Artificial Intelligence Adoption and Econometric Identification: Evidence from Firm-Level Data

This paper examines the causal impact of Artificial Intelligence (AI) adoption on firm-level productivity using a newly constructed, multi-country panel of 11,482 firms across 23 countries and 18 industries over the period 2016–2023. Unlike existing studies that rely on single-country data, sector-level proxies, or fixed-effects estimates that inadequately address selection bias, this paper provides one of the first firm-level, cross-country causal estimates of AI’s productivity effects that explicitly quantifies the magnitude of bias inherent in standard econometric approaches. We document rapid but highly uneven diffusion of AI, with adoption increasing from 6% to 29% over the sample period and concentrated in data-intensive sectors. Using an instrumental-variables strategy based on industry-region diffusion of AI adoption, we show that conventional fixed-effects models substantially overstate the productivity gains from AI. Fixed-effects estimates imply productivity gains of approximately 19 percent, whereas causal IV estimates reduce the average direct effect of AI adoption to near zero. This discrepancy reveals that standard fixed-effects estimates exaggerate AI’s productivity impact by nearly fivefold due to positive selection of high-performing firms into adoption. Crucially, we find that AI generates economically and statistically significant productivity gains only when complemented by high levels of human-capital capability. These findings reframe AI not as a stand-alone productivity enhancer, but as a conditional general-purpose technology whose returns depend critically on workforce skills and organizational readiness. The results carry important implications for empirical research, firm strategy, and public policy aimed at fostering AI-driven productivity growth.