Generative AI at Work
Citation. Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. https://doi.org/10.1093/qje/qjae044
Summary. This paper studies a staggered workplace rollout of a generative AI conversational assistant among 5,172 customer-support agents. Access to AI assistance increased productivity by 15% on average, measured as issues resolved per hour, with the largest gains for less experienced and lower-skilled workers. The evidence suggests that GenAI can disseminate high-performer practices through the workflow, compressing parts of the learning curve rather than simply replacing human labor.
Key claims
- AI access increased issues resolved per hour by 15% on average, with substantial heterogeneity across workers.
- Less experienced and lower-skilled agents improved both speed and quality, while the most experienced and highest-skilled agents saw small speed gains and small declines in conversation quality.
- The authors find evidence that AI assistance facilitated worker learning: gains persisted during software outages, especially among workers with greater prior exposure to AI recommendations and higher adherence to suggestions.
- AI assistance improved English-language fluency, especially among international agents, and moved agent communication patterns toward those associated with higher-skilled workers.
- Gains were largest for moderately rare problems, where human agents had less baseline experience but the AI system still had enough training data to provide useful suggestions.
- The authors report improvements in some work-experience/customer-interaction measures: customers were more polite and less likely to ask to speak to a manager.
Evidence & limitations
- Design: staggered rollout of a GenAI conversational assistant across customer-support agents in a Fortune 500 enterprise software firm and its subcontractors; agent, time, and rollout variation support causal estimates of AI access on productivity and quality outcomes.
- Evidence strength: large workplace sample, real production data, measurable output and quality indicators, and mechanism tests using adherence, outages, topic rarity, text analysis, and customer interaction patterns.
- Limitations: single firm and customer-support setting; the tool was designed to augment agents and withhold suggestions when training data were insufficient; outcomes are short-run and task-specific; the data do not identify long-run wages, labor demand, hiring composition, or whether top-worker knowledge becomes depleted if AI reduces their original contributions.