Generative AI at Work
Source: https://academic.oup.com/qje/article/140/2/889/7990658 ↗
Full text: open-access via OpenAlex ↗
Empirical study with more than 5,000 customer-support agents.
AI increases productivity 15% on average, but the effect is uneven: less experienced workers improve 30% in speed and also in quality, while the most experienced barely improve and can even degrade slightly in quality.
AI compresses the learning curve — agents with two months of experience perform like agents with six months without AI.
Essentially, AI captures the tacit knowledge of the best performers and distributes it.
It is empirical evidence of cognitive decentralisation: something that used to live in a few heads now flows through the entire team.
Central argument
Brynjolfsson, Li, and Raymond study the real-world deployment of a GPT-based chat assistant across 5,179 customer support agents and find a 14% average productivity gain—but the distribution is sharply uneven: novice and low-skilled workers improve by 34%, while experienced and high-skilled workers see minimal or even negative effects. The core argument is that generative AI works by capturing and disseminating the tacit knowledge of top performers, effectively encoding best practices that previous waves of automation could never reach. This makes AI a skill-leveling technology, not a skill-amplifying one—a direct inversion of the skill-biased technical change observed with earlier computerization.
Critique
The study is conducted within a single Fortune 500 firm in a single industry—customer support—where tasks are relatively structured, outcomes are measurable, and AI suggestions are directly embedded in the workflow; generalizing the 'AI democratizes tacit knowledge' thesis to less constrained knowledge work (product strategy, design, engineering judgment) requires assumptions the paper does not validate. There is also a deeper tension in the learning findings: the authors show productivity gains persist during AI outages, suggesting genuine skill transfer, but the convergence in communication patterns toward high-skill norms raises the question of whether workers are actually learning or merely becoming dependent on a narrowing behavioral template that could erode adaptive capacity over time.
Why it matters for product
For a CPO, the 34% productivity gain among novices directly challenges the assumption that AI tooling should be prioritized for senior practitioners—it suggests that onboarding, junior IC ramp-up, and team scaling economics change fundamentally when AI is embedded in core workflows, shifting where you invest in hiring versus tooling. The finding that AI assistance reduces customer escalations and improves retention also reframes how product teams should think about measuring AI feature value: customer sentiment and employee attrition become first-class outcome metrics, not just resolution rates or throughput.