Labor & Automation
The four moats: where work survives the machines
For fifty years, technology has erased routine jobs. This is a data-backed look at the work it can’t reach — and why.
The routine jobs left years ago. Here’s what stayed.
Since 1980, the share of American jobs made up of routine office and factory work has fallen hard. But it didn’t fall into nothing. It flowed into four kinds of work that machines and overseas labor still can’t take. Two of them grew a lot. One is huge and holding steady. And the one we love to romanticize turns out to be a rounding error.
Routine office and factory jobs shrank from nearly half of all work to about a third. Most of the drop happened between 1990 and 2010 — the robots mostly came, and left, a while ago.
The work that expanded was being responsible for a risky decision (up 15%→19%) and being a person people trust (up 7%→11%). Not genius. Not creativity.
Hands-on physical work — aides, trades, drivers — is the largest shielded category and has barely moved in 45 years. A stable floor, not a rocket.
Athletes, musicians, performers — the “do what you love” dream — is less than 1% of all jobs, and it’s the one category machines are now actively copying.
Share of U.S. jobs by type, 1980–2025
If machines keep getting better, what work is actually safe?
“A robot will take your job” is a headline, not a plan. It’s also too simple. Machines and overseas workers have been replacing certain jobs for decades — but not evenly, and not at random. Some work gets automated almost immediately. Other work barely budges, year after year, no matter how good the technology gets.
So the useful question isn’t “will automation happen?” It already has. The useful question is: which jobs keep resisting it, and what do those jobs have in common? If we can name the reasons some work survives, we can make better bets — about careers, about hiring, and about what to teach the next generation.
Not “will the machines win?” — but “where can’t they play?”
Four “moats” protect work from the machines
A moat is the water around a castle — a barrier that keeps attackers out. A job has a moat when there’s a specific reason a machine or a cheaper worker somewhere else can’t do it. We think there are only four such reasons. Every surviving job leans on at least one of them. Everything without a moat is on the automation menu.
Embodiment
Work where a person must physically be in the room, using their hands and split-second judgment in messy, unpredictable places. You can’t email it overseas, and robots are still clumsy at it.
Accountability
Work where a human must own a big, risky decision — and take the blame if it goes wrong. Often the law or a license requires a real person’s name on the line.
Authentic human origin
Work where the entire point is that a person did it. A machine could technically do it “better” and nobody would care. Nobody buys a ticket to watch a robot play the World Cup.
Durable relationship
Work built on a trusted human bond — someone you confide in, rely on, and keep coming back to. The relationship itself is the product, and it can’t be copy-pasted.
The routine stuff — no moat
The routine office and factory work that software and overseas workers can handle: filing, data entry, checkout scanning, assembly lines, and most first-draft, rule-following tasks. It has no moat — and it’s the bucket that keeps shrinking.
Testing the idea against 45 years of jobs
We tracked all five buckets from 1980 to 2025 — the hero chart at the top. The pattern is clear, with one honest exception. Here’s what the lines actually say.
The routine work already left −8 points
Routine jobs fell from 46% of all work in 1980 to 38% today. But most of that drop happened between 1990 and 2010 — and it has nearly flattened since. The automation wave people are bracing for mostly already broke, quietly, over the back-office desk.
Blame and trust soaked it up the winners
Accountability rose from 15% to 19%, and durable relationships from 7% to 11%. Together those two moats absorbed almost all the work the routine bucket lost. The jobs that grew aren’t the flashy ones — they’re “own the decision” and “be the person people trust.”
Physical work is a floor, not a rocket ~31%, flat
Embodiment is by far the biggest protected bucket — about a third of all jobs — but its share has been essentially flat for 45 years. It’s enormous and durable, but it isn’t growing its slice. The robots never actually came for the plumber.
The dream job is a rounding error under 1%
Authentic-origin work sat well below 1% the entire time and never grew. The work we most romanticize is, by the numbers, almost invisible. It’s also the one moat machines are now climbing into — which is the honest caveat, not the headline.
The forecast points the same way to 2034
The U.S. government’s official job forecast to 2034 keeps the trend going: routine work keeps slipping, while accountability and care-and-relationship work keep rising. AI will likely speed this up — but as long as humans have problems, someone has to be paid to own solving them.
Recent years, plus the 2034 forecast
The method, in plain language
A chart like this is only as trustworthy as the sorting behind it. So here’s exactly what we did — and, just as important, where a reasonable person could disagree.
We used real government job counts
Two sources: a long-running national survey of workers going back to 1980 (for the 45-year picture — it counts the self-employed), and a detailed employer survey plus the government’s official job forecast (for recent years and the 2034 outlook). Every yearly total lines up with known U.S. employment.
We sorted every job into one bucket
We took about 1,000 detailed job titles and placed each into exactly one moat — or into “everything else” — based on the main reason it resists machines and offshoring. One job, one bucket, so the pieces always add up to the whole. An automatic check confirms this for every year.
We added up the buckets by year
Then we summed the jobs in each bucket, every year, and drew the lines — both as raw numbers and as a share of all jobs. The share view is where the story lives, because it shows how the mix of work is changing, not just that the country got bigger.
We published everything so you can check it
The sorting file, the raw data, and the code are all downloadable below. Disagree with a call? Change it in the file and the charts redraw. Nothing here is a black box.
Where a reasonable person could disagree
Deciding the “main reason” a job survives isn’t a hard science. We made explicit, consistent rules, but you could move some jobs and shift the lines. That’s why the full sorting is published.
Routine coding is increasingly automatable, so most software work went into “everything else.” But senior security and system-architecture roles — where a human owns high-stakes design decisions — went into accountability.
We counted most managers as “accountability” (they own decisions) and most in-person service jobs as “embodiment” (a body has to be there). These two choices do the most to set how big those buckets look. Both are defensible; both are documented.
Older surveys undercount self-employed artists, musicians, and athletes, so this bucket’s exact height is soft. But even when you count them fully, it stays under 1% of all jobs — the conclusion doesn’t change.
They’re the government’s best projection, not a measured fact. Technology forecasts are often wrong about timing. Treat the direction as the signal, not the decimal.
As long as humans have problems, someone gets paid to own solving them.
That’s the durable point under all the data. Automation doesn’t end work; it keeps narrowing it toward the four things a machine can’t do — show up in a body, be answerable, be trusted, and be authentically human. When we think about the careers ahead for the next generation, the skill that keeps rising to the top isn’t a specific credential. It’s emotional intelligence: the ability to handle other people well, and to be the one others trust with the decision. Two of the four surviving moats are built on exactly that.