Predictions about the workplace impact of artificial intelligence often begin with what the technology can do captured by occupations’ exposure to AI. Using representative data on German workers, this column shows that adoption also depends on whether AI is worth using relative to human labour. Comparative advantage is a much better predictor of AI adoption than absolute advantage based on exposure alone.
Understanding how widely and where artificial intelligence will be adopted, and what its effects will be, has become a central concern in debates about employment, wages, productivity, and inequality (Acemoglu 2025, Bick et al. 2026, Hampole et al. 2025).
Every few months, a new AI model arrives that appears dramatically more capable than the last. As new tasks become feasible, predictions of rapid workplace transformation abound. Yet despite extraordinary progress in AI capabilities, adoption has proceeded more slowly than many expected.
Why? A common explanation is that current AI systems still cannot perform enough economically valuable tasks. According to this view, adoption remains limited because the technology is not yet capable enough.
The standard way to measure AI’s economic impact reflects this logic. Most existing measures focus on exposure. These measures ask whether AI is capable of performing the tasks associated with a particular occupation. If an occupation contains many AI-compatible tasks, it is classified as highly exposed (Felten et al. 2021, Eloundou et al. 2024). But exposure measures what AI can do, not who will actually use it. Understanding who adopts AI is a necessary first step for assessing the labour market effects of this new technology.
In our paper (Lindenlaub et al. 2026), we use representative data on AI adoption among German workers to show that capability is only one side of the equation. We find that the main obstacle to AI diffusion is often not what AI can do, but the costs associated with using it.
Workers and firms do not adopt AI simply because it is technically capable. Adoption requires that AI be worthwhile to use. That is, AI is adopted based on its comparative advantage vis-à-vis a worker, not based on its absolute advantage in task performance. Put simply, the relevant comparison is how much output AI delivers per dollar of user cost versus how much output a worker delivers per dollar of pay. To understand this distinction, imagine a task for which AI performs extremely well. Even when AI is highly capable, adoption may remain limited if workers must spend substantial time verifying outputs, privacy regulations restrict deployment, workflows need to be redesigned, organisational approval is difficult to obtain, or workers themselves are highly capable of performing the task at low cost.
We develop a framework that separates three forces shaping AI adoption: AI productivity, AI user costs, and worker productivity relative to pay. AI productivity reflects what AI can do and, thereby, exposure. User costs reflect the practical burden of deploying AI in production. Worker productivity relative to pay captures the value of existing human labour, meaning how cost-effectively workers can perform tasks without AI.
We estimate all three components by combining direct measures of worker AI use with administrative labour market records and occupation-level task information.
We use the 2024 wave of the Digital Transformation and the Changing World of Work survey (DiWaBe 2.0), a nationally representative survey of 9,835 German workers that records actual workplace AI use and is linked to official worker and establishment records. The combination of reported AI adoption, broad occupational coverage, and reliable administrative records gives us a detailed picture of AI adoption across workers and occupations.
The results are striking. The tasks where AI is most capable are not necessarily the ones where it is most worth adopting, either because user costs are high or workers remain relatively productive relative to their pay. Understanding this gap between what AI can do and where it is cheap and easy to use makes AI adoption far more predictable. An exposure-only model explains just 25% of the observed variation in AI adoption across occupations. Once we allow user costs to differ across tasks, explanatory power increases dramatically. Allowing worker heterogeneity improves the fit further, raising the explained variation to about 60%. Both margins are important for understanding where AI is adopted.
This finding helps explain several puzzles (see Figure 1). Some occupations, like accounting, face massive technical exposure to AI, meaning that AI could automate many of their job tasks. And yet, steep costs of verifying the AI’s work as well as privacy concerns, combined with relatively high levels of productivity among accountants, temper the technology’s true comparative advantage. Conversely, teaching involves core tasks with lower AI exposure but also lower user costs – giving AI a compelling relative edge and prompting surprisingly robust adoption.
Figure 1 Absolute and comparative advantage of AI across occupations
More broadly, exposure and our comparative-advantage measure give opposite high-versus-low adoption predictions for occupations accounting for roughly 30% of employment. This divergence matters for policy. Discussions about who may gain, who may need retraining, and where adjustment costs may arise can focus on the wrong groups if they treat exposure as adoption. Measuring adoption is therefore foundational to debates about AI’s labour market impact.
The implications are important for forecasting the future of AI. Many discussions implicitly assume that adoption will be driven primarily by continued improvements in AI performance and thus increased exposure. Our framework instead separates improvements in AI productivity from reductions in user costs, allowing us to examine how each shapes diffusion.
We use the model to project future adoption over the next three years. The share of workers adopting AI rises from 44% at baseline to about 81%, while the intensity of use among adopters changes only modestly. Decomposing this increase shows that reductions in user costs play a much larger role than productivity improvements. As AI tools become easier to integrate into workflows, easier to verify, and easier to deploy safely, adoption could accelerate even in occupations where AI capabilities themselves improve only modestly. That aggregate movement is also uneven across jobs. Growth is fastest among occupations currently in the middle of the adoption distribution, while lower user costs also bring AI into occupations that appear largely untouched today.
The broader lesson is that technological revolutions are not governed solely by invention. They are also governed by implementation. History offers many examples of technologies whose capabilities arrived well before widespread economic adoption. AI may prove similar. The next phase of diffusion will depend on reducing the organizational, regulatory, and practical barriers that prevent firms and workers from using AI effectively. The question is therefore not just what AI can do; it is whether using AI is worth it.
Source : VOXeu
Foreign investors’ growing holdings of dollar-denominated bonds have created a large demand to hedge dollar…
The Turkish central bank cut interest rates over 2021-2023, on the politically driven view that…
Firms in low- and middle-income countries persistently pay more to borrow than firms in high-income…
Inflation in the US has receded from its 2022 peak. This column uses payroll records…
The US Inflation Reduction Act’s Energy Communities provisions aimed to increase clean-energy investment in areas…
For its entire 15-year life as a regional bank until its collapse in 2023, Silicon…