AI adoption and employment: the company is only one part of the story

Alex Tabarrok’s August analysis draws on a Census survey supplement collected from November 2025 through February 2026. Among AI-using firms, 95.7% reported no effect on their employment, 2.3% an increase, and 2.0% a decrease.

The task-level answers show more movement: 44% said AI supplemented existing work, while 10% said it performed a task previously done by an employee. Those measures describe different things. Replacing some tasks need not immediately reduce a firm’s total headcount.

Zvi’s objection concerns the boundary of the measurement. A more productive company can win business and keep hiring while a competitor loses customers and jobs. Asking businesses about the effects of their own adoption does not automatically capture all those competitive effects, even when the survey covers many firms.

He also points to hiring decisions made in anticipation of future automation. A company might postpone training new staff before it can attribute an actual job loss to current AI use. That is a proposed channel of impact, not a quantity established by these survey results.

There is a separate measurement wrinkle: Census broadened its AI-adoption question in November 2025. The jump in reported adoption around that point cannot all be interpreted as sudden behavioral change.

The survey is useful evidence about what employers reported during a specific period. The dispute is over the larger inference: how direct effects, competitive shifts, and expectations combine across the economy.