Does industrial AI adoption hurt productivity at first?

McElheran, Yang, Kroff and Brynjolfsson's working paper reports a productivity J-curve in industrial AI adoption among US manufacturers: "significant initial productivity losses preceding medium-term gains." Using two Census Bureau datasets — the 2021 Management and Organizational Practices Survey (MOPS) linked to the Annual Survey of Manufactures, and a firm-level panel from the 2018 Annual Business Survey combined with the Economic Census of Manufacturing (2012-2017) — the authors find that "a one-standard deviation increase in our AI index is associated with a 1.33% lower total-factor productivity (TFP)," with instrumental-variable estimates putting losses for marginal adopters at "roughly 44%." Over time, however, they find "growth to be significantly higher among early AI adopters, on average": a one-standard-deviation increase in AI use predicts "roughly 2% higher labor productivity growth among younger plants from 2017 to 2021," and, across firms of all ages, "0.8% employment growth over a four-year period that saw a 1.5% sector-wide decline, overall."

The authors attribute the initial dip to "costly adjustment (not just mismeasurement)," observed directly through "increased work-in-progress inventory, investment in industrial robots, and labor shedding" rather than inferred statistically. The size of the loss and the speed of recovery vary by firm characteristics: age is a key contingency, "with young firms faring better than older incumbents — particularly startups with growth-oriented business strategies" — and "management practices and production-process design significantly shape both the uptake and effects of industrial AI use." Firm structure matters too: the authors observe "significant cross-establishment spillovers inside large, multi-unit firms," meaning the unit that bears the dip is not always the unit that eventually captures the gain.

This measured J-curve gives empirical weight to Does AI adoption require rethinking how organizations coordinate?'s claim that task-level AI adoption inside an uncoordinated organization can make outcomes worse before communication and structural change let a firm capture a larger payoff: here, management practices, production-process design, and multi-unit coordination are exactly the organizational variables that separate firms stuck in the dip from firms climbing out of it. It also sits differently from Why do individual AI gains not translate to business results? and Do AI productivity gains feel larger than they actually measure?: both of those rely on self-reported perceptions of AI's effect, while this paper measures TFP, inventory, robot investment, and employment directly from Census records, so the losses and gains it documents are not a perception gap but an observed change in establishment-level production.

The paper is a preliminary working paper circulated ahead of SIOE 2026 and not peer-reviewed; the authors caution that "data limitations prevent us from tracing out the precise inflection point where net returns become positive or characterizing true long-term implications." Its AI measure also predates "the introduction of generative AI tools in late 2022," and the authors flag "shadow AI" — unmeasured individual LLM use inside a firm — as a risk to their adoption statistics that would, if present, bias effect sizes toward zero rather than away from it. The findings are specific to US physical-goods manufacturing measured through Census surveys; whether a comparable J-curve holds for generative AI adoption, services firms, or knowledge work is not something this excerpt addresses.