Ask a board how much AI will lift productivity and someone will quote a figure with two decimal places. The figures are real, drawn from careful field experiments, not marketing. The trouble is that they range from near zero to more than fifty percent depending on who is doing what, and that spread, not any single number, is the finding that should drive planning.
The same technology, four different numbers
Consider the strongest studies side by side. Peng and colleagues (2023) ran a controlled experiment in which developers were asked to build an HTTP server in JavaScript. Those given GitHub Copilot finished about 55.8 percent faster than the control group. A striking result, and a narrow one: a single well-scoped programming task with a clear finish line.
Move to messier work and the number falls. Brynjolfsson, Li and Raymond (2025), studying 5,172 customer-service agents with a generative-AI assistant, found average productivity, measured as issues resolved per hour, up about 15 percent. Dell'Acqua and colleagues (2025), in their field experiment with 758 Boston Consulting Group consultants, found a 25.1 percent speed gain and roughly 40 percent higher quality, but only on tasks inside the model's capability, and worse results on tasks outside it. Same class of technology, gains of 55.8 percent, 15 percent, and 25 percent, plus an outright negative in one condition. Anyone who lifts one of those figures into a budget is guessing dressed as forecasting.
Where the gains actually pool
The heterogeneity is not random noise. It has a shape, and the shape repeats across every study. The gains concentrate among the less experienced. Brynjolfsson, Li and Raymond (2025) found the strongest effects, up to about 34 percent, among novice and lower-skilled agents, with close to no measurable gain for the most experienced. Peng and colleagues (2023) report, in their later replications, higher adoption and larger gains among less-experienced developers. Dell'Acqua and colleagues (2025) found that AI compressed the performance distribution, with the lowest performers gaining the most.
For anyone doing the sums this matters more than the headline average. If most of your value comes from junior staff on standardized tasks, the returns could be large. If your advantage rests on a small number of senior experts doing non-standard work, the same tool may buy you little, and on the wrong tasks it may cost you, since the BCG experiment showed AI users deferring to confident wrong answers outside the model's range. The productivity number is not a property of the software. It is a property of the fit between the software and the specific job.
Why the average misleads governance
There is a second reason to distrust the single figure, and it is about control rather than magnitude. A 15 percent lift measured in a study was measured under study conditions: a defined task, a supervised rollout, a metric everyone agreed on beforehand. Most firms are deploying agents into far looser settings, and the World Economic Forum (2025), reviewing enterprise readiness, warns that a majority of organizations still do not know how to evaluate, manage, or govern AI agents responsibly even as adoption accelerates. A gain you cannot measure in your own operation is not a gain you can bank.
This is the gap between a productivity statistic and a productivity strategy. The statistic tells you the ceiling under favorable conditions. The strategy is the boring work of specifying which tasks, which workers, which quality bar, and which measurement, so that the number you eventually report is your own and not a citation. Brynjolfsson, Li and Raymond (2025) also found that AI assistance lifted customer sentiment and improved employee retention, effects that never show up if you track speed alone. The value can hide in places your dashboard was not built to see.
The practical read
Treat any published AI productivity figure as a hypothesis about your own operation, not a promise. Run the smallest honest test you can: one task type, one team, a pre-agreed metric, a control. Expect the gain to be biggest on standardized work and for less-experienced staff, and expect it to shrink or invert on the expert, judgment-heavy tasks that often matter most. The firms that will be disappointed are the ones that bought the 55.8 percent and budgeted for it across the whole organization. The firms that win will be the ones that found their own number, task by task, and knew why it was smaller.
Sources
- Peng, Kalliamvakou, Cihon and Demirer (2023). The Impact of AI on Developer Productivity. Controlled experiment where Copilot users finished a coding task about 55.8 percent faster, the high end of the productivity range. https://arxiv.org/abs/2302.06590
- Brynjolfsson, Li and Raymond (2025). Generative AI at Work. Field study of 5,172 support agents finding a roughly 15 percent average gain concentrated among novices, plus improved sentiment and retention. https://academic.oup.com/qje/article/140/2/889/7990658
- Dell'Acqua et al. (2025). Navigating the Jagged Technological Frontier. BCG field RCT showing gains on tasks inside AI's range and losses outside it, illustrating how context-dependent the numbers are. https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838
- World Economic Forum (2025). AI Agents in Action. Governance report warning most organizations cannot yet evaluate or measure agent performance responsibly. https://www.weforum.org/publications/ai-agents-in-action-foundations-for-evaluation-and-governance/