Rather than between distant roles, the true adoption gap lies between teammates working side by side.
Listen to this article:
Imagine two analysts who work on the same floor, hold the same title, and have the same subscriptions to AI tools. One has integrated AI into most of their workday, while the other tried it once, got a mediocre answer, and hasn’t opened it since. This is an example of identical access with opposite behavior, a gap that sits at the heart of the new NBER working paper, “What Work Does Generative AI Do?” co-written by 2026 HBS AI Institute Associate David Deming. Using a nationally representative worker survey, Deming and his co-authors shift the AI adoption question to the actual work AI is being used for. Their findings say as much about the people using the technology as about the technology itself.
Key Insight: AI’s Footprint in the Workplace
“The present distribution of genAI adoption can therefore be characterized as widespread but shallow.” [1]
AI cannot affect the workplace until workers actually use it. Using the Real-Time Population Survey across four waves between August 2025 and May 2026, the researchers asked workers which tasks they regularly perform and whether they actively use generative AI to accomplish them. This methodology deliberately measures realized, operational use, and the results show that genAI reach is broad. More than 80% of occupations and more than 40% of tasks have adoption rates above 20%. Depth, however, lags behind. Only about 15% of occupations exceed 70% adoption, and no individual task does. Only 2.8% of tasks exceed 50%. The highest adoption rates cluster around information-intensive work. Computing occupations dominate the highest-adoption list, alongside management, finance, market research, and related roles. High-adoption tasks include preparing research reports, analyzing data, developing plans, and working with technical information.
Key Insight: The Human Factor
“[S]ome workers are systematically more likely to adopt genAI than others.” [2]
The researchers find that exposure is an incomplete guide to use. Clerical and administrative work makes this especially clear. This work can look highly exposed in a capability-based analysis because much of it involves text, documents, records, and routine information processing. Yet adoption can be modest: receptionists and information clerks, for example, had an actual adoption rate of 7.6%, versus an exposure-based prediction of 54.0%. Medical secretaries and administrative assistants registered 16.8% actual adoption against a 61.0% prediction. Why might the apparent opportunities not translate into use? The authors point to a range of factors: organizations may restrict AI access to sensitive data, established procedures can make experimentation difficult, and workers may differ in both the skills and incentives needed to incorporate a new tool into their routines. Additionally, task identity only explains about 11% of whether a worker uses genAI for a particular task. Factoring in person-to-person differences explains 44% of adoption. A person’s individual traits, habits, and willingness to learn dictate whether they use AI far more than the actual duties assigned to their role.
Key Insight: AI Adoption May Build on Itself
“We have shown that some workers are systematically more likely to adopt genAI than others, regardless of the tasks they perform.” [3]
The researchers find several patterns consistent with the idea that learning to use generative AI in one setting makes it easier to adopt elsewhere. Among current users, people who had already been using generative AI six months earlier reported using it across about 0.4 additional tasks, a roughly 9% increase relative to the 4.7 tasks workers report performing on average. Workers were also more likely to adopt AI for one task when their other tasks had higher adoption, and workplace use was positively related to adoption outside work. The researchers note that these patterns are consistent with learning, where someone pays a one-time cost to work out a skill (like AI prompting, review, and workflow fit). Once that hurdle is cleared, the perceived cost of trying (AI) on a new task drops precipitously. If experience lowers the hurdle for subsequent use, today’s adoption could gradually expand across more of the workday.
Why This Matters
Given these signals, leaders interested in increasing AI adoption should identify high-value tasks with low realized AI adoption, remove unnecessary barriers, and watch how use spreads as employees gain experience. The research suggests that usage can compound: a successful use case may build the confidence and expertise that makes the next use case easier. Because adoption builds momentum this way, ultimate success hinges on an organization’s ability to systematically turn early experimentation into widespread, everyday capability.
References & Notes
[1] Bick, Alexander, Adam Blandin, David J. Deming, and Tyler R. Schumacher, “What Work Does Generative AI Do?,” NBER Working Paper 35677 (2026): 15. https://doi.org/10.3386/w35677
[2] Bick et al., “What Work Does Generative AI Do?,” 31.
[3] Bick et al., “What Work Does Generative AI Do?,” 22.
This article was created through a human-led, AI-assisted process.
Meet the Authors
Alexander Bick is Senior Economic Policy Advisor at the Federal Reserve Bank of St. Louis.
Adam Blandin is Assistant Professor of Economics at Vanderbilt University.
David J. Deming is the Danoff Dean of Harvard College, the William Henry Bloomberg Professor of Economics, the Isabelle and Scott Black Professor of Public Policy at Harvard Kennedy School, and 2026 Associate at the HBS AI Institute.
Tyler R. Schumacher is a PhD student at Vanderbilt University.