Productivity gains from using AI are captured by individuals, not organizations

Companies are seeing small to moderate gains from genAI (https://www.oneusefulthing.org/p/making-ai-work-leadership-lab-and), even though an increasing percentage of people are using AI at work, 30%-60% across many disciplines. A quick look at the list of tasks compiled through a survey by a team of economists in 2025 provides an explanation for this apparent paradox: most of these tasks are defined at the individual level. And most tasks that employees use AI for are individual tasks, for example reducing the time it takes for one person to write a report from 2 hours to 20 minutes.

So the best that an organization can expect from this approach where, for example, everyone is given full access to ChatGPT, is the sum of individual productivity gains, but that is very much a theoretical maximum: the organization-level gain is likely well below that. In other words, the whole is less than the sum of the parts. For example, the number of reports to be written does not necessarily increase sixfold to fill the void, moving the productivity gain into an ill-defined overall gain for the organization: what is the employee going to do with the remaining 100 minutes?

But the more fundamental issue is that genAI-based tools are currently individual tools, be they for writing, spreadsheeting, ideating or even coding. For superlinear gains, where the whole is more (even way, way more) than the sum of the parts, individual tasks must interact, between and across workflows. The current enterprise-grade unreliability of genAI tools means that any agentic architecture may compound errors instead of creating more value. That is where the focus should be: chaining unreliable tools is unlikely to self-correct.

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5136877

The Labor Market Effects of Generative Artificial Intelligence, Jonathan Hartley (Stanford University), Filip Jolevski (George Mason University; The World Bank), Vitor Melo (Clemson University), Brendan Moore (Stanford University)

Average number of minutes to complete a task with and without Generative AI