Figure 2: Theoretical capability and observed exposure by occupational category
Share of job tasks that LLMs could theoretically perform (blue area) and our own job coverage measure derived from usage data (red area).
This week Anthropic published one of the more honest pieces of labour market research to come from an AI lab. Most organisations in this space either catastrophise or dismiss. This paper does neither. It builds a framework, tests it against real data, and publishes the results before the effects are clear enough to be comfortable.
The headline finding will reassure most people. No systematic increase in unemployment for workers in the most AI-exposed occupations since late 2022. The labour market, by conventional measures, is holding.
They are right. And I think that reading misses the entire point.
What the data actually says
The paper introduces a new metric called observed exposure: a measure that combines theoretical LLM capability with real-world usage data, weighted toward automated and work-related applications rather than casual use. It is a more honest measure than most because it distinguishes between what AI can theoretically do and what it is actually doing in professional settings.
The number that deserves far more attention than it is getting: Claude currently covers 33% of all tasks in the Computer and Mathematics occupational category. Theoretical capability, by their own measure, sits far higher. The gap between what these systems can do and what is actually deployed in the economy is vast.
The authors frame that gap carefully. Adoption is slower than capability. Legal constraints exist. Verification steps slow deployment. Organisational inertia holds the door. They are correct about all of that.
What they cannot fully account for is what is sitting inside that gap. And what is sitting inside it is not reassuring.
The gap is not a comfort
I have spent the last several years building intelligence infrastructure for organisations. The context layer for a bloodstock operation. Orchestration systems for luxury hospitality. Automation pipelines for events businesses that had been running on manual data validation for decades.
In every case, the limiting factor was never the technology. The model could do the work. What was missing was the context: the proprietary knowledge, the operational history, the organisational intelligence that transforms a capable model into something genuinely useful. Once that context was in place, the results were not incremental. They were categorical.
I watched this happen again this week with a group of investors and business leaders who understand what great looks like in commercial terms. They set their own bar. They defined the standard, fed the model real context, and tested it against their own measure of what a high quality business plan looks like. What came back met that standard in ways they described as unbelievable. Not unbelievable for an AI. Unbelievable by the measure they apply to the best work they see.
That is what context does. And that is what most of the Anthropic deployment data does not yet reflect.
The 33% coverage figure represents a world where AI is being used for tasks that are visible and bounded: writing, summarising, generating, assisting. The deeper integration, the kind that actually restructures how work gets done, is still concentrated in the hands of the people who know how to build it.
That is the overhang. Not a technology gap. A comprehension and context gap. And it is closing.
Two different things
Most people who have used AI have used it as a capable tool sitting alongside their work. You bring a question. It gives a useful answer. You bring a document. It produces a reasonable summary. You ask it to help you think through a problem. It does. You conclude you understand what AI can do.
You have seen the shell. You have not seen the engine.
What changes when context is properly in place is not a matter of degree. It is a shift in kind. A model that knows your domain, understands the standard you are working to, has access to the proprietary knowledge that makes your business what it is, and is oriented toward a specific and demanding output does not behave like a smarter search engine. It does not behave like a more capable content assistant. It begins to operate at a level that is genuinely difficult to categorise against human work because the comparison keeps coming out wrong.
Most people reach for the word impressive. That is the wrong word. Impressive is what you say when something exceeds your expectation while remaining in the same category. What happens at the deep end of context-equipped AI reasoning is that the category itself shifts. The output is not impressive work for an AI. It is simply work. Produced at a standard that stands on its own terms.
This is almost entirely absent from the Anthropic deployment data. The 33% coverage figure captures tasks where AI is assisting at the surface: generating, drafting, summarising, searching. It does not yet reflect the organisations that have built the context layer and crossed into something else entirely. Those organisations are outliers today. They will not be outliers for long.
The signal hiding in plain sight
There is one finding in the paper that matters more than the unemployment figures.
Hiring of younger workers into the most exposed occupations has slowed. The paper is careful about this: the effect is modest, barely statistically significant, and there are alternative explanations. The authors flag it and move on.
