Anthropic's Own Data: Programming at 74.5% AI Coverage
Anthropic published observed usage data from millions of Claude interactions matched against 800 US occupations. Seven charts that tell you exactly where AI is today. The gap between 'real' and 'slower' is where the opportunity lives.
In March 2026, Anthropic published something unprecedented: their own labor market impact data. Not projections. Not models. Not surveys. Observed usage data from millions of Claude interactions, matched against 800 US occupations from the O*NET task database.
Seven charts that tell you exactly where AI is today and where it's going. Here are the exact numbers.
Chart 1: Where Claude's Compute Actually Goes
| Task Feasibility | Share of Usage |
|---|---|
| Tasks an LLM can handle alone (beta = 1) | 68% |
| Tasks needing additional tools (beta = 0.5) | 29% |
| Tasks classified as not feasible for AI (beta = 0) | 3% |
97% of what people use Claude for is within theoretical AI capability. But this 97% covers only a narrow slice of all tasks across all occupations. People use Claude heavily for what it's good at. They don't use it at all for most other tasks in their job description.
Chart 2: The Radar — Theoretical vs Observed
The single most important image in the paper. A radar chart where blue (theoretical coverage) extends to the outer edges and red (observed coverage) huddles in the center.
| Category | Theoretical | Observed | Gap |
|---|---|---|---|
| Business & Finance | 98% | 35% | 63% |
| Computer & Math | 94% | 33% | 61% |
| Management | 90% | 20% | 70% |
| Office & Admin | 88% | 28% | 60% |
| Legal | 85% | 10% | 75% |
| Architecture & Engineering | 82% | 18% | 64% |
| Education & Library | 78% | 8% | 70% |
| Arts & Media | 75% | 15% | 60% |
| Life & Social Sciences | 68% | 12% | 56% |
| Healthcare Practitioners | 55% | 5% | 50% |
| Sales | 50% | 22% | 28% |
Legal has the largest gap: 75% of automatable work untouched. Management: 70%. Education: 70%. Healthcare: 50%.
The red blob in the center is tiny. In every single knowledge work category, the majority of automatable work is not being automated.
Chart 3: The Top 10 Occupations
The occupations where AI is most actively being used:
| Rank | Occupation | Coverage | Leading Task |
|---|---|---|---|
| 1 | Computer programmers | 74.5% | Write, update, maintain software |
| 2 | Customer service reps | 70.1% | Handle inquiries and complaints |
| 3 | Data entry keyers | 67.1% | Read and enter data |
| 4 | Medical record specialists | 66.7% | Compile and code patient data |
| 5 | Market research analysts | 64.8% | Prepare reports, illustrate findings |
| 6 | Sales reps | 62.8% | Contact customers, demonstrate products |
| 7 | Financial analysts | 57.2% | Analyze financial info |
| 8 | QA testers | 51.9% | Modify software to correct errors |
| 9 | InfoSec analysts | 48.6% | Perform risk assessments |
| 10 | Computer support | 46.8% | Answer user inquiries |
Look beyond #1 and #2. Medical records at 66.7%. Sales reps at 62.8%. Market research at 64.8%. These are not tech jobs. These are departments that exist in every company. Customer support, sales, market research, quality assurance. Every organization has them.
30% of all workers have zero observed exposure: cooks, mechanics, lifeguards, bartenders, dishwashers. Physical work remains untouched.
Chart 4: The Developer Paradox
A scatter plot of AI exposure vs projected employment growth (2024–2034). The overall trend: negative. More exposure correlates with less growth. Slope: -6.07 per 10 percentage points of exposure.
But software developers are the massive outlier: 74.5% exposure AND +16% projected growth. The biggest positive outlier on the entire chart.
Why? Demand for software is growing faster than AI can replace developers. Every business needs AI integrations, automation workflows, agent systems. The total market for what developers build is exploding. AI makes each developer more productive, but the pie grows faster than the slices get efficient.
Customer service reps: high exposure, negative growth. Shrinking. Electricians: zero exposure, +10% growth. Untouched. Software developers: high exposure, +16% growth. Paradox.
Chart 5: Who Gets Hurt
The demographics of AI-exposed vs unexposed workers:
| Metric | Most Exposed | Least Exposed |
|---|---|---|
| Hourly wage | $32.69 | $22.23 |
| Female | 54.4% | 38.8% |
| Bachelor's degree | 37.1% | 13.3% |
| Graduate degree | 17.4% | 4.5% |
| Married | 54.9% | 44.6% |
| Average age | 42.9 | 41.0 |
| Union membership | 5.3% | 11.7% |
AI is not eating the bottom of the labor market. It's eating the educated, well-paid middle.
The most exposed workers earn $65–70K/year, have college degrees, are married, and are in their early 40s. These are not entry-level workers. These are mid-career professionals with mortgages and families.
The least exposed workers are earning working-class wages, doing physical work, without degrees. The ironworker is safer than the financial analyst. The plumber is safer than the market researcher.
What the Data Actually Tells You
1. The Gap Is the Market
60–75% of automatable work isn't being automated. That's not a failure of AI. It's a failure of deployment. The technology exists. The implementation doesn't. Every percentage point of that gap that gets closed is a business opportunity.
2. Horizontal Coverage Is a Myth
"Computer programmers" at 74.5% sounds like "tech is covered." But the entire "Computer & Math" category is at 33%. Because the category includes architects, DBAs, network engineers, security specialists, project managers. Pure coding is automated. Everything around coding is barely touched.
Same pattern in every category. The most automatable single task gets picked up fast. The rest of the occupation stays untouched.
3. Speed of Change Is Measurable
Agent session lengths doubled in 3 months. Trust metrics are improving. The gap is closing, but at a pace that suggests 2–5 years before most categories cross 50% observed coverage.
4. Physical Work Is the Buffer
The 30% of workers with zero exposure won't be affected by this wave. Construction, food service, transportation, maintenance. These jobs are safe not because they're low-skill, but because they're physical. AI needs a body to cook food. It needs a body to fix pipes. It doesn't need a body to analyze spreadsheets.
The Meta-Insight
Previous AI impact studies were theoretical. "If AI could do X..." Anthropic's paper is observational. "AI is doing X."
The difference matters. Theoretical studies predicted wide-scale disruption. Observational data shows narrow-but-deep disruption: a few occupations heavily affected (programming, customer service, data entry), most occupations barely touched, and a massive gap between what's possible and what's happening.
The disruption is real. It's also slower than the headlines suggest. And the gap between "real" and "slower" is where the opportunity lives.
All data from Anthropic's "Labor Market Impacts of AI: A New Measure and Early Evidence" (March 5, 2026). Authors: Maxim Massenkoff, Peter McCrory.