What Makes Exposure Usable at Work
The European evidence shows that occupational exposure is only the starting condition for generative-AI adoption. This chapter asks what happens inside an occupation when workers face broadly similar technical potential but do not have identical capabilities, tasks, or authority over their work. The practical judgment is deliberately specific: an exposed job becomes more adoptable when people can evaluate and integrate outputs and when the workplace gives them room and legitimacy to change how work is organised—not simply when a tool is available.
Education and abstract work shape the conversion
Using generative AI at work is not just a matter of opening a tool. The user must decide which problem is suitable for it, describe the problem well enough to get a useful response, recognise an error, and fit the result into a process that may carry professional or organisational consequences. Those activities create learning and evaluation costs even when the interface looks simple. Formal education can therefore matter as a form of absorptive capacity: it may make it easier to understand a tool’s limits, transfer existing knowledge to a new workflow, and judge whether an output is good enough to use.
The European study tests this idea by looking at how education changes the exposure–adoption relationship within occupations. Tertiary education steepens the gradient when entered on its own and retains an independent interaction in the joint model, where its coefficient is +0.027. In plain terms, workers with tertiary qualifications show a stronger relationship between the exposure of their occupation and their probability of reporting AI use than otherwise comparable workers in the same broad occupational setting. The estimate is a conditional association, not evidence that obtaining a degree mechanically produces adoption.
That qualification matters for both workers and employers. A degree is a proxy for a bundle of resources—formal knowledge, familiarity with complex information, access to professional networks, or confidence in learning new systems. It is not a complete measure of AI skill, and it is not the only way to acquire the relevant abilities. Treating the coefficient as a credential prescription would be a category error. The evidence says that formal human capital is associated with conversion; it does not establish that workers without tertiary education cannot become effective users or that a degree alone makes use safe or productive.
The content of the job matters alongside the person’s background. The study separates abstract task content from routine task content. Abstract work involves activities such as solving problems, learning new things, handling complex tasks, using foreign languages, and making difficult decisions. These are not synonymous with “office work,” and they are not simply a measure of intelligence. They describe what a job asks the worker to do. In the joint model, abstract task content remains a significant moderator of the exposure–adoption gradient, with an interaction of +0.019.
Why would abstract work make exposure more usable? Generative AI can be inserted at several points in a task that already requires the worker to interpret information, compare possibilities, formulate language, or revise a solution. The tool may help produce a draft or organise material, but the worker still has a reason to inspect the result against a problem they understand. That creates an entry point for augmentation: the worker can use the system without handing over the whole responsibility for the task.
Routine work presents a less predictable relationship. Short, repetitive activities may look easy to automate, but a highly routinised job can also leave little room for a worker to experiment with an unapproved tool. A fixed process may make automation technically attractive while making individual adoption organisationally difficult. The European estimates reflect that ambiguity: routine task content does not significantly moderate the exposure–adoption gradient once the relevant conditions are considered. The absence of a significant moderation effect does not mean routine work is unaffected by AI; it means that routineness alone did not explain why exposure translated into reported use more strongly for some workers than others in this analysis.
The same distinction applies to discretion. Task discretion measures control over the order, methods, and speed of immediate work. It is positively associated with adoption when examined separately, but it does not retain independent significance in the joint model. That result suggests that the ability to choose how to perform a task may be entangled with the kind of work and organisational influence the worker already has. A person can choose the sequence of their own tasks yet have no say in whether an AI system is introduced, what data it may receive, or how its output will be judged.
This is why the study’s interaction design is important. The exposure score is attached to the occupation, so adding occupation fixed effects means the main effect of exposure is absorbed. The interaction asks whether workers who share the same occupation-level exposure differ in adoption according to education, task content, or another moderator. It does not eliminate every form of sorting, but it narrows the comparison. The question is no longer simply whether highly educated people work in more exposed occupations. It is whether, among people facing the same broad occupational potential, education or job content is associated with a steeper conversion into use.
That is a more useful question than asking whether one demographic group is “good at AI.” It connects a worker characteristic to a specific opportunity. A person’s knowledge becomes relevant because the job presents a possible use, and the possible use becomes relevant because the person must evaluate and integrate it. The relationship is conditional in both directions: capability without a suitable task may have little workplace expression, while exposure without the capacity to judge outputs may create little usable adoption—or adoption with greater risk of error.
The diagram is not a claim that these components operate as a fixed pipeline or that any one of them is a guaranteed trigger. It is a way to keep the interfaces visible. Occupational exposure describes potential; the worker brings knowledge and judgement; the job supplies concrete places where a tool might fit; and the workplace determines much of the authority, support, and accountability surrounding use. Adoption is observed only after these elements meet.
The workplace decides whether use is usable
Individual capability does not operate in a vacuum. A worker may discover a promising use but still need time to test it, permission to handle information, colleagues who can compare results, or a manager willing to accept a new way of doing the work. The organisation can act as a gatekeeper through restrictions, or as an accelerator through procurement, training, endorsement, and involvement in work redesign. The evidence in the European study is most consistent with the latter being important, while leaving the direction of influence unresolved.
