If you're worried AI could come for your job, you're not being paranoid, just early. This page sets out what the data actually says: which roles are most exposed, which are holding up, and what "exposed" really means.
Before anything else, exposed doesn't mean eliminated. Most jobs on the high-exposure lists will change shape rather than disappear. Knowing where you stand is still the first step to making a plan, and that's what the rest of this site is for.
What "AI exposure" actually measures
Exposure measures how much of a job's daily tasks an AI system could plausibly do. It says nothing about whether you personally get replaced next quarter. A role can be highly exposed and still exist for years, because organisations adopt technology far more slowly than the technology moves.
That gap between what AI can do and what employers actually roll out is your window. It's also why exposure figures describe the work rather than predict anything about the worker.
What the observed-usage data actually shows
The most useful evidence comes from Anthropic's Economic Index, because it's built on how people actually use Claude rather than on theory about what AI might be able to do.
Its headline finding is a share rather than a league table: the proportion of jobs where AI is used for at least a quarter of their tasks has risen from roughly 36% to around 49%. Usage clusters in software development and technical writing, and leans towards mid-to-high-wage jobs rather than the lowest paid.
The more interesting part is what happens when you stop counting raw task coverage. Anthropic's January 2026 report weights tasks by how often they're actually done, how long they take, and how often Claude succeeds at them. That shuffles the order quite a bit:
- Data entry keyers and radiologists come out more affected than raw task coverage suggests.
- Teachers, and notably software developers, come out less affected, even though developers are among the heaviest AI users.
That second point is worth a moment, because it cuts against the most common headline on this whole subject. Heavy AI use in a field isn't the same as high replacement risk. Developers use these tools constantly, and that looks at least as much like the tools helping them as replacing them.
Anthropic also reports that tasks rated fully feasible for a language model make up about 68% of observed usage, while tasks rated not feasible make up roughly 3%. People are, sensibly, using it for what it's good at.
A note on the numbers you'll see elsewhere. Precise exposure percentages for individual jobs get passed around a lot ("78% of this job", "94% of that one") and are often credited to studies that don't actually contain them. When you see one, check whether the cited source publishes that exact figure. Often it doesn't. We used to carry a few of these numbers on this page and have taken them out.
One pattern does hold up across sources: this wave lands on routine knowledge work, not just low-paid work. Clerical and admin tasks cluster at the top of most exposure measures, which is a different pattern from earlier waves of automation. What these jobs share is how routine and text-based the work is, whatever it pays.
The least exposed jobs
At the other end, the roles holding up best tend to involve being physically present somewhere unpredictable, making real judgement calls under pressure, or skilled hands-on work. Across exposure measures, construction, farming, food preparation, the skilled trades and direct caregiving keep turning up at the bottom, and unpredictable physical settings are exactly what current systems handle worst.
Nobody's saying everyone should retrain as an electrician. If your current role is highly exposed, though, the neighbouring roles that lean on judgement, physical skill or human contact are safer ground to move towards.
The nuance most headlines skip
AI is already affecting work. On Anthropic's observed-usage measure, roughly half of jobs now see AI used for at least a quarter of their tasks. You'll also see figures blaming a specific share of recent redundancies on AI. Those usually trace back to announcements from outplacement firms rather than official statistics, so take any precise percentage with a pinch of salt. The direction of travel isn't in doubt, though.
The net job forecasts aren't apocalyptic either. The World Economic Forum's Future of Jobs Report 2025 projects about 92 million roles displaced and 170 million created by 2030, a net increase of roughly 78 million, with structural churn touching some 22% of jobs. It's a net gain, but the new jobs won't necessarily go to the same people or use the same skills.
So, honestly, it looks more like a transition than an extinction. The risk is that work reshuffles faster than people can retrain, and that part is avoidable.
What to do with this information
- Find your real exposure. Look past your job title to the tasks you actually do each day.
- Spot your safer next move. Most exposed roles have a lower-exposure neighbour that reuses a lot of what you already know.
- Close your AI skills gap. Becoming the most AI-confident person on your current team is often quicker and safer than switching careers.
