A better question than "which jobs are safe"

Lists of AI-proof jobs circulate constantly and age badly, because they are usually assembled from intuition rather than measurement. The more durable question is why certain work resists automation, because the reason generalises to occupations that never appear on anyone's list.

Jobs and Skills Australia scored 998 Australian occupations at the four-digit ANZSCO level against two separate measures: whether generative AI could perform each task, and whether it could assist with each task. The occupations that score low on both are the interesting group, and the study identifies what they have in common.

The common factor is physical action in an unpredictable setting

JSA's finding is direct. Occupations with low automating and augmenting potential are those that "capture the presence of tasks that require physical action, and Gen AI could not carry out", and the study notes this holds "at various skill levels". It is not a statement about how skilled or valuable the work is. It is a statement about the medium the work happens in.

The same pattern appears at industry level. Industries "focused on knowledge work, information and services are more likely to have occupations with automating potential, as opposed to industries that involve working more directly with physical materials".

What tends to travel with physical work is variability. A task performed in a different place each time, on objects that are not identical, in conditions that change, with a person present who has their own preferences and needs, resists the kind of structured decomposition that current systems handle well. JSA's own note on agentic AI reinforces this: agents appear most likely to automate tasks that are "structured, predictable, and low risk", and where they are applied to more complex, less routine work, the study's consultations suggest "human involvement will likely continue to be needed for decision-making, judgment, accountability, or contextualisation".

The counterintuitive part: the pressure is on the middle, not the bottom

Two decades of automation commentary trained people to expect the least-credentialled work to go first. The generative AI evidence does not show that.

JSA found that the potential to automate tasks is highest for middle-skill occupations, particularly at ANZSCO skill level 4, which includes several forms of clerical work. Its observation is pointed: many clerical tasks "that were not affected by previous waves of automation could now be undertaken in large part by Gen AI". Meanwhile higher-skilled occupations, ANZSCO skill levels 1 and 2, show the most augmentation exposure, because knowledge work is where the technology slots in most readily as an assistant.

So the exposure profile is neither top-down nor bottom-up. It concentrates on structured symbolic work wherever it sits, and thins out wherever work is embodied.

What the long-range modelling suggests

JSA ran computable general equilibrium modelling to 2050. This is scenario modelling against a base case, not a forecast, and it should be read as a statement about relative composition rather than about absolute numbers. On that basis, the five three-digit occupation groups gaining the most employment relative to the base case are:

  • Cleaners and laundry workers
  • Midwifery and nursing professionals
  • Business administration managers
  • Construction and mining labourers
  • Hospitality workers

The five losing the most relative to the base case are general clerks, receptionists, accounting clerks and bookkeepers, sales, marketing and public relations professionals, and business and systems analysts and programmers. At industry level, the projected gains concentrate in construction, accommodation and food services, manufacturing, education and training, and agriculture, forestry and fishing.

Note what that list is not. It is not a ranking of which jobs are pleasant, well paid or secure. Care work and cleaning appear because their tasks are hard to automate, not because the transition rewards them. The modelling also suggests disruption effects peak in the middle of the transition before improving relative to baseline projections, which is a statement about timing, and about how much room there is to prepare.

Lower exposure is not job security

It would be a mistake to read any of this as reassurance. Several reasons to hold it loosely:

  • The scores are a snapshot of current capability. JSA expects them to shift "especially with more integration of robotics and complementary technologies". Physical work is protected by the state of robotics, which is a moving boundary rather than a permanent one.
  • Low exposure says nothing about demand. An occupation can be technically hard to automate and still shrink for reasons of funding, demography or economic cycle.
  • Low exposure says nothing about job quality. JSA is explicit that exposure "does not by itself predict job quality or complexity". Some of the least exposed work is among the lowest paid and most physically demanding.
  • Mobility matters as much as exposure. JSA warns that some roles "face repeated exposure to automation with limited mobility options", which is a harder problem than exposure alone, because it describes workers with fewer routes out.

What to do with this if it is your job

The practical implication is not to chase a low-exposure occupation. It is to notice which parts of your own role are structured and symbolic, because those are the parts most likely to change first, and to build depth in the parts that are not: judgment applied to unfamiliar situations, accountability for outcomes, and work that depends on being present with other people. JSA's skills analysis found demand rising for both digital literacy and higher-order human capabilities, specifically critical thinking, communication and adaptability, at the same time.

If you are concerned about how AI is being used in decisions about your role, it is worth understanding what your rights are and how to raise it with your employer.

The occupation groupings and projections in this article are drawn from Jobs and Skills Australia's modelling of the Australian labour market. Long-range economic modelling is scenario analysis against stated assumptions, not a forecast, and it does not describe the prospects of any individual role or person. This is general information, not career or financial advice.

Related reading

Sources: Jobs and Skills Australia, Our Gen AI Transition: Implications for Work and Skills (14 August 2025) · Jobs and Skills Australia, Analysis Paper A: Exposure (2 September 2025) · ILO Working Paper 140 (20 May 2025)