The clearest organisational signal is organisational influence, which is different from control over the minute-by-minute performance of a task. The survey measure captures whether employees are consulted before objectives are set, involved in improving work processes, able to influence decisions important for their work, and able to apply their own ideas. These are forms of voice over the organisation of work. They give workers a route into decisions about how a new tool should be introduced and improved, not merely permission to choose a shortcut inside an unchanged process.
Organisational influence steepens the exposure–adoption gradient in both the separate and joint models. The interaction is +0.025 when considered separately and +0.016 when the other moderators are included. It remains significant after accounting for education, job content, computer use, industry, workplace size, and other controls. The effect should still be read as an association. The survey cannot show whether employee voice causes adoption, whether workers who experiment gain more influence, or whether both are consequences of a broader management system that supports learning and digital change.
The distinction between voice and task discretion has a practical consequence. Suppose two workers can both decide the order in which they complete their assignments. That does not mean they can both ask for an approved AI tool, alter a quality-control step, or challenge a manager’s decision about where generated text enters a process. Immediate autonomy concerns the execution of assigned work. Organisational influence concerns the ability to shape the conditions under which work is assigned, evaluated, and changed. The study’s joint results point more strongly to the second relationship.
Workplace size illustrates why a visible correlate is not always an independent mechanism. Large workplaces may have more resources for procurement, information technology, and internal knowledge-sharing. They may also have more formal restrictions and more layers between a worker’s experiment and an authorised deployment. In the European results, workplace size is significant as a separate moderator but does not survive the joint specification. Size therefore helps describe the profile of early adopters, but it is not enough to explain why exposure converts into use once overlapping worker, job, and organisational conditions are considered.
Training has a similar two-level pattern. Recent employer-provided or self-financed training steepens the gradient when entered separately, with an interaction of +0.040. Its independent contribution is not statistically significant in the joint model, plausibly because training is related to education, abstract work, and organisational influence. This does not make training irrelevant. It means that a training-rich worker is often also situated in other conditions that support adoption, so the survey cannot cleanly assign the observed relationship to training alone.
The career and management trade-off is visible here. Giving workers more influence can improve the fit between a tool and real work, but it also distributes responsibility for experimentation. Employees may value autonomy and the chance to build new capabilities; employers may value speed and local knowledge while worrying about confidentiality, quality, or inconsistent practice. A workplace that invites experimentation without clarifying accountability can create informal use that is fast but fragile. A workplace that controls every use may reduce immediate risk while preventing workers from learning where the technology can or cannot help. The evidence supports examining that balance, not promising that one organisational style always wins.
Worker reports and enterprise reports reinforce this point without making the causal direction clear. Across 303 industry-by-country cells, worker-reported generative-AI use and enterprise-reported AI use are strongly correlated, at r = 0.86. The measures nevertheless differ. Enterprise reporting covers a broader set of AI technologies and reflects the organisation’s deployment; worker reporting is anchored to visible tools and the person’s own use. A high enterprise rate can include backend systems that workers do not notice. A high worker rate can include informal, employee-initiated use that a firm survey does not register.
Those mismatches are not merely measurement defects. They reveal different adoption states. A worker may be experimenting independently, a manager may be encouraging use without formal procurement, or an organisation may have deployed an automated process that changes work without asking employees to operate a generative system directly. Each state has different implications for learning, accountability, and bargaining power. The strong aggregate correlation suggests that individual use is connected to firm-level deployment, but it does not tell us whether firms create the conditions for use, workers pull formal adoption into the organisation, or both respond to common digital and managerial conditions.
The safest diagnosis of a low-adoption exposed team is therefore not “the workers are resistant.” Start by asking what crosses the boundary between potential and practice. Can workers access a suitable tool? Do they have time and training to evaluate its output? Can they raise a use case or propose a change to the process? Is there a clear owner for data handling and quality? Are they rewarded for improving the workflow, or exposed to blame if an experiment fails? These questions turn the conversion model into observable workplace evidence without pretending that the survey has tested each intervention.
The same diagnosis also prevents an employer from treating a high adoption rate as proof of success. Reported use may be shallow, informal, or concentrated among workers already able to navigate the organisation. A tool can be widely available while the people most affected by its outputs have little say in how it is used. Adoption is more usable when the worker can evaluate the result and the workplace can learn from the experiment. That combination—capability plus influence—explains why the conversion problem cannot be solved by procurement alone.
The evidence now points beyond the individual workplace. Workers with education and abstract tasks convert exposure more readily; employees with organisational influence do so even after overlapping conditions are considered; training and workplace size help describe the environment but do not operate as simple independent levers. Yet these opportunities are not distributed across national labour markets. The wider question is how computers, training systems, and the composition of jobs determine whether workplaces have the basic capacity to make exposure usable at all.