Which Three Jobs Will Survive AI? What the Research Actually Says

The question has been asked in every office, classroom and dinner table conversation since ChatGPT made AI real for non-technical people in late 2022. Which jobs are safe? What can I do that a machine cannot? Where should my children be heading? The answers offered in much public discourse have ranged from the panicked — "AI will take everything" — to the complacent — "don't worry, humans are irreplaceable." Neither extreme is supported by the research.

The labour economics literature on automation and employment is extensive and methodologically sophisticated. It goes back well before the current AI wave, through debates about robotics in manufacturing, computerisation in clerical work, and the general question of what technologies do to labour markets over time. That literature, now updated with specific research on large language models and the current generation of AI systems, offers answers that are grounded, qualified and — perhaps surprisingly — consistent across different research groups and methodological approaches.

This piece synthesises the best available evidence on which categories of work are most resistant to AI-driven automation, explains why, and draws the practical implications for individuals, families and educational institutions.

What the Research Tells Us About Automation Risk

The foundational study remains Frey and Osborne (2017), originally published as an Oxford Martin School working paper in 2013 and revised for Technological Forecasting and Social Change. Frey and Osborne assessed 702 occupations in the United States for susceptibility to computerisation and found that 47% of US employment was at high risk — defined as greater than 70% probability of automation within ten to twenty years. The study identified three clusters of skills that were most resistant to automation: perception and manipulation in unstructured environments, creative intelligence, and social intelligence (Frey and Osborne, 2017).

The OECD's response, by Nedelkoska and Quintini (2018), applied a more granular methodology using task-level data rather than occupation-level data and arrived at a lower but still substantial estimate: 14% of OECD jobs at high risk of automation, with 32% facing substantial change and 66 million workers across 32 countries affected. The World Economic Forum's Future of Jobs Report 2025 projected 22% churn in the global labour market by 2030, with 170 million new jobs created and 92 million destroyed — a net positive of 78 million, but with enormous disruption to individuals and communities in the process (WEF, 2025).

Goldman Sachs (Hatzius et al., 2023) estimated that generative AI could expose 300 million full-time equivalent jobs globally to some degree of automation, while potentially adding 7% to global GDP over ten years. McKinsey's Generative AI and the Future of Work (2023) put the economic value at $2.6 to $4.4 trillion per year and suggested that 60 to 70% of employees' working time involved activities with potential for technical automation. The IPPR's Transformed by AI (2024), examining the UK specifically, found that 11% of tasks were already exposed in the first wave of AI, rising to 59% in a second wave, with up to 8 million UK jobs at risk in a worst-case scenario.

The First Job Category: Skilled Trades in Unstructured Physical Environments

The single category of work that appears most consistently at the bottom of automation risk tables — that is, the work least likely to be displaced by AI — is skilled manual work in unstructured physical environments. Plumbers. Electricians. Gas engineers. HVAC technicians. Carpenters. Roofers. Specialists in building maintenance, installation and repair.

Goldman Sachs (2023) found that building and grounds cleaning occupations had 1% task exposure to AI automation, installation, maintenance and repair occupations had 4%, and construction and extraction occupations had 6% — among the lowest rates across the entire economy. Frey and Osborne (2017) identified first-line supervisors of mechanics, installers and repairers as having a 0.3% probability of computerisation — the lowest among all 702 occupations studied.

The reason is structural rather than incidental. AI systems — whether software or robotic — perform best in structured, predictable, well-defined environments. The physical world that skilled tradespeople work in is none of these things. Every building is slightly different. Every plumbing fault presents novel spatial constraints. Every electrical installation requires judgement about what a previous engineer intended, what current regulations require, and what the specific circumstances of this building and this customer demand. The dexterity, spatial reasoning and contextual judgement required have proven far more resistant to automation than the most cognitively demanding desk-based work.

The Construction Industry Training Board consistently projects significant shortfalls in trained tradespeople in the UK in the years ahead, driven by an ageing workforce and insufficient apprenticeship uptake. The irony of the AI moment is that the jobs that have historically carried lower social prestige than graduate professions are among those most likely to remain indispensable — and among those that will command increasing wages as their scarcity grows.

The Second Job Category: Healthcare Practitioners with Direct Patient Contact

Frey and Osborne (2017) found that registered nurses had a 0.9% probability of automation, recreational therapists 0.28% — the lowest of all 702 occupations — occupational therapists 0.35%, and physicians and surgeons 0.42%. The ONS (2019), applying a similar methodology to the UK labour market, found medical practitioners among the very lowest risk occupations in England. McKinsey (2023) projected that healthcare professionals would see the largest absolute job growth of any sector in the US by 2030, adding an estimated 3.5 million roles. The WEF (2025) listed care workers among the top roles growing by absolute numbers globally.

The reasons are multiple. The physical element of healthcare — examination, treatment, surgery, physiotherapy — involves precisely the kind of embodied manipulation in unstructured environments that resists automation. The relational element — building trust with a patient, responding to distress, understanding what someone is really saying beneath what they are expressing — involves social intelligence that current AI systems can approximate in text but cannot replicate in person. And the accountability element — a doctor or nurse is legally and ethically responsible for their decisions in a way that an AI system cannot currently be — means that for safety-critical clinical decisions, human professional judgement is structurally required.

