The Gradient Between Potential and Use
This chapter interprets the European evidence on the relationship between occupational exposure to generative AI and reported use at work. Its focus is not to decide which occupations “belong” to the AI economy, but to show how a task-based measure identifies potential, how that potential appears in adoption data, and why the gap between a raw association and a fully adjusted one is itself informative. The result is a better way to read the slope: as a conversion question, not as a causal law.
Exposure is measured through tasks, not tool ownership
An occupation can be exposed to generative AI even when none of its workers uses a generative-AI tool. Exposure begins with the content of the work. The relevant question is whether large language models could plausibly reduce the time needed for some activities, either directly or through software that incorporates them. This is a statement about technical fit between a capability and a task. It is not a statement about access, permission, confidence, quality, or intent.
The European study represents that potential with the Generative AI Susceptibility Index, or GAISI. The index applies an exposure rubric to task information from the OECD’s Programme for the International Assessment of Adult Competencies and weights the ratings by how frequently tasks occur. It then aggregates the result to broad, two-digit occupational groups. A higher score means that a larger share of the jobs in that occupational group contains tasks considered highly susceptible to current LLM capabilities.
That construction does two useful things. First, it moves the analysis below the vague claim that “knowledge work” is exposed. Exposure is tied to activities that workers perform, not to a job title’s prestige or to an industry’s reputation for innovation. Second, it creates a common yardstick for comparing occupations before asking whether workers actually use AI. Without such a yardstick, researchers could mistake high adoption for high technical relevance, even though the two may reinforce each other after adoption begins.
The measure is also deliberately narrower than a forecast of an occupation’s future. It does not predict that a task will disappear, that a worker will be replaced, or that an employer will buy a tool. A task can be technically susceptible while remaining difficult to automate because its output requires context, accountability, physical presence, or human judgement. Conversely, a worker can find a use for a general-purpose system in a job whose average task profile is not especially exposed. The index identifies a potential margin of change; it does not resolve the organisational decision about whether that margin will be used.
There is a second reason to be careful: GAISI is assigned at the occupational level. It is common to workers within a broad occupation–country cell rather than calculated from the exact workflow of each person. That makes it powerful for studying patterns across a large representative sample, but blunt for describing a particular workplace. Two people with the same occupational code can divide their time differently, use different systems, face different rules, or have different authority to alter their work. The index treats those differences as part of the adoption question rather than silently pretending they do not exist.
The adoption measure observes a different point in the process. In the 2024 European Working Conditions Survey, workers were asked whether, at work, they used artificial intelligence that simplified complex mental tasks or made recommendations. Examples such as ChatGPT, Llama, DALL-E, Midjourney, and Jasper helped anchor the interpretation toward generative tools. The question is useful because it asks about the worker’s own workplace use, but its wording is somewhat broader than generative AI in the strict technical sense.
It also measures only whether use occurred—the extensive margin. A “yes” can describe an occasional experiment, a recurring assistant for a narrow task, or a tool woven through a substantial workflow. It does not tell us how often the person uses AI, how much time it saves, whether the output is checked, or whether the system is approved by the employer. Nor does it capture an AI process operating in the background without the worker’s awareness. A worker-level adoption statistic and an enterprise deployment statistic therefore answer related but different questions: one concerns visible individual engagement, the other the organisation’s use of AI systems across its operations.
These distinctions change how the numbers should be read. Exposure is a technical opportunity measure. Worker adoption is a self-reported behaviour measure. Frequency and depth would be measures of intensity. Enterprise deployment would describe an organisational decision or capability. None of these automatically measures output quality or productivity. They can be connected in a workplace, but they should not be substituted for one another.
The study tests whether its exposure measure is overly dependent on one particular rating procedure. When the analysis is repeated with an alternative human-rater exposure score, the substantive pattern remains: exposure strongly predicts reported adoption, although the estimates are somewhat less precise. That robustness supports the broad conclusion that task susceptibility matters. It does not turn the index into a perfect measure of every worker’s technical opportunity. A measure can be reliable for comparing groups and still be incomplete for explaining individual behaviour.
The measurement sequence matters because it prevents a common inferential shortcut. The index does not cause a worker to answer yes, and the worker’s answer does not prove that the occupation has been transformed. The value of putting them together is to ask how much technical potential reaches actual use, and what explains the conversion rate.
The gradient is strong, and its attenuation is the point
The first answer is that exposure does organise adoption. In the pooled European workforce, reported use rose from 1.5% in the least exposed occupational quintile to nearly one quarter in the most exposed. The difference was 23.4 percentage points. Adoption was not zero in the least exposed group, which is consistent with generative AI’s general-purpose character, but the overall pattern was steep enough to show that technical fit was not incidental.