- Build an income backup plan before you need one. Emergency savings, less debt, and a second income you can scale up.
Your AI exposure action checklist
0 of 8 done
Frequently asked questions
Which jobs are most exposed to AI in 2026?
Be wary of any answer that gives precise per-occupation percentages, those figures circulate widely and are often attributed to studies that do not contain them. What the observed-usage data from Anthropic's Economic Index actually shows is a share: roughly 49% of jobs now see AI used for at least a quarter of their tasks, up from about 36% a year earlier, concentrated in software development and technical writing. When tasks are weighted by how often Claude actually succeeds at them, data entry keyers and radiologists rank as more affected than raw task coverage implies, while teachers and software developers rank as less affected.
Does "highly exposed" mean my job will be eliminated?
No. Exposure measures how much of a role's daily tasks an AI system could plausibly perform, not whether your position disappears. Most high-exposure jobs change shape rather than vanish, because organisations adopt technology far more slowly than the technology itself moves. The usual estimate is a two-to-five-year lag between "AI can do this task" and "companies have restructured around it", and that gap is the window you have to adapt.
Which jobs are least exposed to AI?
Roles that combine physical presence in unpredictable environments, human judgement under pressure and hands-on skill. Construction, farming, food preparation, the skilled trades and direct caregiving consistently show little to no high-exposure work. That doesn't mean everyone should retrain as an electrician, the useful reading is that if your role is highly exposed, the adjacent roles leaning on judgement, physical skill or human contact are the safer ground to move toward.
Is AI actually causing job losses yet, or is this speculation?
Partly, though the precision of the usual figures is misleading. Percentages attributing a specific share of recent redundancies directly to AI generally come from employer and outplacement-firm announcements rather than official labour statistics. What is better established is that roughly half of jobs now see AI used for at least a quarter of their tasks. Net forecasts are less alarming than the headlines: the World Economic Forum projects about 92 million roles displaced against 170 million created by 2030, a net gain of roughly 78 million. The honest summary is transition rather than extinction, the real risk is work reshuffling faster than people can retrain.
Is this wave hitting low-paid jobs hardest?
No, and that's the finding that surprises people most. Workers with a bachelor's degree are more likely to be in highly exposed roles (about 27%) than workers with only a high school education (about 12%). The common thread isn't income level, it's how routine the work is, which is why this wave lands squarely on white-collar knowledge work.
Should I change careers because of AI exposure?
Usually not as a first move. Becoming the most AI-fluent person on your current team is faster, cheaper and lower-risk than switching fields, and it builds on domain knowledge you already have. The sequence that makes sense is: work out your real exposure from your daily tasks rather than your job title, identify the adjacent lower-exposure role that reuses most of what you know, close your AI-skills gap where you are, and build the income backup plan before you need it.
Where to go next
- Build a 6-month income backup plan: the practical companion to this page.
- Start freelancing with skills you already have: the fastest route to a second income stream.
- Reskilling paths actually worth your time: which certifications convert to income and which don't.
- AI side hustles that actually work: turning a skill you already have into a second channel.
Sources & methodology
Figures on this page come from the following. Where a number is likely to move, the article says so rather than freezing it.
- Anthropic Economic Index: observed-usage share of tasks (36% rising to ~49%), concentration in software development and technical writing, and the success-rate-weighted reordering reported in the January 2026 report.
- WEF Future of Jobs Report 2025: 92 million roles displaced, 170 million created by 2030, net +78 million, 22% structural churn and ~39% of skills changing. Based on a survey of over 1,000 employers across 55 economies.
What was removed, and why. An earlier version of this article carried precise per-occupation exposure percentages attributed to Anthropic's research, together with figures from several secondary blogs and a 2020-era WEF projection presented as current. The per-occupation numbers are not published in the cited source, and the WEF figures had been superseded. Both have been removed rather than re-sourced, because a precise number with a citation that does not support it is worse than no number. Updated 20 September 2026.
Spotted something still wrong? Tell us: see our sources and corrections policy.
