systematic increase in unemployment has been detected for high-exposure workers since late 2022 - the period after ChatGPT launched.

The baseline unemployment rate for workers in the top quartile of AI exposure is about 3%. Since ChatGPT's release, that number has not moved in a statistically significant way. The same occupations that were most exposed to AI two years ago have not seen a wave of job losses show up in unemployment statistics.

There are three reasons this could be true, and the report is honest about all three:

• AI has not yet caused unemployment. The gap between capability and deployment (the 61-point gap) means AI is augmenting work more than replacing workers. The tools exist but the displacement has not happened at scale yet.
• The signal is too early to detect. Layoffs and workforce reductions take time to show up as sustained unemployment. A 1-percentage-point increase in unemployment would be detectable with this data. Smaller effects would not be.
• Workers are moving between roles. AI may be shrinking headcount within categories while workers transition to other work - which would show up as reduced hiring rather than increased unemployment.

The report is explicit: the researchers cannot tell you which of these three is driving the result. That honesty matters. "No unemployment increase detected" is not the same as "AI is not affecting the labour market." It means the effect, if it exists, is either small enough that current measurement tools cannot see it yet, or it is showing up through hiring slowdowns rather than unemployment spikes.

The young worker signal - the most important early warning in the report

Here is where the data gets genuinely interesting, and where the early signal of labour market change is visible. While overall unemployment in exposed occupations has not moved, the job-finding rate for workers aged 22-25 in high-exposure occupations has dropped by approximately 14% compared to 2022.

What does that mean in plain terms? Before ChatGPT, about 2% of young workers in low-exposure occupations found a new job each month. That rate stayed stable. But for young workers in high-exposure occupations, that rate has declined - they are finding it harder to get hired into those roles.

The report is careful to flag that this finding is "just barely statistically significant" - meaning it is real enough to report but not strong enough to be certain about. It could be noise. But it is the kind of signal that makes sense given what we know about how AI adoption happens in workplaces: employers stop hiring entry-level workers into roles where AI tools are handling more of the workload. The experienced worker stays. The junior position disappears. The result shows up as reduced hiring of young people in those fields before it shows up as any increase in overall unemployment.

What 68% of Claude usage tells you about which tasks are going first

The report contains a specific data point about how Claude is being used that reveals a lot about which tasks are being automated first. 68% of observed Claude usage in work contexts involves tasks that are rated as "fully feasible" for AI - tasks that AI can complete on its own, without a human in the loop. Only 3% of Claude's work-related usage involves tasks that are not feasible for AI at all.

This tells you something important about the nature of AI adoption in workplaces. Employers are not using AI for the hard, ambiguous things. They are deploying it on the easy, structured, high-volume things first. The tasks being automated first are the ones with clear inputs and outputs, consistent formats, and low risk of error. That is precisely why customer service clerks are seeing near-full deployment while ICT professionals are not - a customer inquiry is structured and repeatable. A complex software architecture decision is not.

The implication for workers is direct: if your role is mostly structured, repetitive, rule-following tasks - even if they feel complex to you - you are in the category that employers are deploying AI on first. If your role requires judgment, ambiguity-handling, novel problem-solving, or accountability - those are the tasks that represent the remaining 32% of Claude usage that is augmentative rather than fully automated.

The employment projection signal: slower growth for exposed jobs

The Bureau of Labor Statistics publishes 10-year employment growth projections for every major occupation. The Anthropic report checked whether their observed exposure measure predicts those projections - and it does. For every 10 percentage point increase in observed AI exposure, BLS projects 0.6 percentage points less employment growth over 2024-2034.

This is a modest effect, but it is directionally consistent with what you would expect if AI adoption is beginning to suppress hiring in the most-exposed fields. A 60-percentage-point exposure difference (roughly the gap between an ICT professional and a cook) translates to a 3.6 percentage point difference in projected employment growth. That is meaningful at the scale of millions of workers over ten years.

It also tells you that the BLS - which uses its own independent methodology - is essentially already pricing in some AI effect in its projections. The labour market data and the Anthropic usage data are pointing in the same direction.

The bottom line: what this report is actually telling you

The Anthropic report does not say AI is about to cause a mass unemployment crisis. It does not say AI is harmless. It says something more precise and more useful than either of those things:

• AI is genuinely doing a large share of tasks in some occupations right now - especially customer service and data entry. The disruption in those categories is not coming. It is here.
• In other occupations - particularly ICT and software - the theoretical exposure is high but the actual deployment is low. The gap provides runway. But the direction is set, and the gap is narrowing.
• Overall unemployment has not moved. But hiring for young people entering high-exposure fields has slowed - the first detectable early signal of AI changing who gets hired.
• About 30% of workers are in roles where AI has essentially zero observed usage. For those workers, other concerns (wages, conditions, physical safety) matter more than AI risk right now.

The most useful way to read this report is not as a list of jobs that are "safe" or "at risk." It is as a snapshot of where the AI adoption wave has reached - and how far it still has to travel to reach its theoretical maximum. Your runway is the distance between where the wave is now and where it could theoretically reach. For some workers that is a few years. For others it is a decade or more. The Anthropic data is the best available measurement of that distance.

Frequently asked questions