It does not deserve to be moved past quickly.
The entry-level knowledge worker role has historically been how organisations learn. You hire junior people, they do bounded tasks, they develop expertise, they move up. The organisation grows its own capability. That pipeline is the foundation of how professional knowledge gets transferred across generations.
What I see in the organisations I work with is the same signal at closer range. Senior people expanding their scope with AI while junior headcount stays flat. Roles that would have been filled twelve months ago left open, not because the work has gone away but because one person with the right tools now covers what three people used to do. Nobody announces this. It does not show up in a press release. It shows up in a hiring decision that does not get made.
That is not unemployment. That is the pressure phase. Unemployment is what happens after the pressure releases.
The real friction is cognitive
The Anthropic paper attributes the gap between theoretical capability and observed deployment to adoption friction: legal constraints, verification requirements, the absence of the right software layer.
All true. But the adoption friction I observe most consistently is not legal or technical. It is cognitive. Leaders who have not yet seen what these systems do with real context cannot make the infrastructure investment required to close the gap. They are optimising a mental model that is already out of date.
I was at a sports innovation summit in Düsseldorf this week. A full day of conversations with technology leaders and digital directors: people whose professional identity is built around understanding what is coming next. The debates that broke out when I demonstrated real capability were not hostile. They were urgent. The kind of conversation that happens when intelligent people suddenly feel the ground shifting beneath something they thought they understood. And yet the prevailing mental model across almost every discussion was still the same one I heard two years ago: AI as a generative tool, AI as a feature layer, AI as a way to bolt new capabilities onto existing products.
These were not uninformed people. They were reading the same papers, attending the same summits, and running digital transformation programmes. The gap was not access to information. It was the absence of direct experience with what these systems do when context is properly in place.
That is the most important dimension the Anthropic data cannot capture. You can read about the gap between theoretical capability and observed deployment. You cannot fully understand what is inside that gap until you have seen a model working with genuine organisational intelligence behind it. The comprehension shift that follows is not gradual. It is immediate.
When comprehension catches up, the investment decisions that follow are not incremental. They are urgent.
That urgency, multiplied across the organisations that are currently in the comprehension gap, is what the unemployment figures are not yet measuring. It is what the hiring slowdown for young workers is beginning to reflect. And it is what the gap between theoretical capability and observed deployment in the Anthropic data is holding under pressure.
The overhang is real. It is large. And it does not release gradually.
What to do with this
The investors I spent time with this week gave me a useful frame for what the right first step looks like. They did not ask for a vendor presentation. They brought their own standard, defined what great looked like in their terms, and tested the model against it.
That is the right approach. If you lead an organisation, set your own bar. Define what excellent output looks like in your domain: a business plan, a client strategy, an operational model, a market analysis. Feed the model real context from inside your business. Then judge the result by the same measure you would apply to your best people. The shift in how you think about the next twelve months will be immediate.
If you are building inside the technology space, the opportunity is not in the model layer. The models are a commodity in the making. The opportunity is in context: in proprietary intelligence infrastructure, in the orchestration layer that connects capability to the specific knowledge that makes it useful. The organisations that understand this are not talking about AI features. They are building AI foundations.
If you are a knowledge worker watching this from outside, the most important thing is not to interpret the calm as confirmation that the concern is overstated. The calm is the pressure phase. The people whose roles are not being backfilled, whose entry points are narrowing without announcement, are the earliest signal. The Anthropic data confirms the signal exists. It does not confirm the signal is small.
The window between now and when the overhang releases is the most strategically valuable period in a generation. Not because of what AI will do to the economy. Because of what the people who understand it now will build while everyone else is still forming an opinion.
That window is wider than most people think.
It will not stay that way.
Link to the full research report here
If this landed, subscribe. The next piece goes inside the context layer: what proprietary intelligence infrastructure actually looks like, why it is the real competitive moat of the agentic era, and why most organisations are currently building in the wrong place.
AI strategist and builder. Perplexity Fellow. Former Chief Digital Officer at Art Basel and UEFA. Craig Hepburn helps organisations build useful intelligence for the agentic era.