Will Artificial Intelligence Increase Productivity Across the U.S. Economy?

Filippucci, Gal, and Schief (2026) examine how much generative artificial intelligence is likely to increase productivity across the U.S. economy and whether uneven gains across industries will reduce overall growth. They combine data on AI task exposure, industry employment patterns, AI adoption rates, and U.S. input-output tables covering 65 industries. They estimate that AI could increase aggregate total factor productivity growth by about 0.87 percentage points annually over the decade following ChatGPT’s introduction. Sectoral gains vary widely, from roughly 0.1 percentage points per year in manual-task industries to 2.8 percentage points in knowledge-intensive sectors such as information services, finance, and professional services. They find that structural economic changes only modestly reduce these aggregate gains.

Why This Article Was Selected for The Policy Scientist

Filippucci, Gal, and Schief contribute to one of the most consequential policy and economic questions currently facing governments, firms, and educational institutions: whether artificial intelligence will meaningfully increase productivity and economic growth. The topic is especially timely given the rapid diffusion of generative AI tools following the introduction of ChatGPT and the substantial uncertainty surrounding their long-term economic effects. The authors have written multiple studies on AI-driven productivity growth, allowing this article to build on an established research agenda. Their analysis extends prior work by Acemoglu, Aghion, and others by examining how differences across sectors may affect aggregate outcomes. The dataset is strong, combining task-level AI exposure measures, industry employment data, adoption data, and national input-output accounts. The findings are likely relevant beyond the United States because many advanced economies share similar sectoral structures and face comparable AI adoption decisions. Methodologically, the paper employs economic modeling and productivity projections rather than causal inference techniques or randomized experiments. Future research would be strengthened by incorporating causal designs that directly estimate AI’s effects on productivity and organizational performance.

Full Citation and Link to Article

Filippucci, F., Gal, P., & Schief, M. (2026). Aggregate productivity gains from artificial intelligence: A sectoral perspective. AEA Papers and Proceedings, 116, 31–35. https://doi.org/10.1257/pandp.20261035

Central Research Question

Filippucci, Gal, and Schief examine a central question in the emerging economics of artificial intelligence: How much will generative AI increase productivity across the U.S. economy, and to what extent might differences in AI adoption and productivity gains across sectors limit aggregate economic growth? While many commentators have argued that AI could dramatically transform economic performance, estimates of its macroeconomic effects vary widely. Some scholars, most notably Acemoglu (2025), have projected relatively modest gains, whereas others have suggested that AI could generate productivity improvements comparable to earlier general-purpose technologies such as personal computers and the internet.

The authors seek to move beyond broad aggregate projections by examining 65 U.S. industries individually. Their goal is not only to estimate sector-specific productivity gains but also to determine whether uneven gains across industries might create a Baumol effect. Baumol’s cost disease suggests that when productivity growth is concentrated in only a subset of industries, overall economic growth may be constrained because resources continue to be devoted to sectors experiencing slower productivity improvements. The paper therefore asks whether sectoral disparities in AI benefits meaningfully reduce economy-wide productivity growth over the decade following the introduction of ChatGPT.

Previous Literature

The paper builds upon two related strands of literature. The first concerns the potential economic consequences of artificial intelligence. Recent contributions by Acemoglu (2025), Aghion and Bunel (2024), Haskel et al. (2025), and others have attempted to quantify AI’s likely contribution to productivity growth. These studies differ considerably in their assumptions regarding the magnitude of productivity gains, the proportion of tasks exposed to AI, and the speed of adoption.

The second strand examines structural economic change and Baumol’s cost disease. Earlier research by Nordhaus (2008), Aghion, Jones, and Jones (2017), and Duernecker, Herrendorf, and Valentinyi (2024) emphasized that productivity growth concentrated in specific sectors may not translate directly into aggregate economic growth. If consumers continue demanding output from sectors experiencing limited productivity improvement, resources may shift toward those slower-growing sectors, reducing aggregate gains.

The authors position their work at the intersection of these literatures. Existing AI studies largely estimate aggregate productivity gains using versions of Hulten’s theorem, which focuses on direct productivity effects. However, such approaches generally do not account for the possibility that sectoral differences in productivity growth could alter the structure of the economy. The authors argue that understanding these broader equilibrium effects is necessary for evaluating the true macroeconomic consequences of AI adoption.

Data

The study combines multiple data sources to construct industry-level estimates of AI exposure, adoption, and productivity effects. The authors first rely on task-level exposure estimates developed by Eloundou et al. (2024). These measures identify whether specific occupational tasks can be completed at least 50 percent faster using large language models and related software tools.

