A Danish payroll study of 25,000 workers found AI saved about 2.8% of weekly work hours and shifted no earnings, because the productivity gain is captured at the firm boundary, not the worker.
If a chatbot can hand a knowledge worker one full hour of the week back, where does that hour actually go? On a single twenty-minute task, AI is a rocket: controlled studies report 15% to 55% speedups. In a real job, across a real month, on a real payroll, that rocket is now measured at a small, leaky gain that mostly evaporates before it reaches a paycheck. A study of roughly 25,000 workers across 7,000 Danish workplaces by economists Anders Humlum and Emilie Vestergaard is the cleanest field measurement of that gap to date.
The headline number is about 2.8% of work hours saved, call it one hour per worker per week. The number that should stop a CFO is what came next. According to the Humlum and Vestergaard results as summarized in the explainer that surfaced them, chatbot adoption had no statistically significant impact on earnings or recorded hours in any occupation they studied. Only 3 to 7% of the measured productivity gain reached anyone's pay.
Both findings are true. The lab benchmarks and the Danish payroll are not contradicting each other. They are measuring different things. Lab benchmarks count minutes saved on a structured task with a clean answer. The payroll counts the cash that survives contact with the firm's pricing power, the customer's budget, the manager's calendar, and the worker's own tendency to fill saved time with more work.
The mechanism that makes the gap durable is not adoption speed or measurement lag. It is a boundary. Productivity from AI is generated at the worker level. Compensation is set at the firm level. The two meet in a place where the firm decides whether the saved hour becomes a billed hour, a shipped feature, a new client, or a cost cut, and the worker almost never sits on the firm side of that line.
That is why the same study can report real time savings and a flat earnings line. The time is genuinely being made. The question is who is in the position to convert it into money, and the Danish data suggests the answer so far is: mostly no one. Discussion on Hacker News has been small but pointed, picking up on the gap between lab demos and the payroll result.
If the firm is the value filter, three things follow. First, AI ROI is not a worker problem to solve. It is a billing, capacity, and pricing problem to solve at the operations layer. Second, companies that already have billable utilization, throughput targets, or unit economics that can absorb an extra hour per worker per week will capture the value. Companies that sell salaried time into undifferentiated work will not, and the gain will leak back into meetings. Third, an AI policy that measures itself in hours saved is measuring at the wrong boundary and will systematically under-deliver on the P&L it claims to support.
The next signal worth watching is whether the Humlum and Vestergaard null holds up when replicated outside Denmark, and whether any firm discloses AI-related output per employee moving while headcount stays flat. Until one of those moves, the labor market is telling a cleaner story than the demos.