'Apollo: AI Lowers Wages But Doesn't Cut Jobs'

 Analysis of actual Claude usage data shows workers in AI-exposed occupations are experiencing slower wage growth, while employment levels in these occupations remain unchanged, suggesting companies are capturing AI productivity gains through wage compression rather than workforce reduction.

This paper was written using a difference-in-differences methodology with occupation and year fixed effects across 321 matched occupations from 2015 to today. The paper is available here.

We examine the wage and employment effects of AI adoption across U.S. occupations using observed usage data from the Anthropic Economic Index rather than the theoretical exposure measures that dominate prior work. Using a difference-in-differences design with occupation and year fixed effects across 321 matched occupations from 2015 to 2025, we find that high-exposure occupations experience a 6.7% decline in real wage growth post-2023 with no detectable employment effects, suggesting firms are capturing AI productivity gains through wage compression rather than workforce reduction. The effect is concentrated among the lowest earners: service workers face a 24.3% decline and the bottom wage quartile a 10.7% decline, while top earners show no significant effect. Today, 5.8 million workers are affected, but as AI adoption deepens across corporate America, this figure is likely to grow substantially, with significant implications for income inequality and labor market policy in the years ahead.

Since the release of ChatGPT, researchers, firms, and policymakers have tried to estimate how AI will affect jobs and wages. Early studies projected substantial labor market disruption, but nearly all relied on theoretical AI exposure measures rather than observed AI adoption. The first generation of AI exposure research asked which occupations could be affected by AI using O*NET task descriptions and expert assessments.

AI adoption is accelerating across U.S. firms, yet the distributional consequences for wages and employment remain unknown. Newly available datasets from Anthropic, Microsoft, and Ramp make it possible to measure where AI is actually being used. This shift from potential exposure to realized adoption makes it possible to ask a fundamentally different question: what is the effect of AI adoption once workers are actually using AI? Specifically, we investigate whether high-exposure workers (defined as occupations having an Anthropic Economic Index score ≥0.5) experience significant wage and employment changes compared to low-exposure workers after the release of ChatGPT in late 2022. This paper makes three contributions to the emerging literature on AI and labor market outcomes. First, we use observed AI usage data, i.e. actual Claude interaction logs, rather than the constructed or theoretical exposure scores that dominate prior work (Felten et al., 2021; Webb, 2019; Brynjolfsson, Mitchell, and Rock, 2018; Eisfeldt et al., 2023). As illustrated in Figure 1, the existing literature is concentrated in the expert/model quadrant; this paper is among the first to estimate causal labor market effects from the actual usage/realized quadrant, capturing adoption as it is happening rather than as it is projected. Second, it provides early causal evidence via a difference-in-differences design with occupation and year fixed effects of wage compression in high-exposure occupations post-2023, at the moment adoption is visibly scaling. Third, it produces a conservative lower-bound estimate of aggregate labor income loss ($28 billion annually across 5.8 million workers), grounding an otherwise theoretical debate in a concrete macroeconomic magnitude.

Some 5.8 million U.S. workers, roughly 3.7% of the labor force, are currently employed in high-exposure occupations and bearing some wage costs of AI adoption. This number forms a lower bound for the scope of AI's labor market impact, as it excludes workers in complementary occupations who may also be affected, as well as any employment growth within exposed occupations themselves. Looking ahead, the BLS projects the high-exposure workforce will grow to 5.9 million by 2032, meaning the population of affected workers will expand even as wages within it continue to compress. As AI adoption deepens across corporate America, the true number of workers feeling these effects could grow substantially beyond what current exposure measures capture.

https://www.apollo.com/content/dam/apolloaem/pdf/daily-spark/2026/jul/30/Whitepaper-Impact%20of%20AI%20on%20U.S.%20Labor%20Market-2026-R2%201.pdf