Anthropic Just Mapped Which Jobs AI Will Hit First

Subtitle: The company behind Claude analyzed its own usage data to build an early warning system for AI job displacement. Here’s what they found, and what it means for you.

Key Takeaways

1. Anthropic researchers built a new metric called “observed exposure” that measures how AI is actually being used at work, not just what it could theoretically do.
2. Computer programmers top the list at 75% task coverage. Customer service reps and data entry workers follow close behind.
3. The gap between what AI can do and what it’s actually doing is enormous. In computer and math fields, AI handles 33% of tasks but could theoretically cover 94%.
4. Young workers (ages 22-25) are already seeing a 14% drop in job-finding rates in AI-exposed fields.
5. The researchers warn that a “Great Recession for white-collar workers” is a measurable scenario, not just speculation.

Here’s something you don’t see every day: an AI company studying whether its own product might eliminate your job.

On March 5, 2026, Anthropic, the company that makes Claude, published a research paper that should be required reading for anyone who works at a desk. The paper is called “Labor market impacts of AI: A new measure and early evidence,” and it does something no other study has done before. Instead of guessing which jobs AI might threaten someday, Anthropic’s researchers analyzed actual usage data from millions of Claude conversations to see which professional tasks AI is already performing in the real world.

The result is the most concrete picture we’ve had yet of where AI is headed in the workplace. And the answer is: it’s complicated, but not in the way most headlines suggest.

What Makes This Study Different

Most AI-and-jobs research falls into one of two camps. Either it’s purely theoretical (“AI could do 80% of this job’s tasks”) or it’s based on surveys asking workers how worried they feel. Neither approach tells you much about what’s actually happening.

Anthropic’s researchers, economists Maxim Massenkoff and Peter McCrory, took a different route. They built a metric they call “observed exposure.” It combines three data sources:

• Real usage data from the Anthropic Economic Index, which tracks how Claude is being used in professional settings
• Task definitions from the O*NET database, which breaks down more than 800 occupations into specific tasks
• Theoretical capability estimates from earlier research on what large language models could potentially handle

The key innovation is measuring what AI is actually doing at work, not just what it could do. And the gap between those two things turns out to be the most important finding in the entire paper.

The Jobs at the Top of the List

Here are the occupations with the highest “observed exposure” to AI, meaning AI is already handling the largest share of their day-to-day tasks:

At the other end of the spectrum, jobs with zero AI coverage include cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers, and dressing room attendants. About 30% of all workers fall into this “zero exposure” category, almost entirely in jobs that require physical presence and manual skills.

The pattern is clear: the more your job involves processing information, writing, analyzing data, or communicating in text, the more AI is already being used to do parts of it.

The Enormous Gap Between “Could” and “Does”

Here’s the most surprising finding in the study, and the one that should shape how you think about all of this.

In computer and math occupations, AI theoretically has the capability to handle 94% of tasks. But in practice, it’s currently covering only 33%. That’s a massive gap. For office and administrative roles, the theoretical number is around 90%, but actual usage is far lower.

Why such a big difference? The researchers point to several barriers:

• Legal constraints: Many industries require human oversight by law
• Model limitations: AI still makes mistakes that matter in high-stakes work
• Software gaps: Connecting AI to existing work systems isn’t simple
• Trust and adoption: Many professionals haven’t integrated AI into their workflows yet

This gap is both reassuring and concerning. Reassuring because it means the sky isn’t falling tomorrow. Concerning because it represents an enormous amount of automation potential that hasn’t been tapped yet. Every improvement in AI capability, every new integration, every corporate policy change that encourages adoption chips away at that gap.

Who’s Most Exposed? A Demographic Profile

The study found that workers in the most AI-exposed occupations share a distinct profile:

• 16 percentage points more likely to be female than workers in unexposed jobs
• 47% higher average earnings
• Nearly four times more likely to hold a graduate degree
• 17.4% hold graduate degrees (compared to 4.5% in unexposed occupations)
• Tend to be older and white

This flips the usual narrative about automation. Previous waves of technology displacement (factory robots, self-checkout kiosks) primarily affected lower-wage, less-educated workers. AI is different. It’s coming for the jobs that used to be considered safe because they required education and cognitive skills.

