Anthropic Just Mapped Which Jobs AI Is Actually Replacing — Here's What It Means If You Run a Service Business

Software developers are 75% exposed. Actual usage is a fraction of what's possible. The gap between those two numbers is where your future gets decided.

Anthropic published real usage data from millions of Claude conversations, matched against 800 occupations. Software developers are 75% exposed — but actual usage is a fraction of theoretical capability. The gap is where service businesses either build a moat or get disrupted.

Anthropic published a landmark labour market study today. Not theory. Not predictions. Real usage data from millions of Claude conversations, cross-referenced against 800 occupations. The headline: software developers are the most AI-exposed workers on the planet, with 75% of their tasks already covered by AI.

But here is the number that matters more — actual usage sits at barely a third of what is theoretically possible.

The gap between what AI can do and what businesses are doing with it is enormous. And if you run a service business, that gap is where your future gets decided.

What Anthropic actually measured

The report — Labor Market Impacts of AI: A New Measure and Early Evidence by Maxim Massenkoff and Peter McCrory — introduces something genuinely new. Not another prediction model. An observed measure.

They looked at what people actually use Claude for, matched it against every task in the US Department of Labor's O*NET database (roughly 800 occupations), and built a picture of which jobs are being changed right now. Not in 2030. Now.

The approach combines three things: the O*NET task database, real Claude usage data from the Anthropic Economic Index, and theoretical exposure estimates from earlier research. Crucially, they weighted automated usage (where AI does the task) more heavily than augmentative usage (where AI helps a human do the task).

Previous studies estimated what AI could theoretically do. This one measures what it is actually doing.

The 94% versus 33% gap

The most striking finding compares theoretical AI capability against observed real-world usage across job categories.

Computer and mathematical occupations score 94% on theoretical capability — meaning AI could, in principle, handle nearly all their tasks. But observed coverage? Just 33%.

That is not a rounding error. It is a three-to-one ratio between what is possible and what is happening.

Every other knowledge-work category shows the same pattern. Office and admin roles, business and finance, legal — all sit in the high-theoretical-capability zone. All show actual usage that is a fraction of the potential.

The report frames this directly: as capabilities advance, adoption spreads, and deployment deepens, the observed area will grow to cover the theoretical area.

Translation: the wave has not hit yet. But the water is already rising.

For context, Goldman Sachs estimated that generative AI could affect 300 million jobs globally, with administrative roles at 46% task automation and legal at 44%. McKinsey's research pushed further — 57% of US work hours could be automated with current technology. Anthropic's contribution is not to dispute these numbers. It is to show that reality is lagging theory by a massive margin. The disruption is coming, but it is coming in waves, not all at once.

Who is most exposed

The report ranks the most exposed occupations. Computer programmers top the list at 75% task coverage. Customer service representatives and data entry keyers follow at 67%. Financial analysts, medical record specialists — anything that involves processing information, generating documents, or pattern-matching against data.

At the other end: 30% of workers have zero measurable AI exposure. Their tasks do not appear in Claude usage at sufficient volume to register. These are cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers. Physical work. Hands-on work. The kind of work where you cannot delegate to a language model because the job requires being somewhere, touching something, making a judgment call that depends on physical context.

Construction, agriculture, installation and repair — these industries are not just less exposed. They are effectively immune to the current wave of AI capability.

Where service businesses sit

Compliance consultancies, recruitment firms, training providers, marketing agencies — none of these show up individually in Anthropic's top-ten list. But look at their underlying task composition and they are packed with high-exposure activities: document generation, data analysis, report writing, correspondence, scheduling, research synthesis.

They sit squarely in the business and finance, office and admin, and legal categories — all in the high-exposure zone.

But the exposure is not uniform within these businesses. This is the bit most commentary misses.

The delivery layer is exposed — the methodology layer is not

When I build software for service businesses, I see the same pattern every time. The business has two layers.

There is the methodology — the hard-won expertise, the frameworks, the judgment calls, the thing that makes clients pay premium rates.

And there is the delivery — the reports, the scheduling, the data processing, the compliance checks, the templated outputs that consume 60-80% of billable hours.

AI is eating the delivery layer alive. It is barely touching the methodology layer.

A compliance consultancy's ability to interpret ambiguous regulation and advise a client on the right course of action? AI cannot do that reliably. But generating the compliance report template, populating it with client data, running initial checks against a regulatory framework, flagging anomalies for human review — that is exactly what shows up in Anthropic's Claude usage data.

This maps perfectly to the report's central finding. The 94% theoretical versus 33% observed gap exists because the easy tasks get automated first, and the hard tasks — the ones requiring judgment, context, and expertise — remain stubbornly human.

The businesses that understand this distinction are the ones building moats right now. They are taking their methodology — the thing AI cannot replicate — and productising it into software that uses AI to scale the delivery layer. They are not being replaced by AI. They are using AI to serve ten times more clients with the same team.

The businesses that do not understand this distinction are the ones who will watch a competitor eat their lunch by offering 80% of their output at 20% of their price.

Hiring is already slowing for the most vulnerable workers

The report's most sobering finding is not about unemployment. It is about hiring.

