AI and the Labor Market: What the Data Actually Shows
Every few months a new report arrives claiming to know exactly how many jobs AI will eliminate, and by when. The numbers vary wildly. The tone rarely does. What has been missing, until now, is a rigorous, data-grounded framework that separates what AI is theoretically capable of from what it is actually doing in the workplace today.
In March 2026, Anthropic researchers Maxim Massenkoff and Peter McCrory published "Labor Market Impacts of AI: A New Measure and Early Evidence." The paper introduces a new way of measuring AI's real-world occupational impact and tests it against US employment data. The results are worth reading carefully if you lead a business, manage a technical team, or are thinking about where to invest your hiring capacity over the next few years.
The gap between theory and reality
Previous research on AI exposure has focused on theoretical capability: can a large language model, in principle, perform this task? That framing is useful but incomplete. Just because a model can theoretically draft a patient intake summary or process a refill request does not mean it is doing so at scale in clinical settings today.
The researchers formalized this with a new metric they call Observed Exposure. It combines two inputs: task-level theoretical feasibility scores and actual usage data from the Anthropic Economic Index, which tracks how Claude is used in professional and automated workflows.
The core insight: AI is far from reaching its theoretical capability. Tasks rated as fully feasible for an LLM alone account for 68% of observed Claude usage, yet actual professional coverage of most occupational categories remains a fraction of what the theory suggests is possible.
For example, the theoretical model marks "Authorize drug refills and provide prescription information to pharmacies" as fully feasible for an LLM. In practice, the researchers found no observed usage of Claude performing that task at scale. Legal requirements, verification steps, software integration hurdles, and institutional inertia all slow the translation of theoretical capability into deployed automation.
Which roles are most exposed right now
Among the roughly 800 occupations analyzed, the most exposed under the Observed Exposure measure are concentrated in knowledge work and information processing. The top ten occupations by coverage include roles where AI is already handling core tasks in automated, work-related workflows rather than just assisting humans intermittently.
The pattern aligns with what most practitioners have observed: AI is penetrating structured, text-heavy, and logic-driven work faster than anything requiring physical presence, tactile judgment, or complex interpersonal interaction. Bartenders and motorcycle mechanics are essentially unaffected. Computer programmers are not.
Who is most exposed, and what that means
The demographic profile of highly exposed workers challenges some assumptions. Using pre-ChatGPT baseline data from the Current Population Survey, the researchers found that workers in the top quartile of observed exposure are 16 percentage points more likely to be female, earn 47% more on average, and are significantly more educated. Workers with graduate degrees are almost four times more represented in the highly exposed group than in the unexposed group.
This matters for how businesses think about AI risk. The jobs most exposed to AI are not low-wage, low-skill positions. They are often the anchors of a knowledge-economy workforce: well-compensated, credentialed, and historically considered secure. The displacement risk, if it materializes, will be felt differently than past waves of automation.
What the employment data shows so far
The headline finding is that, as of the data available at time of publication, there is no statistically significant increase in unemployment for workers in the most exposed occupations since late 2022. The difference-in-differences analysis comparing high-exposure to low-exposure workers shows a small uptick that is not meaningfully distinguishable from zero.
The researchers are careful not to interpret this as evidence that AI has no labor market effect. They point out that past disruptions, including the China trade shock and the adoption of industrial robots, also showed delayed and ambiguous effects in aggregate employment data. Their framework is deliberately designed to detect signal early, before effects become unmistakable.
The signal worth watching: young workers entering the labor market
The more pointed finding concerns workers aged 22 to 25 trying to enter exposed occupations. Using the panel structure of the Current Population Survey to track new job starts month by month, the researchers found that hiring into high-exposure occupations has declined by roughly half a percentage point per month since 2024. That translates to an approximately 14% drop in the job-finding rate for young workers in exposed roles compared to the 2022 baseline. The effect is not present for workers over 25.
This echoes similar findings from a separate study by Brynjolfsson and colleagues, which detected a 6 to 16% fall in employment in exposed occupations among workers aged 22 to 25. The researchers interpret this as a hiring slowdown rather than an increase in layoffs: companies are simply bringing in fewer entry-level workers in roles where AI tools are reducing the volume of work that previously required a human.
The practical implication: AI is not yet eliminating roles at scale. But it may already be compressing the traditional entry path into certain technical and knowledge-work careers. If that trend continues, the downstream effects on the talent pipeline could become significant within a few years.
What this means for software teams and tech leaders
For companies building and managing software development teams, this research offers a few concrete takeaways.
The productivity gap is real, and it is widening. Computer programmers sitting at 75% observed exposure are not at risk of mass layoffs today. But they are working with AI tools at a level that is changing how work gets done. Teams that have integrated AI coding assistants, automated code review, and AI-driven testing pipelines are completing cycles faster than those that have not.
Junior talent pipelines need rethinking. The data on young workers is worth taking seriously. If AI is absorbing entry-level programming and data tasks, the traditional model of hiring junior developers to grow them into mid-level contributors over three to five years faces structural pressure. Companies will need intentional strategies for onboarding, mentorship, and skill development that account for a world where many of the repetitive foundational tasks are no longer done by humans.
Observed exposure will keep expanding. The researchers frame their measure as a tracking tool, not a snapshot. As AI capabilities improve and adoption deepens, the red area of actual coverage will continue closing in on the blue area of theoretical capability. Organizations that plan around today's coverage levels rather than tomorrow's trajectory risk being caught off guard.
The skills that remain hard to automate are becoming more valuable. Technical architecture decisions, client relationship management, cross-functional problem definition, and the judgment to know when a model's output is wrong are not showing up in high observed-exposure categories. Engineers and team leads who develop these capacities alongside AI proficiency will be harder to replace than those who treat AI as a background tool.
The broader picture
This research does not support the most extreme narratives, either the "AI will automate everything within five years" claims or the "nothing has really changed" dismissals. What it shows is a careful, evidence-based middle ground: AI is meaningfully present in professional workflows, it is affecting how labor demand is organized in exposed sectors, and the effects are becoming detectable even if they are not yet large enough to register clearly in aggregate unemployment statistics.
The most honest reading is that we are in an early and accelerating transition. The frameworks for understanding that transition are still being built. This paper is one of the better contributions to that effort, because it is grounded in real usage data and designed to be updated as new evidence accumulates.
For business leaders, the implication is not to panic and it is not to wait. It is to build the organizational capacity to adapt, track how AI is changing the actual work your teams do, and invest in the human capabilities that compound alongside AI rather than competing with it.
Our perspective at Alluxi
We have been building software teams at the intersection of US product vision and Latin American engineering talent for over a decade. What we are seeing in our own work tracks with this research. AI is changing the composition of what we build and how fast we build it, but the judgment, communication, and architectural thinking that experienced engineers bring are, if anything, more important now than they were five years ago.
The companies navigating this transition best are not the ones cutting headcount to chase short-term savings. They are the ones investing in teams that know how to work with AI tools effectively, and partnering with development shops that have already done that integration work. That is the conversation we are having with our clients, and we think the data from this paper supports why it matters.
This post is based on "Labor Market Impacts of AI: A New Measure and Early Evidence" by Maxim Massenkoff and Peter McCrory, published March 5, 2026 as part of the Anthropic Economic Index. Occupation-level coverage data is available in the dataset linked from the original paper. All statistics cited here are drawn directly from the published report.
