Editor's Synopsis
Alex Imas presents a comprehensive, continuously updated review of research on AI's impact on productivity, organizing evidence into micro-level studies (controlled experiments and natural experiments) and macro-level studies (aggregate economic indicators). At the micro level, the evidence leans heavily positive: studies document productivity gains ranging from 14-15% for customer support agents to 55.8% faster task completion for developers using GitHub Copilot, with less experienced workers generally benefiting most. However, notable exceptions exist—the METR study found experienced open-source developers were actually 19% slower with AI, and BCG consultants performed worse when applying AI to tasks outside its capabilities.
At the macro level, AI productivity gains have only recently begun appearing in aggregate statistics. A March 2026 update notes that new BLS data revisions show US productivity growth of roughly 2.7%, and economists like Jason Furman now agree with Erik Brynjolfsson that AI may be contributing to these gains. Yet studies from Denmark show precise null effects on earnings and wages despite high adoption rates, and surveys indicate most firms report no measurable impact on employment or productivity.
Imas identifies several factors explaining the micro-macro disconnect: adoption frictions and lack of training (only 36% of workers feel properly trained), bottleneck tasks within jobs that AI doesn't speed up, endogenous selection into AI use by higher-skilled workers (contrary to micro findings showing lower-skilled workers benefit most), and a potential productivity J-curve where initial investments in reorganization suppress measured output before a harvest phase. He argues these frictions will diminish as organizations standardize AI tools and training, and predicts aggregate productivity numbers will reflect AI gains in the near term—especially as AI agents like Claude Code begin automating increasingly complex workflows.
