"Generative AI at Work" is a 2023 empirical study by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond (Stanford / MIT / NBER) measuring the effect of a generative AI conversational assistant on worker productivity. It draws on the staggered rollout of the assistant across 5,172 customer-support agents and is one of the most cited papers on real-world AI adoption effects.
Summary of findings
The study finds a 15% average increase in productivity, measured as issues resolved per hour. The gains are distributed unevenly across the workforce: the largest improvements accrue to the least-skilled and least-experienced workers, while the most experienced agents see small gains in speed and small declines in quality. Less experienced agents improved on both speed and quality.
The authors report that access to the AI assistant accelerates worker learning, increasing the rate at which new workers improve. The effect is particularly large for international (non-native English) agents. Gains are largest for moderately rare problems, where human agents have less baseline experience but the AI has adequate training data. The deployment also changed the work experience: customers became more polite and less likely to escalate issues to managers.
Key claims
- A 15% average productivity increase (issues resolved per hour).
- Largest gains for lowest-skilled workers: less experienced agents improved both speed and quality, while the most experienced agents saw small gains in speed and small declines in quality.
- AI access accelerates the rate at which new workers improve.
- Particularly beneficial for international (non-native English) agents.
- Gains largest for moderately rare problems (human agents have less baseline experience but the AI has adequate training data).
- Customers became more polite and less likely to escalate to managers.
Relationships
- related: AI Labor Disruption — provides empirical evidence that AI disproportionately helps lower-skilled workers, partially challenging the "ability-slicing" prediction from Amodei.
- supports: AI as Normal Technology — consistent with the "normal technology" thesis that AI augments workers rather than replacing them, with productivity gains distributed unevenly.
- related: Labor Disruption Timelines: Who Predicts What and Why — the study documents augmentation rather than displacement, consistent with the "capability-deployment gap" narrative.
- related: Stanford HAI research on why corporate AI projects succeed or fail — complements organizational-adoption findings with direct productivity measurement (Source: hai.stanford.edu).
Provenance
PDF converted to markdown with images on 2026-04-13.