AI can theoretically replace 94% of your job tasks. But in practice, only 33% are actually being covered. What does this 61-percentage-point gap really mean? Anthropic analyzed millions of their own Claude usage data points, and their findings are pretty fascinating.
What is this?
On March 5, 2026, Anthropic economists Maxim Massenkoff and Peter McCrory published a paper titled "Labor market impacts of AI: A new measure and early evidence." The key concept is a new metric called "Observed Exposure." While previous research only looked at "what AI can theoretically do," this study combined actual Claude usage data with the U.S. occupational database (O*NET) to quantify the gap between theory and reality for the first time.
The methodology is quite sophisticated. They first took the theoretical AI exposure scores (beta scores) for each occupational task from Eloundou et al.'s (2023) "GPTs are GPTs" paper, then overlaid actual Claude usage patterns from the Anthropic Economic Index. They weighted automation use cases (as opposed to augmentation) and work-related usage more heavily. So asking ChatGPT for a recipe doesn't count — they filtered for cases where AI is actually being used for work.
The results are striking. 68% of Claude usage was on "fully exposed" tasks (beta=1.0), 29% on "partially exposed" tasks, and just 3% on "non-exposed" tasks. This means AI is being used intensively on the things it can do. The issue is the gap between what it "can do" and what it's "actually doing."
Key distinction: Theory vs. Reality
Theoretical Exposure = percentage of tasks AI is technically capable of performing
Observed Exposure = percentage of tasks confirmed through actual Claude usage data
The difference between these two is the study's core finding. A larger gap means "there's a bigger pie AI hasn't eaten yet," but it also means "adoption barriers are high."
What changes?
Finding 1: 94% vs. 33% — The gap between capability and reality
This is the most eye-catching number. In computer and math occupations, AI's theoretical task coverage is 94%, but actual Claude usage-based coverage is only 33%. Office and administrative roles are theoretically at 90% but only a fraction is covered in practice. The researchers identified four causes for this 61-percentage-point gap: model limitations, legal constraints, the need for additional software integration, and the need for human review of AI output.
Finding 2: Occupation-level "Observed Exposure" rankings
The actual AI exposure rankings from the study paint a quite different picture from theoretical predictions.
| Occupation | Observed Exposure | Theoretical Exposure | Gap |
|---|---|---|---|
| Computer programmers | 74.5% | 94% | 19.5%p |
| Customer service reps | 70.1% | ~90% | ~20%p |
| Data entry specialists | 67.1% | ~85% | ~18%p |
| Computer & math occupations avg. | 33% | 94% | 61%p |
| Chefs, mechanics, bartenders, etc. | 0% | ~5% | ~5%p |
Programmers rank #1 at 74.5%. Customer service (70.1%) and data entry (67.1%) follow. Meanwhile, about 30% of all workers have 0% AI exposure — physical, on-site occupations like chefs, motorcycle mechanics, lifeguards, and bartenders.
Finding 3: Profile of the most AI-exposed workers
This part is somewhat surprising. Workers in the top 25% AI-exposed occupations, compared to the bottom group:
They earn 47% more, are 3.9x more likely to hold a graduate degree (17.4% vs 4.5%), and have 16 percentage points higher female representation. AI isn't targeting low-wage simple labor — it's aiming squarely at highly educated, high-income white-collar workers.
Finding 4: No mass unemployment yet — but there are signals in youth hiring
The researchers' conclusions are cautious. Since ChatGPT's launch (December 2022), no systematic increase in unemployment has been observed in high AI-exposure occupations. But the story is different for 22–25-year-olds. The job finding rate for young workers in high AI-exposure occupations dropped 14%. The researchers acknowledged this figure is "barely statistically significant," but considered it a notable early signal.
A separate study from the Dallas Fed corroborates this. In the top 10% AI-exposed industries, total employment dropped 1% since 2022, with computer systems design down 5% specifically. Meanwhile, wages in the same field rose 16.7% — more than double the national average of 7.5%. The interpretation? Experienced workers get more expensive, while the door narrows for new entrants.
Finding 5: The "White-collar Great Recession" scenario
This is the most talked-about passage in the paper. The researchers referenced how U.S. unemployment doubled from 5% to 10% during the 2007–2009 Great Recession, and warned that something similar could happen in high AI-exposure occupations — a scenario where unemployment doubles from 3% to 6%. Fortune dubbed this the "Great Recession for white-collar workers."
"A shift of this magnitude would be detectable in our framework. It hasn't happened yet, but that is no reason to stop watching."
— Massenkoff & McCrory, Anthropic researchers
"Hasn't happened yet" doesn't mean "won't happen"
Cross-validated with BLS (Bureau of Labor Statistics) independent projections, the researchers found that for every 10-percentage-point increase in AI observed exposure, BLS growth projections drop by 0.6 percentage points. Government forecast data is pointing in the same direction.
The essentials: how to get started
The implications of this research are clear. AI's theoretical capability already covers most white-collar work, but full adoption needs time and the right conditions. So what should you do in the meantime?
- Check your occupation's "Observed Exposure" first
The dataset Anthropic released is available on Hugging Face. Check which of your job tasks are already being covered by AI and which aren't. Even if you're a programmer with "74.5% coverage," the remaining 25.5% could be your core value. - Invest in the 25% AI can't do
The adoption barriers the research reveals — legal constraints, the need for human review, complex judgment calls — are exactly where human value remains. If you're in coding, focus on architecture design and business logic decisions. In customer service, focus on emotional responses and escalation judgment. - Practice using AI as a "colleague," not just a "tool"
As the Dallas Fed data shows, wages for experienced workers in high AI-exposure fields actually rose 16.7%. Experienced workers who leverage AI well are becoming more valuable. Go beyond just asking ChatGPT questions — systematically integrate AI into your workflows. - If you're 22–25, redesign your entry strategy
The decline in youth hiring isn't about "getting fired because of AI" — it's about "companies not hiring new people because of AI." Showcase AI proficiency in your portfolio and emphasize domain expertise that can't be automated. - Reassess every 6 months
The most important message from this research: the fact that 94% theoretical capability sits at 33% actual usage isn't a technology limitation — it's an adoption speed issue. As legal frameworks mature, software integration tools improve, and organizations change, that 33% can climb quickly. According to METR data, the complexity of tasks AI can complete doubles every 7 months.
