In March 2026, researchers from Harvard, the Federal Reserve Bank of St. Louis, Vanderbilt University, and WZB Berlin presented a study at the Brookings Papers on Economic Activity conference that quantified something most business leaders have sensed but couldn’t prove. They tracked generative AI adoption across the United States and six European countries, surveying both workers and firms in two waves during 2025 and 2026, and found that 43% of American workers now use AI on the job. In Europe, the number is 32%.
The gap is real, it is widening, and the study’s most important finding is about what explains it.
What Doesn’t Explain the Gap
The researchers tested every obvious candidate. They ran an Oaxaca-Blinder decomposition, a standard technique for attributing differences between groups to observable characteristics, looking at worker demographics, education levels, occupation, industry composition, and firm size. These factors account for roughly 55% of the US-Europe adoption difference.
That leaves nearly half the gap unexplained by the characteristics of the workforce itself. The workers in France or Germany who are not adopting AI are not, on average, less educated, less digitally literate, or working in less AI-relevant industries than their American counterparts. Something else is happening.
What Does
The study asked workers three additional questions: whether their employer encouraged them to use AI, whether their employer provided access to AI tools, and whether their employer offered AI training. The answers changed the analysis.
Among workers who did not receive AI training or tool provision, 47% adopted AI if their employer actively encouraged it. Among those who received no encouragement, adoption was 10%. That is not a marginal difference. It is the difference between an organization where AI is part of how work gets done and one where it remains a personal experiment.
The researchers then tested each factor independently. AI encouragement is the strongest predictor of adoption. Tool provision matters too, though the effect is smaller: 21% of workers who received tools but no encouragement still adopted, compared to 10% with neither. AI training, by contrast, does not independently predict adoption once you control for encouragement and tool access. A company can train its entire workforce on AI, and if no one in leadership actively promotes using it, the training does not translate into adoption. The distinction matters because some organizations have tried to drive adoption by tracking AI token consumption as a performance metric, which rewards volume rather than the kind of purposeful use that this study links to actual outcomes.
The study found a correlation of 0.81 between firm management quality scores and AI adoption rates, one of the highest in the dataset. But when the researchers controlled for AI encouragement and tool provision, the management index itself no longer predicted adoption. The effect of good management operates through a specific channel: telling people to use AI and giving them the tools to do it.
The geographic pattern is stark. In the US, 42% of workers receive both encouragement and tool access. In France and Italy, 69% to 70% receive neither. When employer encouragement is added to the decomposition model, it explains more than 95% of the US-Europe adoption gap in five of the six countries studied.
The Productivity Payoff
The adoption gap matters on its own. But the study connects it to something bigger. It connects adoption rates to productivity outcomes at both the individual and industry level.
Workers who use AI report saving an average of 5.8% of their work hours, roughly 2.3 additional hours per week for someone working 40 hours. The most common use is writing communications, selected by 55% of AI users and rated as the most useful task by 21%. Searching for facts ranks second at 51%, followed by interpreting, translating, or summarizing at 46%.
Those individual time savings are consistent with a growing body of experimental evidence. The researchers cite studies across software development, legal work, consulting, writing, customer support, radiology, and taxi routing, with productivity gains ranging from 5% to 96% and averaging 31% across all studies. The self-reported 5.8% figure from a nationally representative survey is lower than the experimental average, which is expected since experiments tend to study contexts where AI gains are most plausible.
At the industry level, the relationship between adoption and productivity growth is statistically significant and survives a placebo test. Using data from 29 European countries, the researchers found that a 10-percentage-point increase in firm AI adoption is associated with 1.08 to 2.59 additional percentage points of annual productivity growth from 2022 to 2024. When they ran the same analysis on pre-AI data from 2015 to 2019, the coefficient was small and statistically insignificant, which is what you would expect if the relationship is specific to AI rather than reflecting some pre-existing industry trend.
For the US, where comparable firm-level adoption data is unavailable, the researchers used worker-level AI adoption by industry and found a similar pattern. A 10-percentage-point increase in worker AI adoption is associated with approximately 1.1 percentage points of additional annualized productivity growth from 2022 to 2025.
Scaling these estimates to the actual US-Europe adoption gap of 11 percentage points implies roughly 4.1 percentage points of additional cumulative productivity growth for the US relative to Europe since 2019. That is a meaningful number in a macroeconomic context where annual productivity growth typically runs between 1% and 2%. It also reinforces a pattern emerging from other data: the companies capturing AI’s productivity upside are the ones that measure specific use cases, not the ones that simply report adoption numbers.
The Employment Question
The study also examined whether AI adoption is associated with employment changes, a question that generates more headlines than most. The answer, at least so far, is no. The researchers ran the same industry-level regressions with employment growth as the dependent variable and found negative coefficients, but none that are statistically significant. Industries with higher AI adoption are not, at this stage, showing measurable employment declines.
This is consistent with other recent research. A 2025 study of Danish workers found no meaningful impacts on employment or earnings in occupations particularly exposed to generative AI. The broader academic literature using pre-2020 data shows mixed results depending on the time period and measure of AI exposure.
The researchers are careful to note that this finding comes early in the adoption cycle. The absence of employment effects now does not guarantee their absence later. But the data available today does not support the narrative that AI adoption is currently destroying jobs at the industry level.
What This Means
The signal here is direct. The organizations that will capture AI’s productivity upside are not the ones with the best tools or the largest budgets. They are the ones where someone in leadership made a specific decision: encourage adoption, provide the tools, and set the expectation.
The data shows that this decision accounts for more of the variation in AI adoption than demographics, industry, firm size, education, or even the quality of the management team as measured by standard indices. It is not a general “culture of innovation” that drives adoption. It is a concrete set of actions: telling workers to use AI, giving them access, and creating an environment where experimentation is expected rather than tolerated.
For a CEO looking at this research, the diagnostic question is straightforward. Have you explicitly told your organization to adopt AI, provided at least one tool, and made clear that using it is part of how work gets done? If the answer is yes, the data suggests you are likely already seeing adoption rates in the 40% to 50% range. If the answer is no, the data suggests that no amount of training, hiring, or technology spending will close the gap until that directive exists. And once the directive is in place, the next question is where to point it: the first AI delegation a CEO makes tends to set the ceiling on every pilot that follows.
The productivity stakes are not abstract. Industries where adoption is higher are already growing faster, and the gap between adopters and non-adopters appears to be widening in real time. The management decision to encourage AI use is not just a cultural preference. It is, according to the best data available, the single most consequential variable determining whether an organization participates in the productivity gains that AI is beginning to deliver.