Early Data on AI and Jobs: Reading Anthropic's Labor Market Impacts Paper

In March 2026 Anthropic published “Labor market impacts of AI: A new measure and early evidence,” the first serious attempt we have seen to measure what AI is actually doing in the economy rather than what it could do in theory. The contribution that matters is the methodology: a measure called observed exposure that combines theoretical capability with real-world Claude usage, weighting automated workflows at full value and augmentative uses at half. It is the first measure of its kind that anchors the AI-and-work conversation in something other than survey responses or science-fiction extrapolation.

Our reading of the paper is straightforward: the headline finding is reassuring, the second-order findings are not, and the implications for individual, organizational, sectoral, and national planning are urgent in a way the headline obscures. The skaills position has always been that we are at the dawn of a new era of prosperity, but the threshold to cross can be a painful one if individuals, organizations, sectors, and governments do not plan properly, carefully, and promptly. Anthropic's data is the first piece of empirical evidence we have seen that confirms both halves of that sentence.

What the paper actually finds

Three findings carry the analysis.

First, the exposure is real and uneven. Computer programmers show 75% observed coverage of their work tasks. Customer service representatives and data entry keyers are also in the top quartile. At the other end, roughly 30% of workers — cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers — show zero observed coverage. The economy is not facing a single AI transition. It is facing several transitions at very different speeds.

Second, the headline labor-market damage has not arrived. The authors are explicit: “we find no systematic increase in unemployment for highly exposed workers since late 2022.” If you were expecting an across-the-board displacement signal in the macro data, it is not there. Yet.

Third, the signal that has arrived is the one nobody is talking about. The hiring rate for workers aged 22–25 in highly exposed occupations has dropped roughly 14%. Each ten-point increase in observed exposure correlates with a 0.6 percentage-point lower BLS employment-growth projection through 2034. The labor market is not firing exposed workers — it is quietly closing the door to the next generation of them. That is a slower, kinder-looking displacement than the layoff wave the doomers predicted. It is also the displacement that compounds the most over a decade.

The honest read: neither catastrophe nor non-event

Two readings of this paper are wrong, and they are the readings most people will adopt.

The first wrong reading is the relief reading: no mass unemployment, the doomers were wrong, carry on. This misses the youth-hiring signal entirely. A 14% drop in job-finding rates for 22-to-25-year-olds in exposed occupations is not a non-event — it is the leading edge of a generational allocation problem that will only become visible in headline unemployment numbers years from now, after the cohort has already paid the cost. The macro will look fine for a long time before it stops looking fine.

The second wrong reading is the catastrophe reading: 75% coverage of programmer tasks means the work is gone. This collapses the augment / automate distinction the paper takes care to preserve. Most observed Claude usage today is augmentative — the human is still in the loop, still owns the decision, still produces an output the AI alone could not. Exposure is not displacement. It is a measure of how much of the task surface AI now touches. What happens to wages and headcount depends on whether the augmentation produces productivity that the worker captures, the employer captures, or the customer captures — and on whether the displaced labor finds a productive home.

The honest read sits between those two, and it is the read we have been building skaills around since the platform's repositioning: the transition is real, it is uneven, the worst short-term harms fall on a small number of identifiable groups, and the answer is neither resistance nor surrender. The answer is planning — at the individual, organizational, sectoral, national, and regional levels — done properly, carefully, and promptly enough that the bridge across is one people can cross.

What this means at each layer

Our editorial position has been consistent across every opinion piece on this site: AI transformation is a coordinated problem with at least five layers — individual, organization, sector, nation, region — and a serious response engages all of them. Anthropic's data sharpens what each layer needs to do right now.

At the individual layer, the paper validates a planning instinct that was previously hard to defend with data. If your occupation sits in the top exposure quartile — and software engineers, customer service representatives, data analysts, copywriters, junior consultants, paralegals, and a long list of others all do — you cannot wait. Your planning horizon is shorter than your peers in unexposed occupations, your career path needs more optionality built in, and the skills you compound now should weigh transferability heavily. If you are 22 to 25 and in an exposed occupation, you have already lost a measurable amount of the entry-level opportunity surface; the productive response is to deepen and diversify the moves that get you to the second job, not the first.

