The most consistent finding across the strongest studies of workplace AI is also the most double-edged. Give people an AI assistant and the least experienced improve the most, while the veterans barely move. Read one way, that is a story of fairness and fast onboarding. Read another, it is a warning: the tool that makes a novice productive today may remove the very experience that would have made them an expert tomorrow.
The leveling is real
The pattern is not a single result but a repeated one. Brynjolfsson, Li and Raymond (2025), studying 5,172 customer-service agents, found average productivity up about 15 percent, with the largest gains, up to roughly 34 percent, among novice and lower-skilled workers, and almost none for the most experienced. Their explanation is that the AI captured the tacit knowledge of the best agents and made it available to everyone else. Dell'Acqua and colleagues (2025), in a field experiment with 758 BCG consultants, found the same compression: AI lifted the lowest performers most, narrowing the gap between weak and strong. Peng and colleagues (2023), in their coding experiment and later replications, likewise report higher adoption and larger gains among less-experienced developers.
Three studies, three settings, one shape. AI raises the floor faster than the ceiling. For a firm staring at a hiring bottleneck or a long ramp to competence, that is genuinely good news. A new agent, developer, or analyst becomes useful in weeks rather than months, and the immediate productivity is real.
The rungs it removes
Expertise, though, is not downloaded. It is accumulated, and it is accumulated largely by doing the work that AI now does for the beginner. The junior consultant who struggled through a rough first draft learned something in the struggle. The novice support agent who had to reason through a hard case without a suggested answer built the judgment that later lets them handle the case no script covers. When the AI supplies the draft and the answer, the output improves and the learning that used to come with producing it may not happen.
This connects to the sharpest caveat in the same evidence base. Dell'Acqua and colleagues (2025) found that outside the model's capability, on the hard, non-standard tasks, AI users performed worse, deferring to confident wrong answers. Those out-of-range tasks are exactly where expert judgment earns its keep, and expert judgment is exactly what the apprenticeship ladder was supposed to produce. If AI both automates the rungs a novice would have climbed and struggles most on the tasks that require having climbed them, a firm can find itself with a workforce that is broadly capable on standard work and dangerously thin on the judgment for everything else.
Rebuilding the climb on purpose
None of this is an argument against the tools. The productivity is too clear, and the diffusion of best practice that Brynjolfsson, Li and Raymond (2025) documented is a real good, including their finding that AI assistance improved employee retention and customer sentiment. It is an argument that expertise development, which used to be a free byproduct of doing the work, now has to be designed in, because the work no longer produces it automatically.
Two design moves follow. First, protect deliberate difficulty. Some tasks should be done without the assistant, not because the output will be better but because the person will be. This is where the Human Agency Scale from Shao and colleagues (2025) is useful beyond preference measurement: choosing a higher-agency setting on selected tasks is a training decision, keeping the human doing the reasoning the AI could have done for them. Second, make the boundary explicit. The World Economic Forum (2025) recommends that agents be onboarded with defined roles, permissions, and supervision proportional to risk; the mirror image is defining where humans must retain the hard work so that someone in the organization is still building the judgment the out-of-range tasks demand.
The practical read
The firms that treat AI purely as a productivity lever will get a fast, capable, and quietly hollow workforce, strong where the model is strong and undertrained precisely where it is not. The firms that treat it as both a lever and a threat to their own expertise pipeline will do something less obvious: automate the drudgery, but deliberately preserve some of the hard, unassisted work that turns a competent beginner into someone who can catch the model when it is confidently wrong. AI can train the novice. It cannot, on its own, train the expert. That job is still yours, and it just got easier to forget.
Sources
- Brynjolfsson, Li and Raymond (2025). Generative AI at Work. Field study of 5,172 support agents finding gains up to 34 percent for novices and near zero for experts, as AI diffused expert practice. https://academic.oup.com/qje/article/140/2/889/7990658
- Dell'Acqua et al. (2025). Navigating the Jagged Technological Frontier. BCG RCT showing AI compressed the skill distribution but produced worse results on the out-of-range tasks that require expert judgment. https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838
- Peng, Kalliamvakou, Cihon and Demirer (2023). The Impact of AI on Developer Productivity. Finds larger gains and higher adoption among less-experienced developers, reinforcing the leveling pattern. https://arxiv.org/abs/2302.06590
- Shao et al. (2025). Future of Work with AI Agents. The Human Agency Scale offers a language for deliberately keeping humans doing certain tasks, useful as a training decision. https://arxiv.org/abs/2506.06576
- World Economic Forum (2025). AI Agents in Action. Recommends defining agent roles and supervision by risk, the mirror of defining where humans must retain the hard work. https://www.weforum.org/publications/ai-agents-in-action-foundations-for-evaluation-and-governance/