The Early Warning Signs

Here’s where the paper gets uncomfortable. While the researchers found no systematic rise in unemployment among exposed workers yet, they did find one early signal that’s hard to ignore.

Young workers (ages 22-25) in AI-exposed fields are seeing a 14% decline in job-finding rates.

The baseline monthly job-finding rate for young workers in unexposed occupations is about 2%. In exposed fields, that rate has dropped by roughly half a percentage point since ChatGPT’s release. It’s not a massive number on its own, but the trend line is moving in one direction.

What this likely means: companies aren’t firing existing employees because of AI. Instead, they’re hiring fewer new ones. The entry-level pipeline is narrowing. If you’re a recent graduate looking for your first job in programming, data analysis, or financial services, you might already be feeling this. It’s not that the job disappeared. It’s that the position was never posted because a team of four decided AI tools let them do the work of five.

The “Great Recession” Scenario

The researchers were careful to frame their work as an early detection system, not a prediction. But they did name the scenario that everyone in white-collar work should be watching for.

During the 2007-2009 financial crisis, U.S. unemployment doubled from 5% to 10%. The researchers note that a comparable doubling among highly AI-exposed workers, from about 3% to 6%, would be clearly detectable using their framework. They call this the potential “Great Recession for white-collar workers.”

They’re not saying this will happen. They’re saying they’ve built a system that can detect it early if it starts to happen. Think of it as a smoke detector, not a fire alarm. Right now, it’s not going off. But they wouldn’t have built it if they didn’t think there was reason to have it plugged in.

What the 97% Number Tells Us

One finding that flew under the radar: 97% of observed Claude usage falls into task categories that researchers had already identified as theoretically feasible for AI. And 68% of usage involves tasks rated as fully automatable by language models.

This means AI isn’t being used for unexpected things. It’s being used for exactly what researchers predicted. The surprise isn’t what AI is doing. It’s how much room there still is for it to do more.

What This Means for You

If you work in one of the highly exposed fields, here’s how to read this report:

Don’t panic. The data shows that AI is augmenting work more than replacing it right now. The gap between capability and adoption is your window to adapt.

Pay attention to entry-level hiring. If your company is posting fewer junior positions, or if job boards in your field seem thinner than they used to, that’s the early signal the researchers identified. It doesn’t mean layoffs are coming. It means the shape of the workforce is changing.

Learn the tools. The study essentially measured how much people are already using AI at work. If 33% of computer and math tasks are being done by AI and you’re not using it, you’re the person who refused to learn email in 1998. You might keep your job for a while, but you’re falling behind.

Watch the gap close. That 33% vs. 94% gap? Every time AI gets better at avoiding mistakes, every time your company’s IT team connects a new AI tool to your workflow, every time a regulation is updated to allow AI-assisted work, that gap shrinks. The speed at which it closes is the real story of the next five years.

If you’re a young professional, this is especially relevant. The 14% hiring slowdown is real, even if it’s still small. Differentiate yourself by being the person who knows how to use AI effectively, not the person who competes against it.

An AI Company Studying Its Own Impact

There’s something worth noting about who published this research. Anthropic builds Claude. They’re analyzing their own product’s effect on jobs. You might read that as a conflict of interest, and you’d be partly right. But you could also read it the other way: they have access to usage data that nobody else has. They can see exactly how their AI is being used in the real world, which tasks, which industries, which patterns.

No other company has published this kind of analysis at this level of detail. Whether Anthropic is doing this out of genuine concern, regulatory positioning, or both, the data itself is valuable. And the framework they’ve built, designed to be updated as adoption changes, means this won’t be a one-time snapshot. It’s a monitoring system.

The question isn’t whether AI will change how we work. It already has. The question is how fast the gap between “could” and “does” continues to close, and whether we’ll be paying attention when the smoke detector goes off.

The full research paper, “Labor market impacts of AI: A new measure and early evidence,” is available on Anthropic’s research page. The Anthropic Economic Index is published at anthropic.com/economic-index.