Headcount, not revenue, is where Anthropic’s internal study gets uncomfortable. The company says its own AI agents were putting in six times the working hours of its human staff by August 2026.
March is the pivot the authors point to: from that month on, agent hours exceeded human hours. The gap widened until the ratio reached 6 to 1 company-wide by August, and 19 to 1 inside research departments. Agent runs carry work-hour estimates so both sides can be plotted on a single axis.
Junior hiring moved in the same direction. Technical departments went from 21 percent junior staff in early 2025 to 14 percent. Among people brought in, the fall is sharper still, 27 percent down to 9 percent, and open junior postings now sit at 6 percent, with the steepest decline beginning in December 2025. Other departments held steady.
Output climbed alongside. Per-engineer code commits ran 5.1 times higher. Queries handled by data scientists, and contracts cleared by the legal team, each reached 3.1 times the earlier level. Deals closed per seller rose 1.7-fold, recruiters made 1.4 times the hires, and accounting processed 1.4 times the invoices.
Anthropic sets limits on what the numbers mean. It calls itself an unusually AI-heavy software company, warns that fast growth and voluntary adoption of the tools may be inflating the effect, and notes that the wider economy still shows no measurable AI productivity gains.
What troubles critics is downstream. The dynamic has a name from Nolan Lovett, who researches at the NATO Special Operations University: a tragedy of the cognitive commons. A firm that drops entry-level roles in favor of software pockets the efficiency, while whoever needs seasoned staff later absorbs the cost. Judgement is built somewhere, and the rungs it was built on are being pulled up.
