The Divide Between What AI Can Do and What It Actually Does

The research, published in March 2026 by Anthropic economists Maxim Massenkoff and Peter McCrory, introduces a concept called "observed exposure" (a phrase the researchers use themselves). The measure tracks the share of real work tasks that AI systems are actually performing, versus what they could theoretically perform. They pulled this directly from millions of real Claude conversations matched against 800 occupations.

The result is the most honest picture of AI's labor market impact we have seen from any major AI company, precisely because it uses their own data and they had every incentive to be careful.

The number that resets the conversation

Theoretical exposure asks what AI could do. Observed exposure asks what AI is actually doing. The first generates headlines. The second tells you what is happening inside organizations today.

For computer programmers, observed coverage sits at 75%. For customer service and data entry roles, observed coverage is also high. For business, finance, and management occupations, the divide between theoretical capability and actual observed use is wide. AI could do far more in these roles than workers and organizations are currently asking it to do.

This is worth sitting with. If you are running a company and your team is not closing that divide, someone else's team will.

The hiring signal that deserves attention

The research also surfaces a labor market signal worth watching. After ChatGPT launched, the job-finding rate for workers ages 22 to 25 in highly AI-exposed occupations dropped an estimated 14%. The researchers note this finding is just barely statistically significant. It is a suggestive early signal, not a confirmed trend. But the direction is clear.

Entry-level positions in AI-exposed fields are narrowing as organizations pause junior hiring in roles where AI output is expanding what senior workers can produce. The pipeline into knowledge-work careers is changing shape before the displacement shows up in unemployment numbers.

The researchers name a scenario they call "a Great Recession for white-collar workers." They present this as a possibility, not a prediction. But naming it at all, from Anthropic's own economists, tells you something about where serious researchers are placing the risk.

What the scenario actually describes

Unemployment in the most AI-exposed occupational quartile doubling from 3% to 6% would be a significant disruption to knowledge-work careers, even if it looks modest in percentage terms. We are not there. The current data shows no statistically significant increase in unemployment for highly AI-exposed workers. But the young-worker hiring signal suggests the pipeline is already changing.

What this looks like inside an organization

The Federal Reserve published its own AI adoption monitoring report in April 2026. The picture it shows explains why the Anthropic divide exists.

About 18% of US firms have adopted AI as of year-end 2025. Yet 78% of workers are employed at firms that have adopted AI. The difference between those two numbers reflects a simple reality: large firms adopt first, and large firms employ most people.

Financial sector firms show 30% adoption at the firm level and 63% individual generative AI adoption among workers. Professional services: 33% and 62%. Manufacturing lags both significantly.

So when Anthropic finds that computer and math occupations sit at 33% observed exposure despite 94% theoretical capability, the tools are not the bottleneck. The fluency to use them fully is still being built across organizations. That is a proficiency problem, not a technology problem.

The level where the divide closes

On the 7 Levels of AI Proficiency, the observed-versus-theoretical divide shows up between Level 2 and Level 4.

A Level 2 worker (AI Capable) knows AI exists and uses it occasionally. Their observed exposure is low. A Level 4 worker (AI Architect) has built AI into their workflow architecture. Their observed exposure approaches the theoretical ceiling for their role.

The workers moving between those levels are Level 3 (AI Fluent): actively building the habits and judgment that close the observed-theoretical divide. They are also the workers whose employers see the first productivity returns.

Deloitte's 2026 State of AI in the Enterprise report finds that 66% of organizations report productivity and efficiency gains from AI adoption, but only 25% have moved 40% or more of their AI pilots into production. The productivity is real. The scaling is not. That is still a proficiency bottleneck, not a technology one.

The organizations crossing the observed-exposure divide are not doing so by deploying more tools. They are doing it by developing people at Level 3 and above who know how to use what already exists.

Related reading: Level 4: AI Architect.

What to do with this

You do not need to worry about AI doing 94% of anyone's job next quarter. The Anthropic data makes clear that the actual observed rate is a fraction of theoretical capability, and it has been for years.

What you do need to understand is where your team sits in that divide.

If your knowledge workers are at Level 1 or Level 2, they are in the theoretical exposure zone. AI could handle parts of their work. They are not asking it to. That is both a productivity issue and a retention risk as employers who develop Level 3 and Level 4 workers pull ahead.

The window is open. The divide is real and it is large. The organizations crossing it are not doing so by deploying more tools. They are doing it by building the people who know how to use what they already have.

Harrison Painter

AI Business Strategist. Founder, LaunchReady.ai and AI Law Tracker.

Harrison helps teams build AI systems that cut cost and grow revenue. Nearly 20 years of business experience. 2.8M YouTube views. Founder of LaunchReady.ai and the 7 Levels of AI framework.

Connect on LinkedIn