What the research found
Brynjolfsson, Li and Raymond's Generative AI at Work appeared in The Quarterly Journal of Economics in 2025. The study analyzed a staggered deployment of an assistant among 5,172 customer-support agents. Access was associated with a 15% average increase in issues resolved per hour, with larger gains among less experienced and lower-skilled workers. The most experienced workers saw small speed gains and small quality declines. The published result should not be confused with figures from earlier working-paper versions. [1]
What the design can establish
This is evidence from a particular customer-support organization and deployment, not a randomized healthcare trial. The assistant proposed responses while agents retained responsibility. The results concern task performance and worker differences; they do not establish autonomous clinical capability, healthcare labor savings or the share of economic benefit captured by a software vendor. Transfer to a new workflow is a hypothesis. [1]
My operator interpretation
I would identify where a healthcare organization repeatedly recreates expertise that could be made easier to access. Administrative work with traceable source information may provide a bounded starting point. The design should specify the input, the proposed output, the responsible reviewer and the condition that sends a case to a specialist.
The worker differences make deployment strategy material. I would not assume that the same interface, training or review requirement fits a new employee and an experienced specialist. The evaluation should examine both groups and preserve a route for experts to correct the system. Averages can hide a tool that helps one part of the workforce while imposing work or reducing quality elsewhere.
The operating value depends on the destination of released capacity. It may reduce a backlog, shorten response time, improve service consistency or support more completed work. I would name that destination before recognizing a financial benefit. Faster handling of one step can otherwise move the queue to another team without improving the overall service.
I would negotiate the contract against the source of value the pilot actually establishes. A fee linked to user seats, task volume or completed work distributes adoption risk differently. High utilization can improve the supplier’s revenue while eroding the buyer’s return if review and correction rise with every use. Compare the total cost per correctly completed workflow, including exceptions, with the credible alternative. Then test whether the price still leaves the buyer a reason to renew after initial learning gains diminish.
The capital decision is also relative. If the constraint has moved from drafting to expert review, the next dollar may belong in reviewer capacity or process redesign rather than additional licenses. This is why I would reconcile productivity by role with throughput at the end of the service. The studied improvement is informative; the investment case depends on where scarce capacity is released, where it is consumed and who receives the resulting value.
The decision I would make
I would expand only after a controlled operational pilot demonstrates a useful change in completed work and an acceptable quality profile across relevant worker groups. The financial case should then show how that measured change affects actual capacity or spending. This is an executive application of the evidence, not a healthcare outcome reported in the paper.
What this evidence cannot settle
The study concerns customer support outside healthcare. Its productivity estimate should not be entered as an assumed clinical or healthcare administrative effect.