Which Occupations Have the Highest AI Exposure Score? A 34-Country Comparison

Why occupation matters more than country

Most coverage of AI and jobs focuses on country-level impact: which economies are most exposed, which will adapt fastest, which face the worst outcomes. That framing is useful for policy - but it is the wrong lens if you are trying to understand your own risk.

WorldJobsData covers 34 countries using consistent ISCO-08 occupation codes, which means we can compare the same occupation group across every country in the dataset. What that analysis shows is a clear pattern: the occupation you work in predicts your AI exposure score more reliably than the country you work in.

A Vietnamese clerical worker (Vietnam country average: 3.21/10) has a higher personal AI exposure score than a German agricultural worker (Germany country average: 5.3/10). The country averages move because of the occupation mix - how many people work in high-exposure roles versus low-exposure roles. But the underlying occupation scores are largely consistent everywhere.

Understanding which occupations score highest - and why - is the most direct way to assess your own risk.

The occupations with the highest AI exposure scores

The following scores are drawn from WorldJobsData's analysis of official ILO employment data across 34 countries. Occupation groups follow the ISCO-08 standard used by the International Labour Organization.

Why clerks score 9.0/10 - not 10.0/10

The most common question about the 9.0/10 score for general clerks is: why not 10.0/10? If data entry and document filing are almost entirely automatable, what is the remaining 10%?

The short answer is exception handling. Every automated system encounters edge cases that it cannot process without a human decision - an ambiguous document, a system error, a customer complaint that falls outside the defined workflow, a data entry that does not match the expected format. General clerks spend most of their time on routine tasks (hence 9.0/10) but retain value as exception handlers and as the human interface when automated systems fail. A score of 10.0/10 would imply no remaining human role under any current or near-future scenario, which is not the case.

The 9.0/10 score does, however, mean that the routine core of the job - the majority of the hours worked - is directly replicable by current AI and automation tools. Employers who want to reduce headcount in clerical roles do not need to wait for future AI development. The tools to do it exist today.

The ICT professional paradox

The finding that consistently surprises people is the ICT professional score: 8.5/10 in 28 of 34 countries. Technology workers - the people who build and deploy AI systems - are among the most exposed to AI displacement themselves.

This is not a contradiction. It reflects what ICT professionals actually do. A significant portion of software development involves writing structured, pattern-following code; debugging to documented specifications; generating boilerplate; and translating business requirements into technical specifications. These are precisely the tasks that large language models excel at. GitHub Copilot, Cursor, and Claude can already produce working code for a substantial share of standard software development tasks.

The parts of the ICT role that score lower on exposure are the parts that require architectural judgment under novel constraints, cross-functional communication with non-technical stakeholders, and domain-specific decision-making that AI cannot yet reliably handle. Senior architects, principal engineers, and AI/ML researchers have lower effective exposure than junior developers doing ticket-based implementation work - but the occupation group averages across all seniority levels, which produces a high overall score.

The occupations with the lowest AI exposure scores

The occupations that score lowest on AI exposure share a structural characteristic: they require physical presence in variable, unpredictable environments, direct human interaction, or manual dexterity that current robotics cannot reliably replicate at the price points that would make automation economically viable.

Why trades workers score so low despite high wages

The building trades result is one of the most counterintuitive findings in the dataset. In Germany, an electrician earns approximately $45,971 USD (OECD Average Annual Wages, 2024) and scores 2.0/10 on AI exposure. That combination - high earnings and near-zero AI risk - contradicts the assumption that high-paying jobs must be either knowledge work or managerial roles.

The reason is structural. AI cannot install electrical wiring in a building that does not yet exist. It cannot adjust plumbing to the unexpected pipe configuration behind a wall that was opened this morning. It cannot assess whether a roof structure is safe to walk on in the specific weather conditions of this Tuesday afternoon. These tasks require physical presence, immediate environmental assessment, and decision-making under conditions that vary every single job. Industrial robotics can automate specific, repetitive physical tasks in controlled factory environments. Construction sites are not controlled environments.