The task-level information is then linked to occupational data from O*NET and employment composition data from the U.S. Bureau of Labor Statistics. This process allows the authors to estimate the proportion of work performed in each industry that is potentially exposed to AI. These industry-level estimates are subsequently aggregated into 65 sectors using value-added shares from Bureau of Economic Analysis input-output tables.

To estimate AI adoption, the authors use data from the U.S. Census Bureau’s Business Trends and Outlook Survey. These data indicate substantial variation across industries. For example, AI adoption rates in information-related industries are considerably higher than those observed in transportation and other manual-task sectors. The authors assume that these relative differences persist over time while overall adoption rises.

The resulting dataset combines information on occupational tasks, industry composition, AI exposure, current adoption rates, and sectoral economic output. This provides a comprehensive framework for projecting future productivity growth across the U.S. economy.

Methods

The authors employ a projection-based economic modeling approach rather than a causal inference design. Their estimates of productivity growth are derived using Hulten’s theorem and a series of assumptions regarding AI productivity gains, exposure rates, and future adoption.

The model assumes average task-level productivity improvements of approximately 30 percent when AI is applied to exposed tasks. This assumption is based on a review of experimental and empirical studies examining the effects of generative AI on task performance. The authors combine these productivity estimates with measures of AI exposure and projected adoption rates to estimate industry-specific gains in total factor productivity.

After generating sector-level projections, the authors construct a multisector general equilibrium model to evaluate whether differences across industries create meaningful aggregate constraints. This model incorporates input-output relationships among sectors and allows prices, production, and resource allocation to adjust in response to productivity shocks.

The framework examines several scenarios that vary according to the elasticity of substitution in consumption and the ability of labor and capital to move across sectors. These simulations enable the authors to estimate the magnitude of the Baumol effect under different economic conditions.

Although sophisticated, the methodology remains fundamentally predictive rather than causal. The paper does not estimate the observed effects of AI using experimental or quasi-experimental methods. Instead, it projects future outcomes based on existing evidence and economic theory. Consequently, the results depend heavily on assumptions regarding future adoption patterns and productivity effects.

Findings/Size Effects

The authors estimate substantial productivity gains from AI over the first decade following the introduction of ChatGPT. Across the 65 industries examined, annual contributions to total factor productivity growth range from approximately 0.1 percentage points in agriculture, mining, and other manual-task-intensive sectors to as much as 2.8 percentage points in knowledge-intensive sectors such as information services, finance, and professional services.

When aggregated across the economy, the projected annual increase in total factor productivity growth equals approximately 0.87 percentage points. This estimate is dramatically larger than Acemoglu’s (2025) projection of approximately 0.06 percentage points annually. The difference arises because the authors assume larger task-level productivity gains, greater task exposure, and faster adoption rates than those used in more conservative analyses.

The study also finds substantial heterogeneity across sectors. Industries characterized by cognitive and information-processing tasks receive much larger benefits than industries relying primarily on physical labor. Importantly, sectors with greater exposure to AI already exhibit higher adoption rates, reinforcing these differences.

Despite the pronounced variation across industries, the authors find that the resulting Baumol effect is relatively modest under most plausible scenarios. When consumers can substitute across goods and services and when labor and capital can move freely among sectors, aggregate productivity growth remains close to the weighted average of sector-level gains. Only under extreme assumptions involving highly limited substitution and restricted factor mobility does the Baumol effect substantially reduce aggregate growth. Even then, the reduction amounts to roughly one-quarter of the projected aggregate gains.

These findings suggest that uneven sectoral growth is unlikely to eliminate the broad productivity benefits generated by AI adoption.

Conclusion

The paper concludes that generative AI is likely to produce meaningful productivity gains across the U.S. economy during the decade following ChatGPT’s introduction. Although gains differ substantially across industries, the authors find little evidence that these disparities will dramatically limit aggregate growth through Baumol’s cost disease.

The study’s primary contribution lies in integrating sectoral heterogeneity into estimates of AI-driven productivity growth. Rather than treating the economy as a single aggregate unit, the authors demonstrate how variation in exposure and adoption shapes overall outcomes. Their analysis suggests that aggregate productivity growth may increase by nearly one percentage point annually, a magnitude comparable to some of the most significant technological transformations of the modern era.

At the same time, the paper highlights the uncertainty inherent in forecasting emerging technologies. The projections depend on assumptions regarding adoption rates, task exposure, and productivity effects that may evolve over time. Future research using causal inference methods, natural experiments, or randomized evaluations of AI implementation could provide more direct evidence regarding the actual economic effects of generative AI. Nevertheless, the study offers a rigorous and systematic framework for understanding how AI may reshape productivity growth across industries and the economy as a whole.