Generative AI at Work

Summary

Brynjolfsson, Li and Raymond ran one of the first large-scale field studies of generative AI augmenting human knowledge work. They followed 5,179 customer-service agents at a Fortune 500 software company that rolled out a generative AI assistant trained on the company's own historical chat transcripts. The staggered rollout let them compare workers who had AI access against those who did not, while controlling for the usual differences across teams.

Average productivity rose 14% as measured by customer issues resolved per hour. The interesting story is who gained: the newest and lowest-performing agents saw 34% improvements, while experienced top performers gained little. Customer satisfaction rose, and escalations to managers went down.

The pattern matters because it inverts a common worry. AI did not replace the work; it raised the floor of what a new agent could do. The implication is hopeful and specific. AI can be a leveling tool when it is built on the institution's own knowledge and deployed alongside the people doing the work, rather than over them.