Over 80% of U.S. occupations now have at least 1 in 5 workers using AI — But Only Coders Are Visibly Losing Work
Sixty-one percent of service firms in the New York Fed's region now use AI. Four percent have laid anyone off because of it.
That single ratio — near-universal adoption, near-zero displacement — is the whole story, and it's the opposite of the story everyone thinks they're reading.
Near-universal adoption, near-zero displacement — that single ratio is the whole story, and it's the opposite of the story everyone thinks they're reading.
Because, per Fed Board economists Leland Crane and Paul Soto, the occupation visibly losing work to AI in the employment data is software coders. Employment of software coders has decelerated sharply since the launch of ChatGPT in November 2022, and Fed Board economists Leland Crane and Paul Soto modeled a counterfactual to confirm it's an occupational shock, not an industry one. Coders, specifically, not the industry around them.
So here's the split. Over 80% of U.S. occupations now have at least one in five workers using AI, reportedly. Service-firm adoption leapt to 61% this year, from 40% last year and 25% in 2024, per the New York Fed. And after all that, the only legible casualty in the employment data is the one job that writes the tools.
Last week in this space we argued the Fed can't yet read AI's productivity gains. This week the data got sharper, and the picture inverted: the gains are broad and illegible, the displacement narrow and legible. Coders aren't the leading edge of a wave. They're a special case that's misleading everyone.
Start with why coding is different, because the difference is structural, not temporal.
Two things have to line up for AI job loss to show up in the statistics. The tool has to eat the exact task a worker is paid to do — same-task substitution, not adjacency. And the industry has to already meter its own output, so the substitution shows the moment it happens.
Coding is the one occupation where both are true at once. A model that writes working code substitutes directly for the task, and software already counts commits, tickets, features, headcount per unit of output. The capability curve and the measurement layer are bolted to the same task.
The IFP researchers writing at Noahpinion reportedly note that AI software-engineering capability is doubling roughly every seven months on time-horizon benchmarks. That curve lands hardest exactly where output is already counted. Of course coding shows up first. It's the cleanest possible case.
Now look at the rest of the 61%-adoption economy, where neither condition holds — and watch the pattern reverse.
Among service firms using AI, the median share of workers actually using it is 17%. For manufacturers, 7%. Adoption is a mile wide and an inch deep: the firm has AI, a sixth of the staff touches it.
The workforce response isn't layoffs. Just 4% of service firms cut anyone due to AI — up from 1% a year ago, but still a rounding error. No manufacturers reported layoffs at all, this year or last, per the New York Fed.
And 13% of service firms did the opposite — they hired more workers specifically to use AI. The single most common adjustment isn't firing or hiring. It's retraining: over a third of service firms retrained existing staff to work the tools rather than replace them.
This is not an economy shedding labor. It's an economy reorganizing around a new input, mostly by teaching the people already there.
So why does none of that show up the way coding does? Look at the time data, because it contains the whole measurement problem in two numbers.
The share of work hours assisted by generative AI hit 6.3%, per FRED. The share of work hours actually saved: 2.2%.
That gap is everything. AI now touches 6.3% of hours but frees only 2.2% of them — and even the 2.2% is self-reported, workers eyeballing their own time. "AI touches the work" and "AI frees the hour" are different measurements, and only the first is easy to see.
In coding, the freed hour converts straight into a metered output change, so it registers as employment. Everywhere else, the freed hour dissolves into a slightly faster analyst, a marketing team that ships more, an accounts-payable process that clears quicker — real productivity that no current statistic can attribute to AI.
That's the trap. The displacement signal is legible precisely because it's narrow. The productivity signal is illegible precisely because it's broad.
And this is where it stops being a labor-market curiosity and becomes a rates problem.
The most important number in monetary policy isn't the funds rate. It's potential GDP — how fast the economy can grow without overheating. The Fed sets policy against that ceiling.
The CBO now projects real potential GDP growth of just 2.1% a year from 2026 through 2036. The reason is demographic: aging, falling participation, collapsed immigration mean the economy adds fewer than 10,000 potential workers a month, reportedly.
With labor-force growth near zero, the arithmetic is brutal. Every dollar of future growth above that floor has to come from productivity — from each existing worker producing more. There's no other lever left.
On his 100th day as Fed Chair, Kevin Warsh proposed treating AI as a fourth factor of production, alongside labor, capital, and land. If AI generates output that doesn't scale with more workers or traditional capital, the non-inflationary growth ceiling rises — and a Fed that can tolerate faster growth can eventually hold rates lower for longer.
I'm not adjudicating that call here; the rate-path fight belongs to another piece. What matters is what Warsh's framing actually is: a measurement proposal. He's asking the Fed to raise the growth ceiling on the strength of a productivity signal.
And that's exactly the signal the data says the Fed cannot yet read.
Sit with the asymmetry. To lift potential GDP off the 2.1% floor, the Fed has to believe in the broad, illegible half of the ledger — the 2.2% of hours quietly freed across 80% of occupations. The half it can measure cleanly, coder displacement, is small, narrow, and tells it nothing about the ceiling.
So the Fed is being asked to bet the growth ceiling on the half of the AI story its instruments can't currently see, while the half they can see clearly is a single decelerating occupation.
The optimistic read is that the illegible gains are real and simply early — the FRED Blog notes that if the trend holds, AI could show up in aggregate productivity in the years ahead. The 6.3% of hours assisted is the leading indicator; the 2.2% saved is what's converted so far.
The pessimistic read is that some of that assisted time is novelty, and the true conversion rate stays low — in which case raising potential GDP now imports an inflation error later, because the Fed tolerated growth the economy couldn't actually sustain.
Here's the part that should unsettle the potential-GDP debate. The Fed will get the displacement signal it can read long before the productivity signal it can't.
Occupation-level job losses arrive job by job, cleanly, in the BLS data — coders today, entry-level roles perhaps next, since one study reportedly flags AI substituting for the routine tasks new hires cut their teeth on. That's a legible drip of bad news.
The productivity payoff, if it comes, arrives late, diffuse, and hard to attribute — smeared across a workforce that got retrained rather than replaced.
Which means the sequencing is backwards for anyone hoping AI lifts the ceiling. The evidence of cost lands first and reads clearly. The evidence of benefit lands last and reads poorly, if at all.
Coders aren't the canary. They're the one case where the instruments happen to work. The actual gamble isn't in the layoff count — it's in trusting the half of the story the instruments can't yet see.