At the organization layer, the paper validates a planning instinct that was previously easy to deprioritize. The exposure heatmap inside your own headcount is real, it is asymmetric, and it is now describable. You can build a capability heatmap that matches Anthropic's exposure curve to your actual functions, and you can decide function by function whether the right move is augmentation, redesign, redeployment, or hiring-pause-with-investment-in-reskilling. The CHROs who treat this as one homogeneous “AI strategy” will under-serve both the exposed and the unexposed sides of their workforce. The ones who treat it as a portfolio of differentiated transitions will not.

At the sector layer, the paper's uneven distribution is the entire point. The customer service sector is on a different clock from the hospitality sector, which is on a different clock from the construction trades, which are on a different clock from software engineering. Sector-level coordination — industry associations, employer councils, labor groups — is not a nice-to-have here. It is the only layer at which the cost of the transition can be socialized across firms within a sector and shared portably across worker tenures.

At the national and regional layers, the youth-hiring signal is the urgent one. A 14% drop in job-finding rates for early-career workers in exposed occupations is not something the labor market self-corrects out of within one electoral cycle. It is the signature of a problem that calls for active labor-market policy: apprenticeship subsidies, portable benefits, reskilling vouchers, public-sector training infrastructure, and a serious conversation about which credentials still mean what they used to. Governments that wait for the macro unemployment number to move will be reacting to a problem that, by then, will already have hardened.

Three further lenses the paper does not address — and that any honest reader has to add

Anthropic's data is U.S.-centric, capability-centric, and silent on the question of who captures the gains. Three lenses fill those gaps and change the picture once you apply them.

The developing-world dual effect

AI hits developing economies on two clocks running in opposite directions. The braking clock is structural: heavy reliance on manual, low-skill, and low-wage labor; thinner capital stocks; weaker enterprise IT; political resistance from labor unions and labor ministries; regulatory friction; and the simple economic fact that human labor will remain cheaper than AI for a long list of tasks for many years yet. That clock argues for a slower, gentler exposure curve than the Anthropic numbers imply for similar occupations in higher-cost economies.

The accelerating clock is global demand. The sectors developing economies have built their growth on — call centers, business-process outsourcing, manufacturing in free zones, IT outsourcing — are precisely the sectors where the demand-side decisions are made in headquarters thousands of kilometers away. A North American or European company that decides to automate the work that used to be sent offshore does not need any cooperation from the labor ministry in the receiving country. The first major employment shocks from AI in the developing world will land on outsourcing-dependent sectors well before they land on the domestically consumed labor market. Two clocks, opposite directions, same economy. Reading Anthropic's paper without that asymmetry will lead Latin American, African, and South Asian readers to false comfort on one side and false catastrophism on the other.

Humans are still the pivotal agent — and the values triage matters

AI does not transform an economy on its own. Humans transform economies with AI. That is good news, and it is hard news, because it means the bridge across the transition has to be deliberately built — by people, at every layer — and the cost of not building it cannot be blamed on the technology. The compensating measures that get a country, an organization, or an individual through with limited casualties are not optional add-ons to the transformation; they are part of the transformation itself.

There is a cultural counterpart to the economic transition that the policy literature consistently underplays. Every transition forces a values triage. Some values, traditions, and practices we will drop without regret — the ones that depended on artificial scarcity of information, gatekept access, or labor we now find inhumane. Some we will preserve and carry over — craft, mentorship, the dignity of contribution, the obligation to truth. Some we will nurture and double down on — judgment, collaboration, the human capacity to hold ambiguity that AI systems still cannot. Organizations and individuals who do this triage explicitly come through the transition with their identity intact. Those who let it happen by default lose things they did not realize they wanted to keep.

Productivity — and who captures its gains — cannot stay out of the labor conversation

In the AI era it is no longer honest to talk about work, wages, employment, or employability without addressing productivity and the distribution of productivity gains. The Anthropic paper measures task coverage; it does not, and cannot from outside the firm, measure how much more output the augmented worker is producing per hour, or where the additional value is going. Both questions matter, and the second one matters more.

The structural tension underneath the labor conversation is rarely named out loud. Businesses do not exist to create jobs. They exist to create value for consumers, in the form of goods and services. From the firm's perspective, employment is a resource used to produce and distribute that value — one input among several, optimized against cost and capability. From the perspective of workers, society, and governments, employment is something entirely different: a distribution mechanism — historically the primary way the gains of economic activity get allocated across the population, alongside being a source of meaning, identity, social cohesion, and rights.

