Quantifying the Effect of AI Invention on Productivity

Artificial intelligence (AI) represents a transformative general-purpose technology, yet the magnitude and distribution of its productivity effects across firms remain unclear.
This paper quantifies the impact of codified AI knowledge-captured through patenting activity-on firm-level productivity in the United States.
We construct a novel dataset linking 7.3 million U.S. patents (1990–2024) to Compustat financial data and develop a large-scale text-based classification framework, PaLLaFi (Patent Labeling via Language Models and Fine-Tuned Inference), to identify AI-related patents across ten subdomains, such as machine learning, natural language processing, robotics, and computer vision. Our empirical design combines high-dimensional fixed-effects regressions with dynamic event-study specifications to estimate productivity changes associated with AI invention. Along the intensive margin, firms expanding their AI patent portfolios experience 4-8 percentage point improvements in total factor productivity across specifications. Along the extensive margin, firms entering AI invention for the first time exhibit a gradual and monotonic productivity rise of up to 30% within five years, with no pre-trends, consistent with cumulative capability building.
By integrating patent-based measurement with econometric identification, and introducing the novel PaLLaFi framework, the study provides firm-level evidence that AI-driven innovation yields economically and statistically significant productivity gains.