Anthropic Says AI Can Do 94% of Your Job. Here's What Happens When It Starts Trying.

Anthropic just published the most data-driven study yet on AI and the labor market. The headline numbers are striking, but the real story is in the gap between them.

The paper, titled Labor market impacts of AI: A new measure and early evidence, was authored by Anthropic economists Maxim Massenkoff and Peter McCrory. Unlike most AI-and-jobs research, which estimates what AI could theoretically do, this study measured what people are actually using Claude for at work. They call the metric "observed exposure."

The findings paint a picture of a wave that's building but hasn't crested yet.

The gap

For computer and math workers, AI can theoretically handle 94% of their tasks. In practice, Claude is being used for 33% of them. Office and admin roles show a similar split: 90% theoretical, roughly 25% observed. Business and finance: 85% theoretical, about 20% in use.

Across the board, the red area (actual AI usage) is a fraction of the blue area (theoretical capability). The researchers are blunt about what happens next: as capabilities advance and adoption deepens, the red will grow to fill the blue.

Why the gap exists (for now)

The gap between what AI can do and what it's actually doing comes down to three things: legal constraints that prevent automation of certain tasks, model limitations that make some theoretically possible tasks unreliable in practice, and the need for humans to verify AI work before it counts.

These are real barriers. They're also temporary. Every model release chips away at the limitations. Every new integration removes a friction point. Every team that sees a competitor ship faster with agents accelerates their own adoption timeline.

The researchers note that 97% of the tasks observed in Claude's professional usage fall into categories already rated as theoretically feasible. The capability is there. The adoption just hasn't caught up.

Who's most exposed

The demographics of AI exposure challenge the usual narrative. This isn't a story about factory floors or call centers (not yet, anyway). The workers in the most exposed occupations are more likely to be older, more educated, and better paid. They're 16 percentage points more likely to be female. They're more likely to hold graduate degrees.

Computer programmers sit at the top with 75% of their tasks already covered by AI usage. Customer service representatives follow at 70%. Data entry workers at 67%.

The researchers also found a correlation between higher AI exposure and weaker job growth forecasts from the Bureau of Labor Statistics through 2034. Occupations where AI is doing more today are projected to grow less over the next decade.

The hiring signal

One finding stands out. While the study found no systematic increase in unemployment for workers in highly exposed occupations, it did find something else: a 14% drop in the job-finding rate for workers aged 22 to 25 in exposed occupations, compared to 2022 levels.

Young workers aren't being fired. They're not being hired. The entry points are narrowing before the wave has fully arrived.

Now zoom out

Here's what the research doesn't address, because it's not what the paper was designed to measure: as that gap closes, as agents take on more tasks across more roles, who's watching what they do?

Every percentage point of that gap closing means more AI agents sending emails, accessing databases, modifying files, querying APIs, and taking actions on behalf of workers and companies. Today it's 33% of computer and math tasks. When it's 60%, or 80%, the volume of autonomous agent actions in any given organization will be enormous.

The Anthropic study itself was built on Claude usage data. That means Anthropic has direct visibility into how their AI is being used in professional settings. They can measure which tasks are automated versus augmented, which occupations are most affected, and how usage patterns shift over time.

Your company's agents? The ones you're deploying with OpenClaw, or wiring up to your internal tools, or giving access to customer data? Nobody has that visibility unless you build it.

The governance gap is the real gap

The research community is focused on the gap between theoretical AI capability and observed adoption. That's the right economic question. But there's a second gap that grows in lockstep: the gap between what agents are doing and what anyone can verify, audit, or control.

Right now, most teams deploying AI agents have no structured way to answer basic questions:

What did this agent access? Did it have permission to do that? Who approved it? Can we prove it?

As the theoretical-to-observed gap closes and agents flood into more workflows, these questions become urgent. Not because agents are malicious. Because they're fast, autonomous, and increasingly capable of taking consequential actions without a human in the loop.

What this means for governance

The Anthropic paper frames its work as a first step toward understanding AI's labor market impact, with plans to update the analysis as new data emerges. That's exactly the right approach for the economic question.

For the governance question, teams can't wait for the next paper. Every agent deployed today without permission controls, audit trails, and approval workflows is operating in a gap that grows wider with every capability improvement.

Multicorn Shield sits between your agents and the systems they act on. It enforces permissions before actions execute, logs everything to a tamper-evident audit trail, and gives you approval workflows for high-risk operations. It's the infrastructure layer for the wave the Anthropic research describes.

The red is going to fill the blue. The question is whether you'll have governance in place when it does.