For most of industrial history these two views could coexist quietly, because productivity gains required more labor to capture them. The firm's resource-optimization and society's distribution-mechanism happened to point in the same direction: hire more people, produce more output, distribute more wages. AI breaks that coincidence. When productivity can step-change without proportionally more labor — and the Anthropic data is the first quantitative confirmation that this is exactly what is starting to happen — the firm view and the society view diverge in a way they have not for a century.

That divergence is the political-economy core of the transition. The same productivity gain can show up as higher wages and shorter hours for the worker, as wider margins for the employer, as lower prices for the customer, or as rents accruing to the platform layer that owns the model. Every realistic transformation plan — individual, organizational, sectoral, national — implicitly takes a position on which of those is the target. The honest plans make the position explicit and design the policy levers (collective bargaining, profit-sharing, public-option training, antitrust on the platform layer, portable benefits, sometimes UBI-adjacent transfers) to land the distribution where they say they want it to land. The plans that decline to take a position end up with the default outcome, which historically has not been the worker.

The worker's position in the production chain is moving

There is a deeper structural shift inside the productivity question that the distribution framing alone does not capture. Because AI is a productivity-enhancing technology, it changes the worker's relationship to the production process itself — not only how the gains are split, but where the worker stands in the value chain.

The worker moves from being the producer of the output to being the operator of a productivity multiplier that produces the output. What counts as “skill” shifts from execution to orchestration, judgment, and supervision of AI output. The marginal product becomes increasingly attributable to the tool rather than to the labor, which weakens individual wage-bargaining power — your raise no longer follows from being twice as fast as the next candidate when both of you have access to the same multiplier — and reshapes what collective bargaining is even bargaining over. Unions built around protecting a specific task have to rebuild themselves around protecting access, voice, and value share in a chain where the tasks themselves are dissolving.

The Anthropic exposure data implicitly maps where this relocation is happening fastest. The 75% coverage of programmer tasks is not just an automation statistic; it is a description of a profession whose practitioners are mid-migration from writing the code to directing, reviewing, and integrating code that an AI agent now writes. Some programmers move up the chain — system design, architectural judgment, supervision of agent fleets — and capture far more of the new value than they did before. Others find their old position dissolving without an obvious next position in the chain. Both outcomes are inside the same headline number.

The transition policies and personal plans that work address this structural relocation directly. They create real paths up the chain for the people whose old position is disappearing — not just retraining for an adjacent role at the same level, which is what most “reskilling” still defaults to. Retraining a junior data-entry worker to be a slightly more senior data-entry worker in a function that will be automated next quarter is not a transition plan. Retraining the same worker to be the human who supervises, audits, and corrects the AI that now does the entry — and pricing that role at what supervision is actually worth — is.

Where skaills sits in this picture

We did not build skaills to take a position on whether the AI transition is good or bad. We built it because, having watched the data emerge across two years, we became convinced that the only response that makes sense is plan, at every layer, with the best instruments you can get your hands on. The platform is organized around that conviction in two layers: a Mechanical Layer that absorbs the repetitive scoring, parsing, and paperwork of modern HR so humans can stop doing it, and a Transformative Layer that gives individuals and organizations the instruments to plan the transition deliberately.

Anthropic's paper points at four of those instruments by name.

The Anthropic paper closes with an admission that we take seriously: detecting headline labor-market damage would require roughly a one-percentage-point differential rise in unemployment among exposed workers. We are nowhere near that signal yet. That is not a reason to wait. It is a reason to invest now in the instruments that will let the people who do face displacement — the early-career workers already showing the signal, the high-exposure occupations that have not yet shown it, the sectors that will move next — have somewhere coherent to go.

The dawn of a new era of prosperity and the painful threshold across to it are not separate stories. They are the same story told at two scales. The Anthropic paper is the first quantitative confirmation we have that both are true. The work, at every layer, is to make sure the bridge gets built before anyone has to fall off the edge of it.

Related reading

• Anthropic, “Labor market impacts of AI: A new measure and early evidence” (March 2026) — the Anthropic paper this commentary engages with directly.
• Who Wins, Who Pivots, Who Loses: Honest Categories for the AI Jobs Transition — our framing piece on the three categories of jobs in the AI transition (created / transformed / displaced). This commentary tests that framing against the first round of real usage data.
• Four CHRO Roles, Four Operational Layers — the organization-level counterpart: what a CHRO actually has to build so the bridge across the AI transition is one their workforce can cross.

Expert Perspectives