The Rise of Industrial AI in America: Microfoundations of the Productivity J-curve(s)

Source: Kristina McElheran, Mu-Jeung Yang, Zachary Kroff and Erik Brynjolfsson; working paper (preliminary, SIOE 2026 version) · 2025-12

This study examines the productivity dynamics of artificial intelligence (AI) in American manufacturing. Working with the U.S. Census Bureau to collect detailed large-scale data for 2017 and 2021, we find significant initial productivity losses preceding medium-term gains from industrial AI deployment. We attribute this to costly adjustment (not just mismeasurement), which we observe directly via increased work-in-progress inventory, investment in industrial robots, and labor shedding. Over time, however, early AI adopters exhibit stronger growth on average, conditional on weathering the initial “dip.” Losses vary considerably across firms and establishments. A key contingency is age, with young firms faring better than older incumbents—particularly startups with growth-oriented business strategies. Management practices and production-process design significantly shape both the uptake and effects of industrial AI use. Firm structure also matters, as we observe significant cross-establishment spillovers inside large, multi-unit firms. Overall, our detailed findings provide novel evidence regarding AI-related J-curve effects, unveiling key mechanisms and extending our understanding of this emerging General Purpose Technology.

Introduction. Advances in artificial intelligence (AI) are poised to reshape firm performance, labor demand, and the organization of production. Productivity gains from AI have been identified in narrow settings such as software development (e.g., Cui et al., 2024), certain types of drug and materials discovery (e.g., Lou and Wu, 2021; Strieth-Kalthoff et al., 2024), and customer support (Brynjolfsson et al., 2025). However, evidence outside of modular, data-rich, and already prediction-focused applications (Bresnahan, 2024) remains scarce—particularly beyond the level of a specific task or “point solution” (Agrawal et al., 2024).

Yet a broader firm- or establishment-level view is essential, given the importance of the organizational context in shaping the returns to technology use (e.g., Bresnahan et al., 2002; Bloom et al., 2012) and recent theory emphasizing the need for a “system-wide” approach to deploying AI (Agrawal et al., 2024). That said, representative data on firm-level AI adoption has been in short supply (Seamans and Raj, 2018), while adoption has also lagged expectations (Bonney et al., 2024) and varied significantly across industry and organizational contexts (Calvino and Fontanelli, 2023; McElheran et al., 2024). This has made it difficult to generate broad-based facts about AI’s prevalence and average productivity effects—–let alone pin down the organizational complements and constraints that have historically affected technology adoption and productivity ( David, 1990; Bresnahan and Trajtenberg, 1995; Brynjolfsson and Hitt, 2000; Feigenbaum and Gross, 2024).

We address this gap with new large-scale evidence on the use of AI among U.S. manufacturers and performance data across nearly a decade, drawing on two surveys conducted for this purpose in collaboration with the Census Bureau. In addition to providing representative statistics on both early AI use and barriers to adoption in physical-goods production, we leverage a novel instrumental variable (IV) strategy and visibility to within-firm changes to triangulate on the varied impacts of AI on business performance. Our findings are consistent with a J-curve pattern at the micro-level: the average effects of AI deployment are initially quite negative, followed by growth along multiple dimensions over time. These early losses, however, do not unfold uniformly; they vary substantially by age, business strategy, and other organizational characteristics and practices.

Even when voluntary, technology adoption does not guarantee productivity gains, particularly in the short run (Tambe and Hitt, 2012). With respect to AI, we find the initial relationship between adoption and performance to be negative, despite self-selection into AI use. Specifically, a one-standard deviation increase in our AI index is associated with a 1.33% lower total-factor productivity (TFP), controlling for a large number of time-varying (and usually unobserved) establishment characteristics. Moreover, when we employ quasi-experimental techniques commonly used to address endogeneity and measurement error, losses grow considerably in magnitude. IV estimates yield productivity losses of roughly 44%, which we interpret as a local average treatment effect (LATE) experienced by marginal businesses whose AI adoption is sensitive to our instrument.2 This nuance is important, as we do not interpret such a large productivity hit to be representative for infra-marginal “always adopters,” such as digitally-advanced businesses already poised to reap early returns from AI use in compatible applications. Yet we find negative effects, on average, across a range of specifications—ultimately underscoring the significant short-term costs of AI deployment, in practice. It further underscores the heterogeneity endemic to this phenomenon, reflecting the much greater prevalence in our sample of smaller firms and establishments that lack the advantages of larger, publicly-traded firms central to other studies (e.g., Babina et al., 2024).

Critically, these initial losses do not imply that AI adoption is systematically a “mistake.” With time for adjustment to unfold (and conditional on survival), we find growth to be significantly higher among early AI adopters, on average. This is particularly true among younger firms and establishments. For example, a one standard-deviation increase in AI use is associated with roughly 2% higher labor productivity growth among younger plants from 2017 to 2021. Among firms of all ages, greater AI use is associated with 0.8% employment growth over a four-year period that saw a 1.5% sector-wide decline, overall.

Data limitations prevent us from tracing out the precise inflection point where net returns become positive or characterizing true long-term implications. Yet, our results provide empirical micro-foundations for J-curve patterns previously observed at the macro level (Brynjolfsson et al., 2021b).

Related work. Rising digitization and the diffusion of advanced tools to extract value from data have had profound impacts on firm organization, management, competitive strategy, and performance (e.g., Bresnahan et al., 2002; Bloom et al., 2012; Goldfarb and Tucker, 2019; Adner et al., 2019; Tambe et al., 2020).

In this section, we extend insights from digitization research in economics and management to motivate our analyses of both the adoption and impact of AI use in industrial (i.e., physical goods) production.

