Anthropic has published what may be the most detailed map yet of AI's real penetration into the US labor market, and the findings sit somewhere between reassuring and alarming. The paper, "Labor market impacts of AI: A new measure and early evidence," introduces a new metric called "observed exposure" that compares what AI tools are theoretically capable of doing with what workers are actually using them for. The gap between those two numbers is wide for now. But the direction of travel is clear.
What the Data Shows
The research, authored by Anthropic economists Maxim Massenkoff and Peter McCrory, draws on real-world professional usage data from Claude to quantify AI's actual footprint in the labor market. The headline finding: AI can theoretically handle 94% of tasks performed by computer and mathematics workers, but actual observed usage accounts for roughly 33% of those tasks. In office and administrative support roles, theoretical coverage sits at 90%, while real adoption remains a fraction of that.
The authors name the worst-case outcome directly: a "Great Recession for white-collar workers." During the 2007-2009 financial crisis, US unemployment doubled from 5% to 10%. A comparable doubling in the top quartile of AI-exposed occupations, from 3% to 6%, would register clearly in their framework, they write. It hasn't happened yet. But it could.
The occupations sitting at the top of the exposure list aren't the ones most people expected. Computer programmers, customer service representatives, and data entry keyers face the highest observed exposure. The least exposed jobs are those requiring physical presence: cooks, mechanics, construction workers. And the profile of the most-exposed workers is striking: they are 16 percentage points more likely to be female, earn 47% more than the least-exposed group, and are nearly four times as likely to hold a graduate degree. This isn't a story about low-wage automation. It's a story about the professional class.
Young workers are already feeling it. A separate study cited in the Anthropic paper found a 16% fall in employment in AI-exposed occupations among workers aged 22 to 25, driven primarily by a slowdown in hiring rather than outright layoffs. Job search rates in the most vulnerable professions have dropped 14% since the ChatGPT era began. And per US Bureau of Labor Statistics projections, occupations with higher AI exposure are expected to grow more slowly through 2034. The BLS data didn't start with AI in mind, but it lines up with Anthropic's independent findings regardless.
The researchers attribute the gap between theoretical and observed adoption to legal constraints, model limitations, the continued need for human oversight, and the additional software infrastructure required to deploy AI reliably at scale. That lag isn't permanent.
What This Means
Anthropic CEO Dario Amodei said last year that AI could disrupt half of all entry-level white-collar work. Microsoft's AI chief Mustafa Suleyman put a tighter timeline on it, estimating most professional work will be replaced within a year to 18 months. Both forecasts once sounded extreme. This research makes them harder to dismiss.
What the Anthropic paper shows, more than anything else, is that companies are in a transition window. Adoption is real but incomplete. The jobs most exposed to displacement are also the ones currently in highest demand in AI-adjacent functions: software engineering, data science, financial modeling, legal research. As AI handles more of the routine within those roles, what companies actually need to hire for shifts. The demand isn't disappearing; it's moving.
That shift is reshaping where companies look for talent, and how fast they need to act. In the US, the domestic market for qualified AI and machine learning engineers is effectively locked up. There are an estimated 300,000 qualified ML engineers against more than 1 million open positions. Average time-to-hire locally runs 4-6 months. That constraint doesn't ease by waiting.
India's position in this picture is structural, not incidental. NASSCOM projects AI-related job demand in India will cross 1 million roles by 2026, and India already ranks third globally in AI talent per the Stanford HAI Global AI Vibrancy Tool 2025. Indian GCCs now employ over 126,600 AI professionals working within Fortune 500 operations alone, with AI talent concentration having grown 252% between 2016 and 2024, now sitting 2.5x above the global average, per the Wisemonk India Investment Intelligence 2026 report. More than 2 million IT professionals in India have been upskilled in AI, including 300,000 in advanced AI skills.
This isn't an abstract data point for companies trying to hire AI engineers right now. As Wisemonk's analysis of the global generative AI talent market shows, Bangalore alone hosts one of the world's two largest AI talent pools, and India's pipeline is expanding faster than any comparable market. The Wisemonk India IT Services Analyst Report 2026 puts India's IT/BPM sector at $315 billion in FY2026, employing 5.95 million professionals, with $250 billion-plus in fresh AI infrastructure committed at the India AI Impact Summit in February 2026 alone. For companies that can't compete in a domestic AI hiring market compressed by demand and constrained by supply, India isn't a fallback. It's the primary option.
For companies thinking about how to access that talent without entity setup overhead, the employer of record model has become the default entry point: compliant onboarding in 48 hours, no subsidiary required, with the option to scale into a full GCC structure as team sizes justify it. Wisemonk's EOR and AI developer hiring capabilities are built specifically for this transition, helping US companies access India's engineering depth in weeks rather than months.
What to Watch Next
The Anthropic paper is explicit that its framework is designed as an early warning system: it can detect a white-collar recession if one starts, and it can track adoption rates quarter by quarter. That means the next few reporting cycles matter more than they might otherwise. If observed exposure for the top-quartile occupations starts accelerating, from 33% toward 50%, say, in computer and math roles, the gap between "capability" and "adoption" will stop being a buffer and start being a countdown.
Watch for three specific signals. First, whether young worker hiring rates in high-exposure occupations recover or continue declining in the next two quarters of payroll data. Second, whether corporate layoffs in legal, finance, and software roles accelerate in tandem with published AI adoption expansions. Third, whether federal or state-level policy moves in the US begin treating AI displacement as a labor market risk requiring a policy response, not just a productivity story.
The Anthropic paper closes with a note worth reading carefully: the impact of this technology will be shaped not just by capabilities as they advance, but by the choices companies and policymakers make. That framing matters. Workforce exposure is not the same as workforce destruction, and the firms best positioned are those that treat the transition window as exactly that, a window, not a wall.
The honest read of this research is that the worst-case scenario is possible but not yet inevitable. The gap between AI's theoretical reach and its actual footprint is real. But that gap has been narrowing since 2023, and the pace of adoption in enterprise settings is the one variable that makes the difference between disruption and displacement. Companies waiting for certainty before adjusting their hiring strategies may find the window has closed before they walked through it.