AI Utilization and Changes in Economic Performance
Is artificial intelligence (AI) utilization associated with changes in output and labor market outcomes? That is the question studied in a recent working paper by Tina Highfill and Jon D. Samuels of the U.S. Bureau of Economic Analysis (BEA) and Christos A. Makridis of the Gallup organization, Arizona State University, and Stanford University.
AI has rapidly moved from a specialized research domain into a widely diffused general purpose. Yet the magnitude and timing of its macroeconomic costs and benefits remain uncertain and widely debated. Unlike traditional capital investments, such as machinery or information technology equipment, many AI capabilities are not separately identifiable in most economic data, because they are embedded in software, cloud services, bundled digital products, and organizational routines. As a result, the contribution of AI to output, labor demand, and productivity is difficult to observe directly in conventional national accounts statistics and requires analyzing more disaggregated data.
The authors begin by benchmarking worker-reported AI use from the Gallup Panel against employer-reported AI adoption from the U.S. Census Bureau's Business Trends and Outlook Survey (BTOS). The comparison is useful because the two surveys approach AI diffusion from different perspectives. Gallup asks workers directly how frequently they use AI in their jobs, while BTOS asks businesses whether they use AI in their business functions. The Gallup data can also distinguish frequent users, defined as workers using AI daily or several times a week, from more occasional users. Both surveys show a substantial increase in AI utilization, although their measured levels differ. In the Gallup data, the share of workers reporting any AI use rose from roughly 20 percent in mid-2023 to nearly 50 percent by early 2026, while frequent use increased from about 10 percent to more than 25 percent. BTOS also shows rising adoption, but from a considerably lower base, increasing from approximately 4 percent to 18 percent over a broadly comparable period. Across overlapping survey waves, Gallup's estimates exceed the corresponding BTOS estimates by roughly 15 to 32 percentage points.
Whereas Gallup captures AI use from the perspective of individual workers, including task-level use that may occur even when an employer has not formally deployed or governed an AI system, BTOS instead captures whether firms report using AI in their business functions. Consistent with this interpretation, Gallup respondents are also considerably more likely to report their organizations have begun integrating AI into their practices than firms report adoption through BTOS. The gap appears to reflect, at least in part, differences in what the two surveys measure rather than simple measurement error. The distinction became somewhat smaller after the BTOS question was broadened in late 2025 from AI use “in producing goods and services” to use “in any of its business functions.”
Importantly, however, the two measures tell a very similar story about where AI adoption is occurring. Across industries, worker-reported AI use and employer-reported AI adoption have a weighted correlation of 0.93 in the most recent matched observations. Industries with relatively high AI use in Gallup are therefore generally the same industries reporting relatively high adoption in BTOS, even though the absolute levels differ. This suggests the surveys are capturing a common underlying process of AI diffusion.
Next, the authors link their state-by-industry measures of worker-reported AI utilization to annual data on employment, earnings, and real output from the Census Bureau's Longitudinal Employer-Household Dynamics program and from BEA. They estimate dynamic specifications that compare the evolution of economic outcomes across state-industry cells with different levels of AI utilization, while absorbing persistent state-by-industry differences and common industry-specific changes over time through fixed effects. The results show that state-industry cells with higher levels of worker-reported AI use experienced stronger real-output trajectories after 2020. Employment differences are also generally positive, although they are estimated less precisely. The pattern is therefore more consistent with AI-intensive cells expanding output alongside stable or somewhat stronger employment than with a simple displacement story in which higher AI use is associated with declining labor demand. Estimating similar regressions using only industry-level variation from BTOS produces null results, which suggests the more disaggregate state-by-industry variation is important.
The authors note several limitations, particularly the potential for reverse causality and the thin sample for some states and industries. They nonetheless point to directions for future work, emphasizing that the paper should be read as an early step in a broader measurement project: the evidence is consistent with AI being associated with stronger output and productivity growth at the state-industry level, but specifically identifying AI's causal contribution remains difficult, given the endogeneity of adoption and the uneven, but already expanding, enterprise diffusion of AI before 2022.
