The Safest Jobs from AI in 2026: Evidence from 33 Countries

What makes a job safe from AI?

The question of which jobs are safe from AI has generated a lot of speculation and not enough data. The WorldJobsData scoring model - applied consistently across 33 countries and 2.9 billion workers using ILO ILOSTAT employment data - gives a cleaner answer than most. Three structural factors, when present together, consistently produce low AI exposure scores.

The first is physical presence in unpredictable environments. Current AI systems are software. They can process language, generate images, classify data, and automate structured decision-making. They cannot wire a house, dig a foundation, harvest a crop in the rain, or physically lift and reposition a person with limited mobility. The moment a job requires a human body to interact with an unpredictable physical environment, current AI hits a hard wall. This is why building trades workers score 2.0/10 and agricultural workers score 3.0/10 - not because AI is irrelevant to their industries, but because the core of their work cannot be done without a body in the right place.

The second is direct human-to-human interaction that requires emotional presence. Personal care work - caring for the elderly, supporting people with disabilities, assisting with daily living - requires emotional attunement that is not separable from the physical care itself. A care worker adjusting how they communicate because a patient with dementia is distressed is doing something that AI scheduling software cannot replicate at the bedside. This is the reason personal care workers score 2.0/10 even as adjacent roles in healthcare administration score much higher.

The third factor is the economics of automation. Not every job that is technically automatable is economically worth automating. Smallholder farming across sub-Saharan Africa and South Asia involves such variable conditions, small plot sizes, and low wage levels that the return on deploying agricultural robotics does not exist at current technology costs. This structural economics protection is real but fragile - it erodes as technology costs fall. The other two protections are more durable.

Personal care workers: the most AI-proof major occupation group

Personal care workers (ISCO 53) score 2.0/10 on AI exposure. That is the lowest score in the entire WorldJobsData dataset across all 33 countries measured. To put this in context: general and keyboard clerks score 9.0/10. Customer service clerks score 8.5/10. Even craft and trades workers - the next-safest major group - score 2.5/10. Personal care workers are in a separate category.

The reason is structural. A care worker supporting an elderly person with dementia is performing physical repositioning, medication assistance, nutritional support, and real-time emotional response to a person whose needs change minute to minute and whose communication is often non-verbal. AI tools can help with rostering, record-keeping, and alerting. They cannot do the care itself. The physical and emotional labour is the job - and it is irreducibly human.

The demand signal for care workers points strongly in one direction. Skills for Care's 2024 State of the adult social care sector report found 131,000 vacancies in adult social care in England at any given time. The World Health Organization projects a global health and care worker deficit of 10 million by 2030 (WHO, 2023). These shortages are structural - driven by aging populations, not by oversupply. Japan alone needs an additional 570,000 care workers by 2040 (Japan Ministry of Health, Labour and Welfare).

This does not mean care work is without challenge. Wages in the sector are often low relative to the skill and emotional demand involved. But the AI risk is genuinely minimal - and the combination of low AI exposure and structural labour shortage gives care workers a position that very few other occupations can claim.

Craft and trades: safe and in demand

Craft and trades workers (ISCO 7) score 2.5/10 on AI exposure globally. Building trades workers specifically (ISCO 71 - electricians, plumbers, carpenters, structural steel workers) score 2.0/10 - matching personal care workers for the lowest AI exposure in the dataset.

The spatial complexity argument is real and specific. An electrician on a residential job is navigating a building where every run is different - different wall compositions, different distances between junction boxes, different load requirements, different local code interpretations. They are making hundreds of small spatial and judgement decisions per hour in a physical environment that has never been modelled digitally. Current AI cannot handle that task. Current robotics cannot either - not at the cost and reliability required for commercial deployment. This is why building trades score 2.0/10 on AI exposure even as AI adoption in the construction industry accelerates for design, estimation, and project management.

The demand signal is as strong as in care. Associated Builders and Contractors (2026) estimates the US construction industry needs 500,000 more workers on top of normal hiring to meet projected demand. Germany's Bundesagentur fur Arbeit (2024) reported a shortage of 86,000 apprentices in skilled trades - a structural gap that has been widening for a decade. An electrician in Germany earns approximately $46,000 USD per year with low AI displacement risk and structural demand. That combination is rare in any labour market.

Agricultural workers: scale and structural protection

Agricultural workers (ISCO 6) represent the single largest occupational group in the world - 525 million workers globally - and score 3.0/10 on AI exposure. This is the largest pool of structurally AI-resistant workers on earth.

Their protection comes from two distinct sources. The first is environmental variability. Outdoor agricultural work involves weather, soil variation, crop disease, animal behaviour, and biological systems that are genuinely unpredictable in ways that structured AI tasks are not. A corn harvester in Iowa operates in conditions different from a rice paddy in Vietnam, a smallholder plot in Nigeria, or a vineyard in Spain. The range of physical and biological variables is enormous.

The second source is economic. Automating large-scale industrial farming (combine harvesters, drone spraying, precision irrigation) is already happening and reflects in the robotics risk score of 6.5/10 for agricultural workers - meaningfully higher than the AI-specific score of 3.0/10. But this applies to a narrow subset of large-scale commercial agriculture. Nigeria has approximately 71 million agricultural workers (AI average 3.31/10 nationally) and India has approximately 476 million workers with a national AI average of 3.26/10. The economics of deploying autonomous agricultural machinery in these contexts do not exist at current technology costs. The smallholder farmer in these economies has structural protection that is not going away in the near term.

The important caveat is this: the robotics risk for agriculture (6.5/10) is much higher than the AI-specific risk (3.0/10). Agricultural workers are not safe from all forms of automation - they are specifically safe from AI-driven cognitive task displacement. Physical automation of large-scale farming is a different and real threat over a longer time horizon.

The safest jobs: data table

Sources: ILO ILOSTAT employment data. WorldJobsData AI and robotics exposure scores. Robotics score reflects physical automation risk separately from AI-specific cognitive task displacement.

What this means for you

If your occupation scores below 3.0/10 on AI exposure, you have structural protection for at least 8 to 12 years based on current AI capability benchmarks and observed deployment rates. The three safest groups - personal care, building trades, agriculture - share characteristics that are genuinely hard to replicate with current and near-term AI systems. This is not a prediction about 2040; it is an assessment of what is technically possible today and what the deployment timeline looks like based on observed adoption rates.

If your occupation scores above 7.0/10, the pressure is real and already materialising. General and keyboard clerks (9.0/10) and customer service clerks (8.5/10) are not facing a future risk - they are facing a present one. Anthropic's 2026 research shows 70% of customer service clerks' theoretical AI exposure has already been observed in practice. That is not a projection. It is a measurement.

The data suggests a clear direction for anyone evaluating their options: the safest moves are toward the three low-exposure groups. Not because those jobs are easy - personal care work and trades work are physically and emotionally demanding - but because the structural protection is real, the demand is structural, and AI cannot close the gap in any near-term scenario. Use the WorldJobsData explorer to check the score for your specific occupation group in your country.