What jobs are safe from AI? Official projections

The US Bureau of Labor Statistics projects employment occupation by occupation, and its 2025 to 2035 projections give the clearest available answer to which jobs are safe from AI. The fastest-growing occupations are almost entirely hands-on healthcare and physical installation work. The steepest declines are clerical, transactional and desk-based. That split is the answer, and it is more specific than any list of predictions.

Key facts

The US Bureau of Labor Statistics projects occupational employment over 2025 to 2035, and publishes both the fastest-growing and the largest-declining tables.
Nurse practitioners lead growth at 41.0 per cent, followed by solar photovoltaic installers at 36.5 and data scientists at 34.6.
Eight of the ten fastest-growing occupations involve physically touching a person or installing equipment. The other two build the technology.
Cashiers face the largest numeric decline at 200,600 fewer posts, a fall of 6.5 per cent over the decade.
Data entry keyers face the steepest proportional decline in the top ten at 25.5 per cent, followed by tellers at 13.2 per cent.
Exposure and job loss diverge: computer and information research scientists are highly exposed to AI and are the eighth fastest-growing occupation.

What the official projections actually say

The BLS publishes a table of the fastest-growing occupations over a ten-year horizon, and the 2025 to 2035 edition reads as follows at the top.

Nurse practitioners, projected to grow 41.0 per cent. Solar photovoltaic installers, 36.5 per cent. Data scientists, 34.6 per cent. Wind turbine service technicians, 29.5 per cent. Medical and health services managers, 24.2 per cent. Physical therapist assistants, 23.0 per cent. Psychiatric technicians, 22.3 per cent. Computer and information research scientists, 21.8 per cent. Occupational therapy assistants, 21.5 per cent. Ophthalmic medical technicians, 21.4 per cent.

Eight of those ten involve either physically touching a person or physically installing equipment. The two that do not, data scientists and computer and information research scientists, are the jobs that build the technology in question.

And the jobs going the other way

The same projections include a table of the largest declines, and the pattern there is just as clear.

Cashiers, down 200,600 posts or 6.5 per cent. Office clerks, down 156,200 or 6.0 per cent. Customer service representatives, down 141,800 or 5.3 per cent. Secretaries and administrative assistants, down 114,100 or 6.0 per cent. Bookkeeping, accounting and auditing clerks, down 85,600 or 5.6 per cent.

Further down the list, two occupations stand out for the rate rather than the volume. Tellers are projected to fall 13.2 per cent, and data entry keyers 25.5 per cent, which is the steepest proportional decline in the top ten.

Read the two tables together and the dividing line is not intelligence or education. It is whether the work consists of handling information at a desk, or of being physically present and accountable in a room.

Spacious empty office featuring modern desks, chairs, and partitions — Cashiers, office clerks, customer service representatives and secretaries account for the four largest projected job declines to 2035. Photo: cottonbro studio / Pexels

Why physical presence is the strongest protection

Three reasons, and none of them is about how clever a model is.

The first is that the work requires a body in a specific place. A nurse practitioner examining a patient, a technician climbing a wind turbine or an installer on a roof is doing something no software can do, and the robotics needed to do it is a separate and much slower problem than language models.

The second is licensing and liability. Many of the fastest-growing roles are regulated: someone must hold a licence and carry responsibility for the outcome. Automating the cognitive part of such a job does not remove the requirement for a licensed human to sign for it.

The third is demography. Healthcare growth in these projections is substantially driven by an ageing population, which is a demand story independent of technology. That is why nursing and therapy assistance appear near the top in projections made across many different years.

Does education protect you?

Less than people assume, and the projections make that visible.

The fastest-growing list spans a very wide range of required preparation. Nurse practitioners need a graduate degree and a licence. Solar photovoltaic installers and wind turbine service technicians typically need neither. Psychiatric technicians, occupational therapy assistants and ophthalmic medical technicians sit between them. All five are in the top ten.

“Eight of the ten fastest-growing occupations involve either physically touching a person or physically installing equipment. The two that do not are the jobs that build the technology.”

Reading of the US Bureau of Labor Statistics 2025 to 2035 projections

The declining list is similarly mixed. Secretaries, bookkeeping clerks and first-line retail supervisors are not uneducated occupations, and several of them require specific training and certification. What they share is that the work product is information handled at a desk.

So the useful variable is not years of study. It is whether the output is a document or a person, and whether someone must be physically present and legally responsible for it. That reframing is uncomfortable because it cuts across the usual assumption that more education means more protection, and the data does not support that assumption in this particular case.

Where AI exposure and job loss are not the same thing

This is the distinction most coverage of this question misses.

An occupation can be highly exposed to AI without shedding jobs, if the technology changes what the job consists of rather than whether it exists. Software engineering is heavily exposed by any measure, and computer and information research scientists are the eighth fastest-growing occupation in these projections.

Conversely, the declining occupations have been declining for reasons that predate generative AI. Cashiers have been falling with self-checkout and online retail. Tellers have been falling with cards and apps. Data entry has been falling since optical character recognition. AI accelerates those trends; it did not start them.

That matters practically. A worker deciding what to retrain into should look at whether an occupation is growing, not at whether it sounds futuristic. Our explainer on what AI washing is covers the parallel problem of judging employers by how much AI they claim to use.

The jobs most likely to be reshaped rather than removed

A middle category is doing most of the work in this debate and rarely gets named.

Professional work with a verification step. A radiologist, a lawyer, an auditor and a translator all now have tools that produce a first pass. In each case somebody qualified still has to check it, and the liability sits with the checker. The volume of output per person rises, which affects how many people are needed, and the role does not disappear.

