The Jobs AI Is Already Replacing (According to New Data)

There’s a version of the AI-and-jobs conversation that has been running for about a decade now, and it goes roughly like this: AI will eventually automate many tasks, the transition will take time, workers will adapt as they always have, and the net effect on employment is genuinely uncertain. This version is sober, balanced, and calibrated. It is also, at this point, missing something important.

The theoretical debate is over. The empirical one has begun.

Anthropic recently published its Economic Index, a systematic analysis of millions of real interactions with Claude across the full range of people who use it. The methodology is straightforward: take actual AI usage data, map the tasks being performed against the O*NET occupational database, and ask a simple question. Which types of paid work are people already offloading to AI?

The answer is not what most people expected, including many economists who spent the last decade confidently predicting that knowledge work was the safe harbor.

The tasks that are already going

The Anthropic data shows AI usage concentrated heavily in a specific cluster of work types. Not physical work. Not the trades. Not the roles everyone spent years assuming would automate first. The highest-volume categories are writing, analysis, coding, and information processing. The cognitive middle layer of the modern office economy.

Writing and content production is the largest single category. Articles, marketing copy, emails, summaries, documentation, reports. Work that used to require a person to sit down and produce it now frequently involves a person reviewing and editing what AI produced. Copywriters, content writers, marketing assistants, documentation specialists: the demand signal for these roles is weakening in ways that show up in hiring data.

Data analysis and reporting is the second major cluster. Summarizing data sets, identifying trends, producing research summaries, writing the narrative layer over numbers. Junior analysts, research assistants, business analysts whose primary output is synthesized information: these roles are not disappearing overnight, but the headcount required to produce a given volume of analytical output is shrinking.

Programming follows a specific pattern worth understanding precisely. AI is not replacing senior engineers making architectural decisions. It is handling code generation from clear specifications, debugging, refactoring, documentation, and explaining what existing code does. The effect concentrates at the entry level.

Administrative and coordination work rounds out the picture. Meeting notes, email drafting, document preparation, information lookup, scheduling, first-pass triage. The work that filled administrative assistant and office coordinator roles is now substantially automatable.

The non-obvious finding

The Anthropic research makes a point worth sitting with: AI is most commonly used not to replace entire jobs but to automate portions of jobs. This sounds reassuring until you follow the logic to its conclusion.

If a junior analyst spends 70% of their time on tasks that AI now handles, the demand for junior analysts doesn’t drop to zero. It drops to roughly 30% of what it was. The remaining work, the judgment, the client relationships, the problem framing, still requires a human. There just aren’t enough hours of that remaining work to justify the same headcount.

Organizations don’t fire everyone and replace them with AI. They hire fewer replacements when people leave. They restructure teams around a smaller human core. The change is gradual enough that it doesn’t trigger dramatic headlines. It shows up in hiring freezes for entry-level roles, in teams that used to have six people running well with four, in a job market where junior roles are harder to find.

The danger is that this process is invisible to the people it’s happening to until it has already happened.

The trades are laughing. They shouldn’t be.

Every time a new wave of concern about AI and jobs circulates, there’s a predictable response from workers in physical trades. Plumbers, electricians, HVAC technicians, farmers, construction workers. The argument is intuitive: AI can write your emails but it can’t snake your drain. Physical work is safe because it requires a body in a space.

This is true right now. It is not a permanent condition.

The trades face a different automation timeline, not immunity from automation. Robotics in physical spaces is advancing along a separate curve from language AI. Automated systems for warehouse picking, construction site surveying, agricultural monitoring, and pipe inspection are already in commercial deployment.

Agriculture is particularly exposed over a medium-term horizon. Precision agriculture systems already automate planting, irrigation, and monitoring decisions. Autonomous harvesting machinery is in field trials for multiple crop types.

Electrical and plumbing work is more protected because it involves unstructured physical environments and genuine problem-solving under uncertainty. But the diagnostic layer—understanding what’s wrong before the physical repair happens—is increasingly addressable by AI systems that can analyze sensor data, building schematics, and fault patterns. The skill requirement doesn’t disappear. It shifts.

Construction faces robotics from the repetitive end (bricklaying, rebar placement, concrete finishing) and AI from the management and planning end simultaneously. The physical craft at the skilled center is protected. The semi-skilled labor at the edges is not.

The timeline for physical work automation is longer than for cognitive work. The answer is ten years, not fifty. And the workers who are currently most protected are the ones with the least time and probably the least institutional support to prepare for the transition when it accelerates. (For those considering career changes at 40 or at 50, understanding this timeline is essential.)

The new structure of the labor market

What AI creates, across both cognitive and physical domains, is a new productivity model that changes the relationship between headcount and output.

The old model: one specialist produces one unit of work. The new model: one specialist with AI or robotic assistance produces several specialists’ worth of output. The headcount required to produce a given volume of work decreases. The value of each remaining worker increases. The middle tier faces the most pressure because AI assistance raises the floor while the ceiling on truly exceptional work remains human.

The practical consequences compound:

Entry-level roles thin out because the output they used to provide is being handled by AI tools that senior workers operate directly. The pipeline that developed people from junior to senior runs slower. Compensation at the top increases as leverage per person grows. The middle tightens as work concentrates upward.

This pattern has already run its course in manufacturing over the previous two decades. It is now running in knowledge work. It will run in physical trades over the decade after that.

The question that replaces “what should I do with my career?”

There used to be a relatively stable way to think about career decisions. Pick a field with good prospects. Develop expertise. Advance. The field might change, but the change was slow enough that expertise accumulated over years remained valuable for years after it was built.

That calculus has changed. The new question is not “what profession should I pursue?” It’s “what skills remain valuable as the task composition of every role shifts under pressure from AI, and which adjacent occupations can I move to if the one I’m currently in contracts faster than I expected?”

PathScorer approaches this by combining O*NET task composition data with AI adoption curves to produce exposure scores and three-year forecasts for each occupation. When you run your profile through the system, you see not just which roles your skills match but how each of those roles is projected to hold up as AI capabilities develop. The same skills might qualify you for three different roles, one of which sits on a steep automation trajectory, one of which is stable, and one of which is growing. The salary difference between those options might be modest today. The career difference over ten years is not.

The biggest risk in the AI labor market transition is not that jobs disappear overnight. It’s that people spend years in roles where the underlying demand for their task output is quietly declining, without the information to see it coming or the map to navigate somewhere better positioned.

The data on what AI is already doing exists. The question is whether you look at it before or after you need to.