A 2023 NBER study of 5,179 support workers shows generative AI inverts the experience curve, raising a sharp question about how the next generation of experts will be trained.
When a customer-support agent with three months on the job starts using a generative-AI assistant, their output jumps 34 percent. When a colleague with a decade of experience adopts the same tool, the line on the chart barely moves. That asymmetry, documented in a 2023 NBER working paper by Brynjolfsson, Li, and Raymond, is the wrong way around for every assumption about how workforces develop. (NBER WP 31161)
The researchers tracked 5,179 customer-support agents and found a 14 percent average gain in issues resolved per hour, with a striking redistribution: the least experienced agents captured most of the lift, the most experienced agents captured almost none. The mechanism the authors propose is direct. AI disseminates the conversational patterns of top performers, which means newcomers can borrow the playbook without serving the years of apprenticeship that built it. The technology does not raise the floor by spreading skill. It raises the floor by encoding skill and serving it on demand. (NBER WP 31161)
That reading inverts the usual AI-and-jobs frame. The dominant narrative treats large language models as a substitute for routine work, with the canonical estimate from Eloundou et al. finding that roughly 80 percent of U.S. workers could see at least 10 percent of their tasks affected and 19 percent could see at least half of their tasks affected, and that higher-wage occupations show greater exposure. (arXiv:2303.10130, "GPTs are GPTs")
In that frame, exposure is the proxy for risk, and higher exposure means higher risk. The Brynjolfsson finding suggests a different shape. The tasks AI is most capable of substituting are also the tasks that build expertise. Customer support at three months looks like answering tickets. Customer support at ten years looks like diagnosing the unusual ones, anticipating the angry ones, and knowing when to bend the script. AI compresses the first part of that arc, which means fewer hours of low-stakes practice, and fewer chances to discover what the rare cases feel like.
The labor-market data is starting to move in the same direction. The World Economic Forum's Future of Jobs Report 2025 surveys more than 1,000 employers covering over 14 million workers across 22 industry clusters and 55 economies, with projections through 2030 tied to technological, geoeconomic, demographic, and green-transition drivers. (WEF Future of Jobs Report 2025)
The report's headline numbers mix displacement and creation, but the experience-curve inversion offers a way to read them more sharply. If the productivity gain from AI is concentrated in the bottom of the experience distribution, the immediate return to firms from adoption is high, while the return to the labor market, in the form of newly minted experts, is delayed and probably smaller.
A second-order effect hides in that math. The paper's authors also report an early improvement in agent retention and customer sentiment at the firms in the study, which they read as a sign that AI assistance is making the work more bearable, not just faster. If so, the experienced agents who might otherwise churn are staying in roles where their expertise is being compressed into patterns the system can replay. The economic value of that expertise may still show up in the firm's balance sheet. The economic value of the next generation of experts, who never got the reps, may not show up at all.
There is a separate, more technical way to describe what is happening. A 2026 paper in AI & Society argues that tokenization, the process of breaking text into the units language models actually price and process, is not a neutral preprocessing step. It is a measurement system for labor. (Springer AI & Society, "The cost of language: tokenization as a metric of labor")
The authors, citing earlier work by Ahia et al., note that tokenization costs are not uniform across languages. Telugu users, for example, pay roughly five times the per-token cost of English users on OpenAI's APIs. The implication is that the unit of work inside an AI-mediated workflow is the token, and the firm that buys the workflow is buying tokens, even when the human on the other end of the call is doing something the token system cannot price. Legal signature, embodied witness, accountability for a wrong call: these are residuals the routing layer does not measure.
The frame is not yet consensus. The token-as-labor reading is a working hypothesis in a 2026 journal article and a cluster of 2025 and 2026 working papers, several of them still in preprint form and not yet individually verified. The Brynjolfsson productivity study is a working paper from 2023, widely cited, not a settled benchmark. The WEF numbers are employer surveys, not payroll measurements. None of this is enough to declare a labor-market regime change.
It is enough to ask one question. If the next generation of customer-support experts gets to ten years of experience having leaned on AI for the routine 80 percent, what is the curve they will sit on? The current curve says they will be roughly 14 percent more productive than their predecessors by the end of their first quarter, and roughly indistinguishable from them by the time they hit their second decade. That is good news for this quarter. The cost is the quarter after that, and the quarter after that, where the institutional memory of how the rare cases feel starts to thin.