In 1987, Robert Solow famously remarked that the computer age was visible everywhere except in the productivity statistics (Solow 1987). Nearly four decades later, the rapid diffusion of generative artificial intelligence (AI) raises related questions about whether official statistics adequately capture the productivity effects of a new general-purpose technology. Micro-level evidence points to substantial efficiency gains in a wide range of cognitive and service-oriented tasks: experimental and field studies report productivity improvements of 14 to 56 percent across customer support, professional writing, software development, consulting, and legal analysis (Brynjolfsson, Li, and Raymond 2025; Noy and Zhang 2023; Peng and others 2023; Dell’Acqua and others 2023; Choi, Monahan, and Schwarcz 2024). Recent aggregate data are also more encouraging than they appeared a year ago: average labor productivity growth across the OECD doubled to 1.2 percent in 2024, with the United States standing out at +2.2 percent, while the euro area remained close to zero (OECD 2026a). The OECD itself interprets these developments as possibly providing “early, tentative signals consistent with a productivity-enhancing role of AI” (OECD 2026a, p. 11). Even so, the magnitude of measured aggregate gains remains modest relative to the emerging micro evidence, and several structural features of the statistical framework continue to limit how much of the underlying improvement can be captured.
This paper argues that two features of the System of National Accounts (SNA) constrain the measurement of productivity gains from generative AI. First, a significant share of the sectors most exposed to AI—particularly public administration, education, and health—rely on non-market output measurement, where output is valued using a sum-of-costs approach (United Nations and others 2009; United Nations and others 2025). Under this framework, output is by construction closely tied to inputs, implying that measured productivity growth is constrained even when underlying efficiency improves (Atkinson 2005; Schreyer 2010). Second, in market-based service sectors, the absence of adequate quality adjustment in price indices can result in improvements in service quality—such as faster delivery, greater accuracy, or expanded functionality—being insufficiently captured in real output measures (Triplett and Bosworth 2004; Byrne, Fernald, and Reinsdorf 2016; Syverson 2017). These features do not preclude the measurement of AI-driven gains entirely—and the 2024 data suggest that some gains are now being picked up in countries where AI investment has been concentrated—but they imply that the share of true productivity improvement reflected in official statistics is likely to remain incomplete, particularly in service-heavy economies.
The paper contributes to the literature by (i) linking the sectoral incidence of generative AI to known measurement challenges in the national accounts, drawing on recent task-based exposure studies (Eloundou and others 2024; Cazzaniga and others 2024; OECD 2026b, Box 1.1); (ii) providing an accounting framework that clarifies how non-market output measures and unadjusted service prices dampen measured productivity growth; (iii) assessing whether the inclusion of a return to capital in non-market output—as recommended in the 2025 SNA—can mitigate this gap; and (iv) outlining potential avenues for improvement, including outcome-based volume indicators, enhanced quality adjustment in service price indices, and exogenous multifactor productivity adjustments applied through the deflator. The findings have important implications for policy: if AI-related productivity gains are even partially underestimated, assessments of economic performance and potential output may be biased downward (Brynjolfsson, Rock, and Syverson 2021). Broader welfare gains may also be missed unless supplemented by measures outside the core national accounts.
