Anthropic’s job exposure data shows an enormous gap between what AI can do and what AI is actually doing. The composition of that gap is the most interesting part of the dataset.

Anthropic published a paper in March called Labour Market Impacts of AI: A New Measure and Early Evidence. Most of the coverage focused on the headline numbers – which jobs are most exposed, which are least, projected impacts on employment. Worth reading on its own.

The part that didn't get enough attention is the structural finding underneath those numbers.

For every major occupation, the paper distinguishes between two metrics:

• Theoretical AI capability: what AI could do based on task analysis
• Observed AI coverage: what AI is actually being used for right now, measured from real Claude usage data

The gap between those two is enormous and consistent across sectors:

The headline reading is "AI capability is way ahead of adoption." That's true but it's the surface reading. The more interesting question is what specifically lives in that gap, and whether the things in the gap are temporary or permanent.

The composition of the gap, based on the paper's analysis:

1. Legal and compliance constraints. Tasks AI could do but isn't being used for because regulations require a human in the loop, or because liability frameworks haven't caught up. This is a large chunk of legal, healthcare, and financial work.
2. Software integration friction. Tasks AI could do but currently can't because the data is locked in legacy systems that don't expose APIs, or because workflows require human handoffs between tools that aren't connected. Large chunk of administrative and back-office work.
3. Verification overhead. Tasks AI could do at machine speed but in practice take human time to check, which eliminates most of the speed advantage. Common in coding, research, and data analysis.
4. Workflow inertia. Tasks AI could do but where the existing process is socially embedded – meetings, decisions, established communication patterns – and changing the process is harder than the technology problem. Common in sales, management, and consulting.
5. Quality threshold effects. Tasks where AI output is technically possible but consistently 10-15% below the quality bar that matters in practice. Common in creative work, complex writing, and any task where edge cases dominate.

The paper is clear that the researchers consider all five of these temporary – barriers that are eroding rather than holding. Categories 2 and 3 (integration friction and verification overhead) are eroding fastest, because they're being addressed by infrastructure investments and tooling improvements. Categories 1, 4, and 5 are eroding more slowly because they involve law, social dynamics, and quality thresholds rather than just engineering.

Why this matters more than the headline numbers:

If you're trying to forecast how AI exposure will play out for any specific role, the headline number (current observed coverage) is misleading. What you actually want to know is which of those five gap categories your role's protection is built on.

A role currently at 20% observed coverage is in a different position depending on whether the remaining 80% is:

• Locked behind compliance constraints (slow erosion)
• Locked behind integration problems (fast erosion – probably gone within 2-3 years)
• Locked behind quality thresholds (medium erosion – improving with each model generation)
• Locked behind workflow inertia (slow erosion – but cliff-edge once it goes)

Two roles at the same observed exposure level can have very different future trajectories depending on which category their protection lives in. The headline number doesn't tell you that. The composition does.

The rough framework I use to read my own role through this:

For each task in your work, ask: if AI couldn't do this task today, why not? Then categorise the answer into one of the five categories above. The mix tells you how durable your current position is, more accurately than any single exposure number.

Tasks protected by compliance or workflow inertia are durable for a few years even at high theoretical exposure. Tasks protected by integration friction or verification overhead are exposed soon, even at low current observed exposure. Tasks protected by quality thresholds are middle – improving model generations close those gradually rather than suddenly.

A note on the data source:

Anthropic measured observed coverage from real Claude usage. That means the dataset reflects what early adopters and AI-native workers are doing, not the average worker. The actual gap is probably larger than the table suggests, because Anthropic's user base skews toward people already using AI heavily. The 33% observed coverage for computer & mathematical occupations is what Claude users in that field are doing. Across the field as a whole, the number is lower. This makes the gap conclusion stronger, not weaker.

I built a free resource that runs your specific role through this framework – takes your tasks, scores each one against the five categories above, and gives you a durability assessment alongside the raw exposure score, here if it helps.

If you do nothing else after reading this, run the five-category test on your own role. The composition of your protection matters more than the level of it.