Labour-market brief · 24 job profiles · 23 public sources
The work that stays.
AI is not wiping out jobs. It is splitting them: the same job can become more skilled and better paid, or simpler and worse paid. Below is what the evidence says about which is which — with the source printed next to every number.
workers worldwide are in an occupation with some generative-AI exposure — but only 3.3% sit in the highest exposure band.
ILO 2025how far 22–25-year-olds in the most AI-exposed occupations sit below the path they were on. It was 15% a year earlier; experienced workers show no such gap.
Stanford 2026average wage premium for workers whose jobs require AI skills, up from 57% a year earlier.
PwC 2026home health and personal care aide jobs projected by 2035 — the largest addition of any occupation in the US economy.
BLS 2026Some jobs are hard to automate
They need hands in a place that keeps changing, a person to answer for the result, or a relationship with someone. Care work, trades, teaching, medicine. see them
Some are already shrinking
Mostly routine office work — typing, filing, booking, first-draft writing. The loss lands hardest on people just starting out, because the entry-level version is the first to go. see the risk
New jobs are appearing
Building AI, installing it inside a company, labelling and checking its output, and watching it once it is live. They pay well and they are growing fast. see them
The short answer
Tasks die. Jobs change shape.
Employers surveyed by the World Economic Forum expect AI and information processing to create 11 million roles and displace 9 million by 2030, while robots displace 5 million more than they create.WEF 2025 That is churn, not collapse — and it is a forecast, not an observation.
The observed data is sharper and more uncomfortable. In payroll records, employment for workers aged 22–25 in the most AI-exposed occupations now sit 19% below where they would be had they kept pace — up from 15% a year earlier — while experienced workers in those same jobs show no comparable gap. The adjustment came through hiring rather than pay, and it concentrated where AI automates rather than augments.Stanford 2026
So the useful question is not "is this job safe". It is which half of it survives — and PwC's analysis of a billion job advertisements gives the clearest answer available: AI moves a job's required expertise up or down, and the two directions have opposite outcomes.PwC 2026
The two-track labour market
AI raises the expertise bar — or lowers it
AI does the boring parts, so the job gets harder — and pays more.
AI takes the routine tasks; the human does the harder, more expert ones
These roles are growing twice as fast as democratised ones, with 42% higher wage growth since 2021.
Jobs created and displaced, by technology driver
What to notice
-
The biggest driver employers name is not AI at all — it is people getting online.
19M jobs created, 9M displaced, for broadening digital access.
-
In employers' own projections, AI adds more than it takes away.
11M created against 9M displaced — a net +2M, and still a forecast.
-
Robots are the only driver where displacement wins.
A net −5M, with no creation figure published to balance it.
The four tests
Four things that keep a job out of a machine's reach.
They come up again and again in the research. A job with none of them is the one at risk. Each job page says how much each one applies, in plain words — our reading, not a measurement.
The work needs a body, in a place that keeps changing
Work that changes shape daily, where a mistake injures someone. Robots reach about 13% of US work hours against 44% for software agents.
A person is legally or morally answerable for the result
Nobody insures a model or puts one on the stand. Indivisible bundles survive: every school still needs a principal.
It depends on a relationship with a person
New tasks in AI-exposed roles lean 2.5× harder on empathy, presence, judgement and creativity.
The role is one job, not a pile of tasks
Most of today’s skills apply to work that could be automated and work that cannot. The seat stays; the mix changes.
Who is losing ground
Why the four tests are not a slogan
Unpredictability plus consequence
The most useful ranking in the literature scores occupations on six O*NET abilities: stamina, static strength, reaction time, spatial orientation, time sharing and problem sensitivity. Firefighting tops it — not because the work is manual, but because it needs real-time decisions in conditions that change, where the outcome hurts someone. Resume Now 2026
That combination, not physical labour by itself, is what the evidence keeps pointing at.
Profiles
Six roles that carry the argument
Home health and personal care aides
The largest source of new jobs in the US economy — and the one AI helps least.
Nurse practitioners
The fastest-growing occupation in the BLS table, at 41% — more than ten times the economy-wide rate.
Firefighters
Ranked the single most AI-proof occupation in a study of O*NET physical abilities.
Data entry keyers
Fully specified input, verifiable output, no judgement — the pure case of automatable work.
Customer service representatives
The clearest case of AI eating the entry rung rather than the ladder.
Software developers
Split in two: the ones who direct agents are thriving; the ones who wrote boilerplate are not.
Junior data analysts
The role that proves the entry level is being redesigned rather than deleted.
AI engineer (applied / LLM)
The single fastest-growing job title in the US, two years running.
Forward deployed engineer
Postings up 729% in a year: the job that exists because AI projects fail for organisational reasons.
AI governance, risk and audit
Regulation creates a function; the function creates the jobs.
Where to go next
Care, trades, accountability: the work with direct evidence behind it.
What is exposed, how exposure is measured, and why it is not displacement.
The roles being created, with the growth rates attached.
Jobs that don’t exist yet — the patterns the evidence is already pointing at.
Four tests, one instruction, and the three ways this brief could be wrong.
What would change this conclusion
Four assumptions this brief rests on.
Stated plainly, because none of the forecasts above are destiny. Each one names the assumption it attacks.