The Rise of Industrial AI in America: Microfoundations of the Productivity J-curve(s)
Source: https://www2.census.gov/ces/wp/2025/CES-WP-25-14.pdf ↗
Full text: author page ↗
Firm-level empirical evidence for the productivity J-curve associated with AI adoption.
Companies that adopt AI initially see no productivity gains — they can even get worse — because they need complementary investments in reorganisation, training and process redesign.
The benefits come later, and they are substantial.
It empirically confirms Brynjolfsson's thesis of the productivity paradox applied specifically to AI.
A useful reality check for any executive who expects a clean quarter-over-quarter line from an AI rollout.
Central argument
The paper argues that share repurchases are not causally responsible for the observed decline in labor share in U.S. corporations. Using Census firm-level data from 1982–2016, the authors find no evidence — across multiple empirical specifications, including an instrumental variable approach exploiting EPS-motivated buybacks — that increases in share repurchases lead to decreases in labor compensation relative to value added. The policy implication is direct: taxes on buybacks, such as the 1% excise tax in the 2022 Inflation Reduction Act, are unlikely to improve workers' share of income.
Critique
The paper's scope is confined to labor share as a ratio, which means it cannot fully adjudicate whether buybacks suppress absolute wage growth even when value added grows faster — a distinction that matters politically and welfare-wise even if the accounting identity holds. Additionally, the sample ends in 2016, predating the post-2017 tax reform buyback surge and the concentration dynamics in platform-driven industries, where the mechanisms linking capital returns to labor outcomes may differ structurally from the manufacturing-era firms that dominate Census longitudinal data.
Why it matters for product
This work is a cautionary model for product leaders who face pressure to attribute complex organizational outcomes — stagnant team productivity, poor retention, underinvestment in tooling — to a single proximate cause like budget allocation to one function over another. The paper's use of instrumented variation to isolate causality mirrors the discipline required when interpreting A/B test results or OKR correlations in product analytics: concurrent trends in your metrics dashboard do not imply the causal story your stakeholders want to tell. More concretely, when making the case for headcount or tooling investment, CPOs should resist the narrative trap of correlation-as-causation and instead design the organizational equivalent of a natural experiment to build credible arguments.