Quick answer: The exposure data leans one way. Analysis of how AI is actually used across occupations puts hands-on work at the bottom, with construction around 16.9% theoretical coverage and installation and repair around 18.4%, far below knowledge work. Meanwhile recent graduate unemployment has hit roughly 5.6% to 5.7%. But AI-proof is doing a lot of work in that sentence, and the physical job is only half of what a tradesperson actually does.
There's now real measurement behind the hunch that physical work is harder to automate. It's worth looking at properly, including the part that complicates the story.
What the exposure data shows
The Anthropic Economic Index, which combines what AI models can theoretically do with how people actually use them at work, finds the occupations least represented in AI usage are the ones involving serious physical labor.
Ground maintenance sits lowest at roughly 3.9% theoretical coverage. Construction and food service around 16.9%. Installation and repair around 18.4%. Production around 19%. All of it well under knowledge work, and the distance between computer and mathematical occupations at the top and farming, fishing and forestry at the bottom runs to more than three orders of magnitude.
The US Bureau of Labor Statistics now publishes AI exposure categories of its own, which tells you how seriously this is being tracked.
Meanwhile, the safe path got less safe
The other half of the story is what happened to the track that was supposed to be reliable.
Entry-level white collar work took the first real hit, because the tasks juniors traditionally cut their teeth on, drafting documents, reviewing files, cleaning data, writing basic code, are precisely what current models handle well. Recent graduate unemployment climbed to roughly 5.6% to 5.7%, at points running above the rate for the workforce overall. That's close to unprecedented. Computer science graduates specifically sat near 6.1% in 2025.
Employers have also been quietly reshaping junior roles rather than cutting them outright, loading entry-level postings with expectations that used to sit two rungs up. Which makes the first rung harder to reach even where the job still exists.
Why the physical work genuinely holds up
The protection isn't magic. It comes down to something quite specific: unstructured physical environments.
A model can draft a report from anywhere. It cannot get under a house, work out why a drain is backing up in a crawlspace nobody has ever mapped, and fix it in conditions that are different on every job. Diagnosing a fault in a building that three different people modified across forty years is a fundamentally different class of problem from summarising a document.
Robotics keeps improving, and I'd be careful about promising anyone this lasts forever. But the bottleneck for trade work has never been intelligence. It's dexterity in unpredictable spaces, and that has moved far more slowly than language ability has.
The part the AI-proof pitch leaves out
Here's where I'd push back on my own headline.
Notice who tends to be loudest about trades being AI-proof. Trade schools with seats to fill. Recruiters. Workforce programs with funding targets. That doesn't make them wrong, but it does mean the claim arrives pre-polished, and it's usually stated as though a trade is one single thing.
It isn't. Running a trade business is roughly half physical work and half estimating, scheduling, quoting, chasing invoices, answering customers and handling dispatch. The physical half looks genuinely durable. The other half is exactly the sort of routine text, numbers and coordination work that AI is already good at, and there are products selling into it right now.
That's not a reason to avoid the trades. It's a reason to be precise about the claim. The work is safe. The margin sitting on top of the admin around it is not automatically safe, and the tradespeople who do best over the next decade will be the ones using those tools rather than being priced out by competitors who do.
Also worth saying plainly: safe from automation and well paid aren't the same thing. Ground maintenance has the lowest AI exposure in the entire dataset. Nobody's arguing it's the best career available.
So what would I actually tell a teenager
That the risk calculation which made college automatically safer deserves rechecking, which is a smaller and more defensible claim than trades are AI-proof.
We covered the education side in college or the trades, and where the money actually sits in the highest paying trade jobs in 2026.
If the trades are where the work is, get found for free
Qiggz is a free, US-based marketplace connecting tradespeople, apprentices included, with local homeowners and employers. No lead fees, no commissions, one profile for gig work and full-time roles.
Create your free Qiggz profile or browse open trade jobs. New to this? Start with apprenticeship versus trade school.
Frequently asked questions
Are skilled trades really safer from AI than office work?
The exposure data supports it for the physical work. Analysis of AI usage across occupations places hands-on roles lowest, with construction around 16.9% theoretical coverage and installation and repair around 18.4%, far below knowledge work. The reason is that unstructured physical environments remain hard to automate.
Which jobs are most exposed?
Knowledge work built on routine text, code and data handling. The impact has landed hardest on entry-level roles, because the tasks juniors traditionally start with are the ones current models handle best.
Is graduate unemployment rising because of AI?
Recent graduate unemployment reached roughly 5.6% to 5.7% and at points exceeded the overall workforce rate, which is highly unusual. AI adoption is widely cited as a driver, particularly through the contraction of entry-level roles, though it isn't the only factor in a complex labor market.
Is any part of a trade business exposed to AI?
Yes, and this gets skipped a lot. The physical work looks durable, but estimating, scheduling, quoting, dispatch and customer communication are routine coordination tasks that AI already handles, and products are selling into exactly that. Expect tradespeople who adopt those tools to out-compete those who don't.
Will robots eventually replace tradespeople?
The constraint is dexterity in unpredictable physical spaces rather than intelligence, and that has advanced far more slowly than language ability. It's a real long-term question rather than a near-term one, and anyone promising you a lifetime guarantee is overreaching.
Does low AI exposure mean a job is a good career?
No, and it's worth separating the two. Ground maintenance has the lowest AI exposure in the dataset and isn't the best-paid work available. Low automation risk is one input, not the whole decision.
Sources
- AI usage and exposure by occupation, including physical labor occupations being least represented and the coverage figures cited: Anthropic Economic Index, 2026.
- US government AI exposure categories: Bureau of Labor Statistics, 2026.
- Recent graduate unemployment of roughly 5.6% to 5.7%, at points above the overall workforce rate, and computer science graduate unemployment near 6.1% in 2025: Federal Reserve Bank of New York data as reported by Washington Monthly and related 2026 coverage.
- Restructuring of entry-level roles toward senior-level expectations: 2026 labor market reporting.
- Qiggz no lead fees and no commissions: first-party (Qiggz product); see /for-contractors.