The regression estimates express the same pattern on a continuous scale. In the baseline specification, a one-standard-deviation increase in occupational exposure was associated with a 10.1 percentage-point increase in the probability that a worker reported AI use at work. For perspective, the pooled adoption rate was about 12%, so the unadjusted association is large relative to the early average. It would be tempting to stop there and say that exposure drives adoption. The sequence of controls explains why that would be too quick.
When the analysis adds education, seniority, and computer use, the estimated association falls to 6.6 percentage points. When it adds employment status, workplace characteristics, and industry fixed effects, it falls again to 4.9 points. The decline is not a technical nuisance. It tells us that exposed occupations contain a distinctive mix of workers and settings: people with different formal qualifications, different degrees of computer-mediated work, different employment relationships, and different kinds of workplaces. Some of what looks like an exposure effect in the raw comparison is a difference in who and what is located in exposed work.
This is best understood as a decomposition of the question, not as a sequence proving that one variable mechanically mediates another. Education and computer use are related to occupational exposure, but they are also related to many unmeasured features of a job. Workplace size and industry can stand for procurement capacity, management practice, client demands, or the kinds of data a worker handles. Once these variables are included, the coefficient answers a narrower question: how much association between occupational exposure and reported use remains after the specified observable differences are taken into account?
The answer is still positive and precisely estimated. That persistence is important, but its meaning is conditional. It does not show that exposure independently causes adoption. Workers may sort into jobs in ways the survey cannot fully observe. Employers may encourage, restrict, or ignore AI use. Workers may differ in their ability to evaluate outputs or in their willingness to accept the risks of entering information into a model. The occupation-level index also assumes that the task content of a broad occupation is reasonably comparable across European labour markets. Those limitations qualify the estimate without making it useless.
The estimate becomes more informative when we stop treating it as a ranking of occupations and read it as a slope. Imagine two settings with the same distribution of potentially exposed work. In one, exposed workers have frequent computer access, time to learn, and room to adapt their workflows. In the other, the same workers face rigid procedures, limited training, or uncertainty about whether AI use is acceptable. The exposure level may be similar, but the share of workers who adopt can differ. The slope captures that difference in conversion more directly than a list of “high-AI” jobs.
The distinction between a level and a slope is especially useful for interpreting the country differences reported in the European evidence. A country could have a modest amount of AI use spread evenly across occupations. Another could have a similar average but a much sharper concentration among exposed workers. Those situations imply different mechanisms. The first might reflect broad experimentation wherever tools are available; the second suggests that technical fit is being converted more effectively in the jobs where it should matter most. The conceptual point is already visible in the pooled gradient: the average adoption rate does not tell us how exposure is being translated.
A slope also makes room for low-exposure adoption. General-purpose systems can help with writing, searching, summarising, planning, or other occasional activities outside the occupations with the highest measured susceptibility. That does not invalidate the index. It means the index is answering a question about relative technical potential, while adoption reflects both broad tool availability and the particular problems a worker chooses to solve. A measure can predict a large difference between groups without predicting every individual case.
The same reasoning explains why the alternative exposure measure matters. If the gradient disappeared when the researchers changed how exposure was rated, the result might be an artefact of one index’s construction. Instead, the pattern remains, with somewhat different precision. The evidence is therefore stronger for the proposition that occupational task susceptibility is related to early uptake than for any exact claim about the size of that relationship.
The most defensible reading has three parts. First, technical exposure is a meaningful starting condition: it identifies where the incentive or opportunity to use generative AI is more likely to arise. Second, the raw relationship partly reflects the composition of exposed work and the people who occupy it. Third, a positive adjusted association remains, but the study’s cross-sectional design does not establish a causal effect of exposure on adoption. The gradient tells us where the conversion problem is concentrated; it does not, by itself, tell us which mechanism solves it.
That boundary is practical, not merely statistical. If a manager sees that an exposed team has low use, the result does not justify assuming that workers lack motivation. The missing condition could be training, access, authority, a workflow that leaves no room for experimentation, or a concern about quality and accountability. If a policymaker sees higher adoption in one country, the result does not justify copying a single programme without examining digital readiness and job structure. And if a worker in a low-exposure role uses AI, that use is not evidence that exposure measures are irrelevant. Each observation becomes clearer when potential, behaviour, intensity, and outcome are kept separate.
The gradient has therefore done its job when it leaves us with a sharper question. Occupational exposure helps explain where adoption is more likely, but the attenuation and the remaining association both point beyond the index. To understand why similarly exposed workers convert potential into use at different rates, we have to look inside the occupation—at what workers can evaluate, what their tasks demand, and how much influence the workplace gives them over changing the work itself.