A peer-reviewed NBER working paper (Brynjolfsson, Li, Raymond) uses a staggered difference-in-differences design across 5,179 customer-support agents at a Fortune 500 firm to measure the real-world productivity impact of a GPT-based chat assistant deployed between late 2020 and early 2021. Access to the tool increased issues resolved per hour by 14% on average (138 log points in the preferred specification), with the bottom skill quintile gaining 34% and experienced top-performers seeing near-zero or slightly negative effects; the authors attribute the pattern to AI encoding tacit best practices of high performers and disseminating them to novice workers. Ancillary results show improved customer sentiment (+0.18 points, ~0.5 SD), a ~25% decline in manager-escalation requests, and a ~40% reduction in attrition among agents with under six months' tenure. For financial institutions, the study provides the first large-scale causal estimate of enterprise AI productivity ROI in a labor-intensive service operation, with measurable retention and quality co-benefits that bear directly on cost-per-contact and training-cost line items.
368 other tracked developments touching the same claims.