AI at work: what the measured studies show

Most numbers about AI and productivity come from surveys that ask people how much time they think they save. A smaller set of studies measured the work itself, with a comparison group. This note summarizes four of them, then two federal data sources on how businesses use AI today. Every figure links to its source.

Customer support: 15% more issues resolved per hour

Researchers followed 5,172 customer support agents at one large software company as an AI assistant was rolled out in stages. Agents with the assistant resolved 15% more issues per hour on average. The gains were not even. "Less experienced and lower-skilled workers improve both the speed and quality of their output, while the most experienced and highest-skilled workers see small gains in speed and small declines in quality."^[1]

Consultants: faster and better, until the task is outside AI's range

In a preregistered randomized experiment, 758 consultants at a global consulting firm worked on 18 realistic tasks chosen to be within what AI does well. Those using AI completed 12.2% more tasks, 25.1% faster, with better quality. On a separate task chosen to be outside that range, the group using AI was less likely to reach the correct answer than the group working without it. The authors call this boundary a "jagged frontier": it is uneven, and it is hard to see from the inside.^[2]

The whole workforce: about 2 hours a week, for those who use it

A nationally representative US survey by the Federal Reserve Bank of St. Louis found that workers using generative AI reported saving 5.4% of their work hours, about 2.2 hours a week on a 40-hour schedule. Counting everyone, including people who don't use it, the saving was 1.4% of all hours.^[3] These are self-reported figures.

How US businesses use AI today

The Census Bureau's business survey found that 19.8% of US businesses had used AI in the previous two weeks (as of May 3, 2026). Use rises with size: 37% of firms with at least 250 employees, and fewer than 20% of firms with four or fewer.^[4]

A Census research paper on the same survey looked at how firms use it. Most users (66%) rely on AI only to assist people with tasks, and staff reductions related to AI were rare, at 2% of firms.^[5]

What we take from it

This part is our reading, not the studies'.

• The gains are real when the task fits. Repetitive, well-defined work with a clear right answer is where every study above found improvement.
• Pick the task before the tool. The consultant study shows the same tool helping on one task and hurting on another. Deciding which work goes to AI is the important decision.
• Keep a person on the exceptions. Outside the frontier, AI was confidently wrong more often. A check that routes uncertain cases to a person is what keeps the gains without the errors.
• Most firms use AI to help people, not replace them. That matches what the federal data shows.

We wrote more about choosing the work in How to decide which business processes to automate and about the checks in Deterministic before intelligent.

Sources

1. Brynjolfsson, E., Li, D., and Raymond, L. "Generative AI at Work." Quarterly Journal of Economics 140(2), 2025. doi.org/10.1093/qje/qjae044
2. Dell'Acqua, F., et al. "Navigating the Jagged Technological Frontier." Organization Science 37(2), 2026. doi.org/10.1287/orsc.2025.21838
3. Bick, A., Blandin, A., and Deming, D. "The Impact of Generative AI on Work Productivity." Federal Reserve Bank of St. Louis, February 2025. stlouisfed.org
4. US Census Bureau. "Large Firms With at Least 20 Employees Biggest AI Users." May 2026. census.gov
5. Bonney, K., et al. "The Microstructure of AI Diffusion." US Census Bureau, Center for Economic Studies, CES-WP-26-25, April 2026. census.gov (PDF)