Skilled trades with a diagnostic element. An electrician or a car mechanic can use AI for diagnosis and still has to do the work. The tool makes a less experienced person more capable, which changes training more than it changes employment.

Creative work with a client. The first draft is cheap now. The judgement about which draft is right for a particular client, and the accountability for the choice, is the part being paid for.

What the projections cannot tell you

Three honest limits.

They are projections, built on models of demographic and economic trends, and they are revised. A ten-year projection made in 2025 about 2035 is a considered estimate rather than a forecast anybody should treat as settled.

They are occupation-level and your job is not. Two people with the same job title can spend their days doing completely different proportions of automatable work, and the projection says nothing about either of them specifically.

And they are United States only. The occupational mix, the licensing regimes and the demographic pressures differ enough between countries that the US table is a guide rather than a global answer.

What to actually do with this

Four things the data supports, none of which is dramatic.

Move toward work that requires presence, licensure or accountability, because all three resist automation for reasons unrelated to model capability. Within an exposed occupation, move toward the verification and judgement end of it rather than the production end, because that is where the liability and therefore the pay is going.

Learn the tools in your own field rather than in general. The occupations growing fastest that are not physical are the ones building the technology, and the ones being reshaped reward people who can operate the tool competently.

And treat any list of AI-proof jobs that does not cite employment data with suspicion. The BLS tables are free, published, and specific, and almost nothing written on this subject uses them.

The short answer

The jobs safest from AI, by the US government's own ten-year projections, are hands-on healthcare roles, physical installation and maintenance trades, and the technical roles that build AI itself. The jobs most at risk are clerical, transactional and desk-based information handling, where data entry keyers face a projected 25.5 per cent decline and tellers 13.2 per cent.

Presence, licensure and accountability are the three things that protect an occupation. None of them is about being smarter than a model.

This article reports published government projections and is not career advice. Figures were verified on 20 October 2026.

Frequently asked questions

What jobs are safe from AI?

By the US Bureau of Labor Statistics 2025 to 2035 projections, the fastest-growing occupations are nurse practitioners at 41.0 per cent, solar photovoltaic installers at 36.5, data scientists at 34.6, wind turbine service technicians at 29.5 and medical and health services managers at 24.2. Eight of the top ten involve physically touching a person or installing equipment, and the two that do not are the roles that build the technology itself.

Which jobs will AI replace first?

The BLS projections for the largest declines to 2035 point to clerical and transactional work. Cashiers lose the most posts at 200,600, down 6.5 per cent, followed by office clerks at 156,200, customer service representatives at 141,800, secretaries and administrative assistants at 114,100 and bookkeeping clerks at 85,600. Data entry keyers face the steepest proportional fall at 25.5 per cent, and tellers 13.2 per cent.

Will AI take my job?

The honest answer is that occupation-level projections cannot tell you about your job specifically. Two people with the same title can spend their days doing very different proportions of automatable work. What the data supports is a direction: work that requires physical presence, a professional licence or accountability for an outcome is growing, and work that consists of handling information at a desk is declining.

Why are healthcare jobs growing so fast?

Three reasons, only one of which is about technology. The work requires a body in a specific room, which is a robotics problem rather than a language-model one. Many of the roles are licensed, so a qualified human must sign for the outcome whatever software was involved. And an ageing population drives demand independently of technology, which is why nursing and therapy assistance appear near the top of projections made in many different years.

Is software engineering safe from AI?

It is heavily exposed and it is also growing. Data scientists are the third fastest-growing occupation in the 2025 to 2035 projections at 34.6 per cent, and computer and information research scientists are eighth at 21.8 per cent. That is the clearest illustration that exposure and job loss are different things: AI can change what a job consists of without changing whether it exists.

Were the declining jobs already shrinking before AI?

Most of them, yes, and that is important context. Cashiers have been falling with self-checkout and online retail. Tellers have been falling with cards and banking apps. Data entry has been declining since optical character recognition became reliable. Generative AI accelerates those trends rather than starting them, which means attributing the whole decline to AI overstates how fast the change arrived.

What kinds of jobs get reshaped rather than removed?

Professional work with a verification step, where a tool produces a first pass and somebody qualified must check it and carries the liability. Radiology, law, audit and translation all fit. Skilled trades with a diagnostic element, where AI helps identify the problem and a person still does the work. And creative work with a client, where the first draft is cheap and the judgement about which draft is right is the paid part.

Can I trust lists of AI-proof jobs online?

Only if they cite employment data, and most do not. The BLS tables are free, published, specific and revised on a regular cycle, and almost nothing written on this subject uses them. A list built from guesses about which jobs sound automatable will systematically misclassify occupations that are highly exposed to AI and growing anyway, which is a large and important category.

What are the limits of these projections?

Three. They are projections built on demographic and economic models and they get revised, so a ten-year estimate is considered rather than settled. They are occupation-level, so they say nothing about any individual job. And they cover the United States only, where the occupational mix, licensing regimes and demographic pressures differ enough from other countries that the table is a guide rather than a global answer.

What should I retrain into if I am worried about AI?

The data supports moving toward work requiring presence, licensure or accountability, because all three resist automation for reasons unrelated to how capable models become. Within an exposed occupation, move toward the verification and judgement end rather than the production end, since that is where liability and pay are heading. And learn the tools in your own field specifically rather than in the abstract.

What jobs are safe from AI? The official projections, not the predictions