Fewer than one in four U.S. manufacturing plants reported using AI in 2021, with uptake concentrated among larger, digitally modernized facilities; cost, unclear use cases and skill gaps—not past productivity—appear to limit diffusion.

The Adoption of Industrial AI in America

Kristina McElheran, Mu-Jeung Yang, Zachary Kroff, Erik Brynjolfsson · Fetched May 23, 2026 · AEA Papers and Proceedings

semantic_scholar correlational medium evidence 7/10 relevance DOI Source

In 2021 only 22.8% of U.S. manufacturing plants reported any AI use, with adoption concentrated in larger plants that have more recent digital infrastructure (cloud and predictive analytics) and structured production-process management, while cost, lack of applicable use cases, and expertise are the main cited barriers and prior productivity does not predict adoption.

Using a mandatory, purpose-designed Census Bureau survey of approximately 28,500 establishments, we provide new evidence on industrial AI adoption in US manufacturing. Despite widespread digitization, only 22.8 percent of plants report any AI use as of 2021; intensity-weighted adoption is far lower. Adoption correlates with more-recent digital infrastructure—cloud computing and predictive analytics—rather than legacy on-premises IT or descriptive analytics. Structured production-process management and size are significant predictors. Cost and lack of applicable use case are the most cited barriers, followed by expertise. Prior productivity does not predict use, pointing to organizational readiness as a key barrier to AI diffusion.

Summary

Main Finding

Using a mandatory, purpose-designed Census Bureau survey of ~28,500 U.S. manufacturing establishments in 2021, the paper finds that AI adoption is limited: only 22.8% of plants report any AI use, and adoption measured on an intensity basis (share of activity/workforce using AI) is substantially lower. Adoption is associated with recent digital infrastructure (cloud computing and predictive analytics) and with structured production-process management and plant size; it is not associated with legacy on‑premises IT, descriptive analytics, or prior productivity. Firms most often cite cost and lack of an applicable use case as barriers, followed by lack of expertise—pointing to organizational readiness as a primary bottleneck for AI diffusion.

Key Points

Prevalence: 22.8% of manufacturing plants report any AI use in 2021; intensity-weighted measures show much lower penetration within plants.
Complementarities: AI adoption correlates strongly with more-recent digital infrastructure (cloud services, predictive analytics), implying complementarities between newer digital investments and AI.
Non-complements: Legacy on-premises IT and descriptive analytics do not predict AI adoption.
Organizational factors: Structured production‑process management practices and larger plant size significantly predict adoption.
Barriers: Top-cited barriers are cost and lack of an applicable use case; expertise shortages are also important.
Productivity selection: Prior productivity is not predictive of AI adoption, suggesting adoption is not concentrated among already-high-productivity plants.
Survey strength: The data derive from a mandatory, purpose-built Census survey, reducing nonresponse bias and improving representativeness compared with voluntary surveys.

Data & Methods

Data source: Mandatory Census Bureau establishment survey covering roughly 28,500 U.S. manufacturing plants (2021).
Measurement: Self-reported AI use (binary) and intensity-weighted adoption (share of activity/workforce exposed to AI).
Analysis approach: Statistical analysis of adoption correlates—assessing relationships between AI use and plant attributes (digital infrastructure types, management practices, size, prior productivity) and self-reported barriers.
Identification: Results are correlational evidence about predictors and reported barriers; the mandatory census design strengthens external validity and population coverage.

Implications for AI Economics

Diffusion constraint: Low prevalence and low intensity imply limited near-term aggregate productivity impact from AI in manufacturing unless adoption accelerates.
Importance of complementarities: Policies or investments that promote cloud and predictive-analytics infrastructure can unlock broader AI adoption—AI tends to follow more recent digital upgrades rather than legacy IT.
Organizational readiness matters: Because prior productivity does not predict adoption but management practices and firm size do, interventions should target organizational capabilities (managerial practices, implementation capacity) in addition to technical access.
Policy levers: Effective policy could include subsidies or tax incentives to lower adoption cost, publicly supported pilots to demonstrate use cases, technical assistance for identifying applicable use cases, and workforce training to address expertise gaps.
Measurement and research: Mandatory, establishment-level surveys are valuable for accurately tracking diffusion and intensity; future work should focus on longitudinal evidence of adoption dynamics and causal impacts on productivity.

Assessment

Paper Typecorrelational Evidence Strengthmedium — Uses a large, mandatory, purpose-designed Census Bureau survey that yields high-quality, nationally representative descriptive evidence on AI adoption rates and correlates in U.S. manufacturing; however, all results are cross-sectional and observational, so claims about causal effects (e.g., of digital infrastructure or management on adoption or productivity) are not identified and remain vulnerable to omitted variables, reverse causation, and measurement error from self-reports. Methods Rigormedium — High rigor in sampling and survey design (mandatory Census survey, large N ≈28,500) and careful reporting (binary and intensity-weighted adoption measures, multiple covariates including digital infrastructure and management practices); downgraded because analysis is correlational, relies on self-reported adoption and barrier measures, and lacks quasi-experimental or longitudinal strategies to address endogeneity. SampleA purpose-designed, mandatory U.S. Census Bureau survey of approximately 28,500 manufacturing establishments (plant-level) in 2021, reporting AI use (any use and intensity-weighted measures), digital infrastructure indicators (cloud computing, predictive analytics, on-premises IT, descriptive analytics), production-process management structure, establishment size, self-reported barriers to adoption (cost, lack of use case, expertise), and prior productivity measures; cross-sectional snapshot for 2021. Themesadoption org_design productivity GeneralizabilityOnly manufacturing establishments — excludes services and non-manufacturing sectors, U.S.-specific context — institutional, regulatory, and labor-market differences limit extrapolation to other countries, Cross-sectional 2021 snapshot — may not reflect faster adoption dynamics after 2021, Establishment-level data may not capture firm-level coordination or parent-company investments, Self-reported AI use and intensity may suffer from measurement error or heterogeneous definitions of 'AI'

Claims (8)

Claim	Direction	Confidence	Outcome	Details
The analysis uses a mandatory, purpose-designed Census Bureau survey of approximately 28,500 establishments. Adoption Rate	null_result	high	survey_sample_size / data source	n=28500 0.5
Only 22.8 percent of plants report any AI use as of 2021. Adoption Rate	negative	high	share of plants reporting any AI use	n=28500 22.8 percent 0.5
Intensity-weighted adoption is far lower than the 22.8 percent headline rate. Adoption Rate	negative	high	intensity-weighted AI adoption	n=28500 0.3
AI adoption correlates with more-recent digital infrastructure—cloud computing and predictive analytics—rather than legacy on-premises IT or descriptive analytics. Adoption Rate	mixed	high	association between AI adoption and types of digital infrastructure/analytics	n=28500 0.3
Structured production-process management and size are significant predictors of AI adoption. Adoption Rate	positive	high	AI adoption (predicted by management structure and size)	n=28500 0.3
Cost and lack of applicable use case are the most cited barriers to AI adoption, followed by expertise. Adoption Rate	negative	high	reported barriers to AI adoption (cost, applicability, expertise)	n=28500 0.3
Prior productivity does not predict AI use. Adoption Rate	null_result	high	predictive relationship between prior productivity and AI adoption	n=28500 0.3
The lack of a relationship between prior productivity and AI adoption points to organizational readiness as a key barrier to AI diffusion. Adoption Rate	negative	medium	organizational readiness as a barrier to AI diffusion	n=28500 0.03