What changed
A Fortune 500 software company introduced a chat assistant that suggested replies to customer-support agents while they worked. The tool had learned patterns from previous conversations, including responses associated with strong outcomes.
Across 5,179 agents, access to the assistant raised productivity by 14% on average. The least experienced and lowest-skilled workers improved by 34%, while highly experienced agents saw little change. The gains showed up in issues resolved per hour, faster handling, and the number of chats managed at once.
What this could change for you
The striking result is not that the best workers became superhuman. It is that newer workers appeared to absorb some of the habits of experienced colleagues much faster.
Used well, that could make the first months in a job less punishing and give customers more consistent help. It could also free managers from repeating the same coaching—if workers still have room to think, disagree, and learn rather than blindly copy suggestions.
What it does not prove
This was one company, one kind of customer-support work, and a staggered rollout rather than a conventional randomized trial. The findings may not transfer to every industry or tool.
Productivity is not the whole job. The study does not settle questions about surveillance, deskilling, pay, job loss, or who benefits financially from the added output.
The bottom line
Generative AI looked most useful as an on-the-job coach for people still climbing the learning curve. That is a more grounded promise than replacing experts—and it gives employers a clear standard: measure whether the worker actually benefits.
Primary research
Generative AI at Work
NBER Working Paper 31161 · 2023 · DOI 10.3386/w31161
