For the past few years, we've been hearing the same question: "Will AI take our jobs?" The answers tend to cluster at two extremes: either "everything will change and millions will be unemployed" or "it's overhyped, don't worry." Anthropic's research team brought something different to this debate: real data instead of theoretical projection.

Published on March 5, 2026, the study titled Labor Market Impacts of AI: A New Measure and Early Evidence combined Claude's real usage data with the US national occupational database to develop a new metric. They called it "observed exposure." And the findings show that both the optimistic and the alarming predictions are wrong.

The Gap Between Theoretical Capacity and Real-World Usage

The most critical finding of the research: AI is not yet doing most of what it theoretically could do.

According to a 2023 study by Eloundou and colleagues, the share of work tasks where a language model could provide a speed advantage approaches 90% in some occupational categories. In computer and math occupations, that figure is 94%. In office and administrative roles, it's 90%.

What does actual usage show? According to Anthropic's own data, only 33% of tasks in the computer and math category are actively being handled by Claude. One third of theoretical capacity.

The reasons for this gap are varied: legal restrictions, the need for human verification, dependencies on proprietary software, and the fact that the technology hasn't fully diffused yet.

"AI is far from reaching its theoretical capability: actual coverage remains a fraction of what's feasible." — Anthropic, March 2026

The Most Exposed Occupations

The research ranked the ten occupations with the highest "observed exposure" scores. This list differs significantly from earlier studies because it's grounded in real Claude usage data, not abstract projections.

• Computer Programmers — 75% coverage

• Customer Service Representatives — approximately 70% coverage

• Data Entry Operators — 67% coverage

• Financial Analysts — approximately 60% coverage

• Cooks, Bartenders, Lifeguards — 0% coverage

What Does the Unemployment Data Say?

This may be the most surprising finding in the research. Since ChatGPT's launch, no systematic increase in unemployment has been observed among workers in the highest AI-exposure occupations.

Researchers used Current Population Survey data to track unemployment rates for the highest and lowest exposure workers since 2016. Aside from the large divergence during COVID (when physical jobs were hit harder), the two groups have moved almost in parallel.

But there is one warning signal: a statistically significant slowdown was detected in entry into high-exposure occupations among younger workers aged 22 to 25. In the post-ChatGPT period, new worker entry into these occupational groups declined by approximately 14%. The same effect is not observed in workers over 25.

Why All the Predictions Have Been Wrong

The research offers an instructive historical example. In 2009, Blinder and colleagues calculated that roughly one quarter of US jobs were "susceptible to offshoring." A decade later, most of those jobs were still around.

The contribution of the new study is this: unlike previous metrics that tried to answer the theoretical question of "what can AI do," it answers the empirical question of "what is AI actually doing" using real platform data. And the distance between those two questions is currently very large.

What This Means for Businesses

The research uses data to refute the fear that "AI will make everyone unemployed" — for now. But that doesn't mean you should sit still. Quite the opposite.

Consider: according to the research, AI is still operating at a small fraction of its theoretical capacity. In computer and math occupations, there's 94% theoretical capacity and 33% actual usage. That gap will close over time. And as it closes, those who are ready will grow, and those who are not will shrink.

Look at the picture for businesses. Customer service representatives rank second with approximately 70% exposure. A company's customer communications, reporting, content preparation, and campaign analysis processes sit squarely in the highest-exposure categories of the research. And that's just one department example.

The question is not "should I use AI?" The question is: will you become a company that uses and adapts to AI, or will you become a task that AI covers?

The hiring slowdown among young workers that the research points to is also an important signal. Companies are hiring fewer entry-level employees because AI is taking on some of those tasks. For businesses, this is actually an opportunity: doing more with your existing team without adding headcount is exactly what AI transformation promises.

Research Limitations and What Comes Next

The researchers frame their findings with appropriate caution. The effects they've detected are still small, and some results sit at the edge of statistical significance. The hiring slowdown among younger workers is open to more than one interpretation: are these young people shifting to different sectors, returning to education, or actually exiting the labor market?

The team plans to update the study periodically. One of the most important next steps they highlight: how new graduates trained in high-exposure fields advance in the job market. A student studying computer science today who graduates in 2027 — what environment will they enter?

Is Your Business Ready for This Transformation?

The research is clear: the gap between AI capacity and real-world usage is closing. Right now, during this transition period, let's work together to understand which of your company's processes could be strengthened with AI and how you can scale without growing your team.

At WhiteGate, we analyze company departments and build AI systems tailored to their specific processes. In a free 15-minute consultation, we provide an analysis specific to your company — with no commitment required. If you want to stay ahead of your competitors in this era of transformation, get in touch.

SOURCE

Massenkoff & McCrory (2026), "Labor Market Impacts of AI: A New Measure and Early Evidence", Anthropic. anthropic.com/research/labor-market-impacts