Since our study takes place in the manufacturing sector, it is useful to bridge insights from the study of “digital firms” in the information sector with longstanding intuitions about industrial production, which has become increasingly automated and digitized in the United States. The strategic implications of economies of scale in physical-goods production have been understood since at least the early 1960s (Arrow, 1962; Henderson, 1968). One is early-mover advantage.

Economies of scale can be an important isolating mechanism, protecting early movers (Rumelt, 2005). First-movers are by definition able to accumulate production experience earlier. Under economies of scale, this increases value creation (by reducing unit costs, increasing quality, or both), and feeds back into yet more demand and higher production volumes (Cabral and Riordan, 1994). While industrial firms do not have the same demand-side economies as firms whose products and services are also digital (Giustiziero et al., 2023), we anticipate that “traditional” economies of scale combined with the non-rival attributes of AI technologies will mutually reinforce early-mover benefits from AI adoption.

A key limiting factor, however, arises from the general-purpose nature of the technologies in question. General-purpose technologies (GPTs), such as the steam engine, electricity or digital computers, are a class of economically and socially transformative technologies that share characteristics of being widely used, constantly improving, and innovation-spawning (Bresnahan and Trajtenberg, 1995). When we look at the family of technologies increasingly referred to as “AI” (e.g., machine learning, machine vision, speech recognition, etc.) they are already observed in every sector of the economy, in far-ranging use cases, advancing at unprecedented rates, and closely linked to innovation (Cockburn et al., 2018; Iansiti and Lakhani, 2020; Felten et al., 2021; Miric et al., 2023; McElheran et al., 2024).

Method. To provide robust estimates and study dynamics, we use two datasets on AI use by U.S. manufacturers at different time periods. Our main dataset is the 2021 Management and Organizational Practices Survey (MOPS), supplement to the Annual Survey of Manufactures (ASM). The ASM is one of the oldest large-scale Census Bureau data collections, sampling 10% of the roughly 300,000 manufacturing establishments across the country. Stratified by size and industry to provide representative statistics, the ASM nevertheless oversamples large establishments, covering over 70% of the sector’s total value added in the certainty sample. A government-mandated survey, the MOPS had a response rate of 68% in 2021.15 Panel A of Table 1 reports that the average establishment in our main sample has roughly 170 employees, is 29 years old, ships more than 60% of its products via e-commerce, and has approximately $11 million in profits.

Linking the MOPS to the broader ASM yields a useful breadth of organizational measures.

In addition to fine-grained data on the use of AI-related technologies, it provides information on the use of other digital technologies (predictive analytics, descriptive analytics, cloud computing, digitalization of information, IT capital) and structured management practices (Bloom et al., 2013, 2019). Dimensions of organizational and production design are captured in the ASM-MOPS linked data, as well (see below).

Our second main dataset is a panel of approximately 55,000 manufacturing firms combining data from the 2018 Annual Business Survey (ABS) with the Economic Census of Manufacturing (CMF) from 2012 to 2017. This panel dataset allows us to estimate the effects of AI adoption on performance while controlling for firm fixed effects. The ABS-CMF data are more representative of the large population of single-unit firms in the U.S. economy (Zolas et al., 2020). This different size coverage adds robustness to our findings, while also supporting an identification strategy focused on within-firm changes over time.

Panel B of Table 1 describes manufacturing firms in the ABS–CMF sample. The average firm in this sample is by no means small or young, with mean employment of 344 workers and age of 25. However, when comparing the MOPS–ASM data to the ABS-CMF sample, it is useful to keep in mind that the latter is a firm-level dataset whereas the former is at the establishment level.

Moreover the majority of establishments in MOPS–ASM belong to larger, multi-unit firms (hence our ability to identify within-firm spillovers).

Members of the author team worked with the Census Bureau to create new, designed-for-purpose measures of AI-related technologies and their applications for the 2021 MOPS. Measuring fastemerging technologies systematically and at scale is a perennial challenge. Given the recency and some lack of consensus about what “AI” encompasses, our measurement approach is multifaceted.

First, we clearly define what we mean by “AI” in the survey, as well as “predictive analytics,” which we separately measure to avoid confusion (and to distinguish from prior work, e.g., Brynjolfsson et al., 2021a). These definitions underwent systematic cognitive testing involving members of the research team and dedicated Census experts, as well as respondents pulled from the official ASM sampling frame, to ensure validity and reliability of responses.17 According to the survey, “Artificial Intelligence is a machine-based system that can perceive and learn about its environment and then make relevant predictions, recommendations or decisions for an objective that is determined by humans.” This definition is close to the legal definition of AI in the recent EU AI Act.18 Two points are worth emphasizing about this definition of AI. First, it shares with the EU AI Act the emphasis on systems that learn from their environment and adapt. Second, the last part was added because respondents were resistant to the idea that adaptive technologies might actually be choosing the objectives (more in line with “artificial general intelligence”).

Limitations. Although the timing of our survey precedes the introduction of generative AI tools in late 2022, the issue of “shadow AI” is an increasing concern when it comes to measuring AI diffusion and its impacts. The worry is that there could be AI use taking place in the firm, either undetected in machinery or software— or, as seems to be common with the rise of LLLMs, in use by individual employees— without managerial knowledge. We think the risk of this is lower with the Industrial AI applications studied here (compared to newer generative AI technologies). That said, it is worth considering the implications for our analyses. Ultimately, widespread “shadow AI” might bias downwards our adoption statistics. However, if AI use is prevalent and undetected in our sample, this should work against finding any significant productivity differences between firms based on observed AI usage.