The NHS Long Term Workforce Plan (2023) projected the need for 300,000 additional NHS staff in England by 2036/37, even accounting for full productivity gains from AI. This is not a profession contracting under automation pressure. It is a profession whose demand is growing faster than any conceivable automation-driven reduction in the number of roles it requires.

The Third Job Category: Early-Years and Primary Teachers, and Religious Educators

Frey and Osborne (2017) found that elementary school teachers had a 0.44% probability of automation, preschool teachers 0.74%, and special education teachers less than 1.2%. The ONS (2019) found special-needs teachers, headteachers and education advisers among the UK's lowest-risk occupations. The WEF (2025) listed teaching and educational professionals among the top growing roles globally by absolute numbers.

Teaching, at its core, is a relational act. The question "has this child understood this concept?" cannot be answered by observing whether they can reproduce the right output. It requires reading a child's face, responding to their confusion, adapting the explanation, managing the emotional dynamics of a classroom, and holding thirty children in mind simultaneously while attending to each one individually. These are not tasks that AI can perform. AI can provide content, practice problems, feedback on written work, and individualised pacing through educational material. It cannot replace the teacher as the locus of the learning relationship.

For Raedan Institute, there is a specific and important point here. Frey and Osborne (2017) found that clergy had a 0.81% probability of automation, and directors of religious activities and education 0.83% — among the lowest in the entire study. The work of an imam, a religious educator, a pastoral figure within a faith community involves precisely the combination of relational intelligence, accountability and moral authority that AI cannot replicate. The guidance of individuals through significant life events, the transmission of tradition in ways that make it alive rather than merely archived, and the pastoral care of a community in need are forms of work that will not be automated.

What About the Jobs That Are Most at Risk?

Intellectual honesty requires noting that alongside the jobs most resistant to automation, there are jobs that face genuine disruption. The WEF (2025) identified postal clerks, bank tellers, data-entry clerks, cashiers, administrative assistants and executive secretaries, and accountants and auditors as among the fastest-declining roles. Frey and Osborne (2017) found telemarketers, tax preparers and insurance underwriters at 99% probability of automation.

Crucially, the current AI wave has inverted the older displacement pattern. Previous automation waves — robotics in manufacturing, computer automation of clerical work — disproportionately affected lower-skilled, lower-wage workers. The current AI wave, as Eloundou et al. (2023) demonstrated in their analysis of large language models and the US workforce, affects higher-skilled, higher-wage white-collar work most directly. Paralegals, junior coders, junior consultants, market research analysts, junior copywriters — these are the roles whose tasks are most automatable by current-generation AI, and they are roles that have historically required significant education and carried professional status.

The Practical Implication: What Should Young People Do?

The evidence points clearly toward a number of practical conclusions. Skills in working with physical systems in unstructured environments will be valuable for the foreseeable future — skilled trades are not a second-best option but a strategically sound choice in a world being transformed by AI. Skills in human care — nursing, therapy, physiotherapy, social work, early years education — are in growing demand and resistant to automation. Skills in religious education and pastoral care remain as irreplaceable as ever.

At the same time, the most important skill of all may be adaptability — the capacity to work alongside AI tools rather than competing with them, to use AI to amplify human capability rather than being displaced by it. Brynjolfsson, Li and Raymond (2025) found that AI assistance raised the productivity of customer service workers by an average of 14%, with the largest gains — 34% — among novice workers who received AI guidance that helped them approach expert-level performance. This is the human-AI collaboration model, and it is where most of the economic value of AI will be captured in the near to medium term.

The future of work is not a simple story of replacement. It is a complex story of transformation — some roles disappearing, new ones emerging, most changing in their day-to-day content. Navigating that transformation well requires accurate information about where the risks and opportunities actually lie, and that information is now available. The research is clear. The question is whether educational institutions, employers and individuals will act on it.

References

Brynjolfsson, E., Li, D. and Raymond, L.R. (2025) 'Generative AI at work', Quarterly Journal of Economics, 140(2).
Eloundou, T. et al. (2023) 'GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models', arXiv preprint, arXiv:2303.10130.
Frey, C.B. and Osborne, M.A. (2017) 'The future of employment: How susceptible are jobs to computerisation?', Technological Forecasting and Social Change, 114, pp. 254–280.
Goldman Sachs (2023) The Potentially Large Effects of Artificial Intelligence on Economic Growth. New York: Goldman Sachs Economics Research.
IPPR (2024) Transformed by AI: How Generative Artificial Intelligence Could Affect Work in the UK and How to Manage the Transition. London: Institute for Public Policy Research.
McKinsey Global Institute (2023) Generative AI and the Future of Work in America. New York: McKinsey and Company.
Nedelkoska, L. and Quintini, G. (2018) Automation, Skills Use and Training. Paris: OECD Publishing.
NHS England (2023) NHS Long Term Workforce Plan. London: NHS England.
ONS (2019) The Probability of Automation in England. Newport: Office for National Statistics.
World Economic Forum (2025) Future of Jobs Report 2025. Geneva: World Economic Forum.