Anthropic found tentative evidence that hiring of workers aged 22-25 has slowed in high-exposure occupations. Job finding rates for young workers entering exposed roles decreased by roughly 14% compared to 2022 levels, while entry into less-exposed jobs remained stable.

This aligns with Stanford research by Brynjolfsson et al., which used ADP payroll data covering millions of US workers and found significant declines in employment for early-career workers in AI-exposed roles since late 2022.

The unemployment rate itself has not moved. Anthropic is clear on this — the average change since ChatGPT's release is small and statistically insignificant. But as they point out, young workers who cannot find jobs in exposed occupations might simply leave the labour force entirely rather than showing up as unemployed. The signal is in the hiring data, not the unemployment data.

What does this mean practically? If you are running a service business and your hiring model depends on bringing in graduates to handle delivery work — writing reports, processing data, doing research — that pipeline is getting complicated. Not because the graduates are not available. Because the work they would be doing is increasingly handleable by AI tools.

This is where I think the smartest move is counterintuitive. Instead of hiring three junior staff to handle delivery, hire one sharp generalist who knows how to use AI tools effectively, and invest the difference into productising your methodology into software.

Trades and physical businesses are not at risk — but they are missing an opportunity

If you run a plumbing business, an electrical contracting firm, or a landscaping company, this report confirms what you probably already sensed: AI is not coming for your core work. Anthropic's data shows near-zero exposure for physical occupations. You cannot send a language model to fix a boiler.

But your work is safe. Your admin is not.

Scheduling, quoting, invoicing, compliance documentation, customer follow-ups — these are all high-exposure tasks that consume hours of your week without generating revenue. They sit in the same exposure zone as office administrator roles.

The opportunity for trades businesses is not to fear AI. It is to automate the admin layer and free up time for the billable work that AI cannot touch.

What the report validates about productising services

I have been saying for a while that the smartest move for service businesses is to turn their services into software. The Anthropic report does not use that language, but the data supports the thesis from multiple angles.

The theory-practice gap creates a window. If actual AI usage is only a third of theoretical capability, there is a massive window right now where human expertise still matters enormously. But that window is closing. The report is explicit: the observed exposure area will grow over time. If you wait until AI can do 90% of your delivery work before you act, you have waited too long.

The valuation implications are significant. Service businesses typically trade at 1-3x revenue. Software businesses trade at 4-8x or higher. A service business doing £500K in revenue might be worth £1M. Productise that same methodology into software and you could be looking at £2.5M-£4M. Same expertise. Different packaging.

The talent economics are shifting. If hiring 22-25 year olds into delivery roles is getting harder — and the data says it is — then the alternative is to encode your delivery process into software and use AI to handle what junior staff used to do. Not to eliminate humans. To change the ratio. One experienced person with AI-powered software can now do what used to take a team of five.

What to do about this — practically, this quarter

If you run a service business, here is what I would take from this report.

Audit your delivery layer. Map every task your team does in a typical week. Separate methodology tasks (client judgment, strategy, complex decisions) from delivery tasks (report generation, data processing, scheduling, templated outputs). The delivery tasks are your AI-exposure surface area.

Productise one workflow. Pick your most repeatable, highest-volume delivery process and build it into software. Not a prototype. A production-ready tool that your team or your clients can use.

Rethink your hiring model. The Anthropic data on 22-25 year old hiring slowdowns is a leading indicator. Build your team around experienced practitioners with AI fluency, not junior staff doing manual delivery work.

Do not panic if you are in trades. Your core work is safe. But productise your quoting, scheduling, and compliance processes while the tools are cheap and the competition is not paying attention.

The report's most important finding is not that AI is coming for jobs. It is that there is a massive gap between what is possible and what is happening — and that gap is where smart service businesses build their advantage. The question is not whether AI will automate your delivery layer. It is whether you will be the one who productises it first, or the one who gets disrupted by someone who did.

Frequently asked questions

How reliable is Anthropic's data given they only measured Claude usage?

It is the best real-world usage data we have right now, but it is one platform. The researchers acknowledge this limitation. That said, Claude is one of the most widely used AI tools in professional settings, and the findings align with independent studies from Stanford, McKinsey, and Goldman Sachs. The directional picture is solid even if the specific percentages might shift with broader data.

Does this mean I should replace my team with AI?

No. The report explicitly shows that most exposed workers are higher-paid, more educated, and more experienced — exactly the people whose judgment AI cannot replicate. The move is not replacement. It is restructuring around the delivery-versus-methodology split. Keep the experts. Automate the routine delivery work.

My service business is not in tech — does this still apply to me?

If your business involves producing reports, analysing data, managing compliance, writing proposals, or processing information, you are in the high-exposure zone regardless of your industry label. The exposure follows the tasks, not the sector.

When will this actually start affecting employment numbers?

Anthropic's data shows no significant unemployment impact yet, but hiring for 22-25 year olds in exposed roles is already slowing. The honest answer: macro-level effects are probably 2-3 years away, but individual businesses and roles are being affected right now. If you wait for the unemployment data to move before you act, you have already lost the initiative.

What is the single most important thing I can do this month?

Map your delivery tasks. Spend two hours listing every repeatable process in your business — every report template, every data pull, every compliance check. That is your exposure surface area, and it is also your productisation opportunity. The businesses that move first on this are the ones that will own their market in three years.

---