The economic protection this provides is substantial and underappreciated. Germany, the Netherlands, Sweden, and Switzerland all face genuine shortages of qualified tradespeople - plumbers, electricians, carpenters, HVAC technicians. A qualification in any of these trades combines the lowest AI exposure scores in the dataset with labour markets that have more demand than supply. The personal care picture is similar in every aging European economy.

How much do scores vary by country for the same occupation?

Less than you might expect. The ISCO-08 scores are largely stable across countries because they are driven by task content, and the same job title involves broadly the same tasks everywhere. A data entry clerk in Vietnam enters data; a data entry clerk in Germany enters data. The AI exposure of that task does not change based on geography.

What does vary by country is risk velocity - how fast AI tools actually get deployed into workplaces. Germany has a Risk Velocity score of 9.6/10, meaning it has the enterprise infrastructure, broadband penetration, and capital investment patterns to deploy AI at scale faster than almost anywhere else. Vietnam has a Risk Velocity of 1.2/10, meaning even highly exposed jobs there face slower actual displacement because adoption is slower.

This is why the country-level weighted averages differ so significantly (Vietnam 3.21/10 vs Singapore 5.62/10) despite the underlying occupation scores being similar. The weighted average reflects the occupation mix - how many workers are in which occupation categories - and the risk velocity adjustments that slow or accelerate actual deployment.

What the gap between 9.0 and 1.5 means in practice

The 7.5-point gap between the highest and lowest occupation scores in this dataset is not just a statistical observation. It translates into a concrete difference in trajectory for workers in those roles.

A 9.0/10 score for general clerks means that the majority of tasks in that role can be replicated by tools available today - not in 2030, not when AGI arrives. Microsoft 365 Copilot, Google Gemini for Workspace, and a dozen specialist tools already automate the core of the data-entry, correspondence, and document-handling tasks that make up most of a clerical worker's day. The question for these workers is not "will this happen?" but "how long until my employer deploys it and how can I position myself in the meantime?"

A 2.0/10 score for building trades workers in Germany means the opposite. Not only is there no technology today that can replace an electrician doing new-build wiring - there is no plausible technology roadmap in the next ten years that makes this economically viable at scale. The job is structurally protected, it is in structural shortage, and it pays well. The challenge for these workers is not AI - it is keeping skills current with evolving materials, building codes, and energy systems.

The occupations in the middle: what 5.0-7.0/10 means

The most uncertain territory in the dataset is the 5.0-7.0/10 range. These are occupations where AI exposure is real and growing but where the human elements of the job are significant enough to create genuine ambiguity about the outcome.

Teaching professionals (3.0-4.0/10 in most countries) are the clearest example of this middle ground. AI can generate lesson plans, grade multiple-choice tests, provide personalised practice exercises, and answer student questions on a large scale. These are legitimate productivity gains for teachers. But classroom management, identifying which student is struggling and why, adjusting instruction in real time to a room full of different emotional states, and building the trust relationships that make learning possible - these are not automatable tasks, and they represent a large share of what effective teachers actually do.

Health professionals (2.5-3.5/10 for non-admin roles) are in a similar position. AI diagnostic tools already outperform radiologists on specific image-recognition tasks in controlled settings. But the clinical consultation, the patient relationship, the judgment call about when a test result means something different for this patient than it would for the average patient - these remain human. The overall score reflects that the administrative burden of healthcare (scheduling, documentation, billing) is highly exposed, while the clinical core is much less so.

What this means if you are in a high-exposure role

If your occupation scores above 7.0/10, three things are worth knowing.

First, the timeline is shorter than most people assume. The tools that automate the core tasks of clerical and administrative work are not in development - they exist and are being deployed by large employers now. The question is not "will this happen in five years?" but "is my employer already evaluating this?"

Second, within any high-exposure occupation there are elements that score differently. A financial clerk who handles exceptions, communicates directly with clients about complex situations, and makes judgment calls about edge cases is less exposed than a financial clerk who processes standard transactions. Moving toward the exception-handling and judgment-requiring elements of a role buys time and reduces immediate exposure.

Third, the adjacent safe options are identifiable. The skills adjacent to clerical work that are not highly AI-exposed include direct client relationship management, physical service delivery, care work, and trades - not technology, which also scores 8.5/10. The transition is not easy, but the destination is identifiable from the data.