Do executives overstate AI productivity gains?
About 750 US finance executives say AI raised labour productivity 1.8% in 2025. Calculated from the same firms' own revenue and staffing figures, the gain is 0.6%.
What executives say AI saved is about three times what their own revenue and staffing estimates imply. In a survey of nearly 750 US finance executives, firms said AI raised their labour productivity by 1.8% in 2025. Calculated from the same firms’ reported revenue and staffing effects, the gain is 0.6% [1].
The numbers
-
- Reported gain 1.8%. Gain implied by the firms’ own revenue and employment figures 0.6%.
- 2026, expected. Reported gain 3.0%. Implied gain 1.8%.
- Large firms expect AI to cut their headcount by 0.7% in 2026.
- The authors find little evidence of near-term declines in total employment due to AI.
What the study did
The authors, Salomé Baslandze, Zachary Edwards, John Graham and colleagues from the Federal Reserve Banks of Atlanta and Richmond and Duke University, surveyed senior finance executives, mostly chief financial officers. The main survey ran from 11 November to 16 December 2025 and drew 603 responses. A supplementary survey of finance executives, run from mid-December 2025 to mid-January 2026, added 145 [1].
The survey asked two kinds of question. First, by how much did AI change your firm’s output per worker? That gives the reported figure. Second, by how much did AI change your firm’s revenue, and by how much did it change your employment? Taking the staffing change away from the revenue change gives the implied output per worker. Both come from the same executives.
This is a working paper from the National Bureau of Economic Research, issued in March 2026. It has not been peer reviewed.
Our reading
An executive who is asked about productivity answers with a feeling about time saved, quality and speed. Asked about revenue and headcount, the same executive answers in terms the firm counts. When the two disagree by a factor of three, the second is closer to a measurement.
The authors give a fair reason for the gap. They attribute it to delay, since gains such as faster decisions and better quality have not yet reached revenue. It is also a claim an investor cannot check. Until the revenue appears, a claimed gain is a forecast.
For a manager, the recommendation is simple. Before you budget a saving, fix a baseline. Pick the output you care about, such as cases closed, reports delivered or cost per unit. Record it for a team using the tool and a comparable team that is not. Then compare the change in each. The comparison method is the same one we describe in our note on difference-in-differences.
For a board, ask for the measured result next to the claimed one on the same page. For a funder or a buyer of AI services, put the gain in the contract as a measured outcome with a stated baseline.
Measure before you believe.
Limits
Both figures are self-reported. The implied number is the executives’ own estimate of AI’s effect on revenue and staffing, not audited accounts. The sample is not a random draw of US firms, so the authors reweight their aggregate results to match the US Census. The paper reports what firms said about 2025 and what they expect for 2026.
Sources
- Baslandze, S., Edwards, Z., Graham, J., McClure, T., Meyer, B. H., Sparks, M., Waddell, S. R. and Weitz, D., "Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives", NBER Working Paper 34984, March 2026, https://www.nber.org/papers/w34984 (checked 2026-10-29)