75% of a programmer's daily tasks. Already covered by AI — not theoretically, not eventually. Right now.
Anthropic's March 2026 labor report, authored by economists Maxim Massenkoff and Peter McCrory, is the first major study using observed Claude usage data across 800 US occupations. That distinction matters more than any headline stat it generated. It's the difference between what AI can do in a lab and what it's already doing inside your company.
The Exposure Split Every Worker Should Know
Occupations with highest AI task coverage:
- Computer programmers: 75% of tasks covered by observed AI usage
- Customer service representatives: high exposure, measurable displacement pressure
- Data entry keyers: near-full task coverage
- Financial analysts: significant usage across core workflows
- Medical records specialists: substantial and growing exposure
Occupations with near-zero AI exposure (~30% of all US jobs):
- Construction trades
- Cooks and kitchen workers
- Mechanics and skilled trades
- Physical safety roles (lifeguards, etc.)
The remaining 70% of occupations register some measurable AI exposure. "Some" covers a wide range — from "AI assists occasionally" to "AI handles most of this." Most knowledge workers sit somewhere on that spectrum whether they've noticed yet or not.
By the Numbers
- 800 US occupations analyzed using real Claude usage data
- 75% task coverage for computer programmers — highest of any occupation tracked
- 14% drop in monthly job-finding rate for workers aged 22–25 in high-exposure roles since ChatGPT's emergence
- 6–16% fall in employment among the same cohort, per separate Brynjolfsson et al. research
- 3% current unemployment among highly exposed workers — vs. 4.1% for the broader US labor market
- 80% reduction in task-completion time Claude delivers in certain workflows
- 10.1% projected BLS growth for computer and mathematical occupations from 2024 to 2034
Why the Entry-Level Hiring Data Scares Me More Than the Unemployment Rate
Don't wait for unemployment to move. It won't — not yet.
A 14% drop in the job-finding rate for 22–25-year-olds in high-exposure occupations isn't dramatic in isolation. But pair it with Brynjolfsson's 6–16% employment decline in the same cohort, and a pattern forms. Entry-level hiring is where AI disruption registers first — not in layoffs, but in fewer open doors.
Here's the mechanic: when Claude reduces task-completion times by up to 80% in certain workflows, firms don't immediately fire staff. They stop backfilling vacancies. Four hours of analyst work compressed into under 50 minutes means the junior hire that would've done that work simply doesn't get recruited. That's what a 14% slowdown looks like from the inside of an HR team.
The overall unemployment rate for highly exposed workers sits at 3% — barely different from the broader 4.1% market. That gap doesn't tell you the story. The front door is closing. The unemployment number won't reflect it until the cohort that couldn't get entry-level jobs starts aging into unemployment statistics two to three years from now.
| Cohort | Job-finding rate trend | Change vs. 2022 |
|---|---|---|
| Low-exposure occupations | ~2% per month | Stable |
| High-exposure occupations (ages 22–25) | Declining | Down ~14% |
| Broad US labor market unemployment | ~4.1% | Minimal change |
What the 2001 China Trade Shock Teaches Us
Between 2001 and 2007, roughly 3.5 million US manufacturing jobs disappeared after China joined the WTO. Mainstream forecasters consistently projected those workers would absorb into service roles. That reabsorption mostly didn't happen.
What's striking isn't the parallel in job losses — it's the timing problem both shocks share. Economists spent over a decade underestimating the China shock's damage because official unemployment statistics didn't cleanly signal it. Workers left the labor force, accepted lower wages, or relocated rather than appearing as unemployment data points. Anthropic's researchers explicitly reference this dynamic: their early-warning index is designed to catch what aggregate statistics will miss until it's too late to act.
One critical difference: the China shock concentrated in specific geographies and hit manufacturing workers. AI exposure is landing on educated, higher-paid, geographically dispersed knowledge workers — a demographic with more exposure to financial markets, mortgages, and consumer spending. The transmission to the broader economy is likely faster.
The Bull Case — and Why I Think It's Being Overweighted
BLS projects computer and mathematical occupations to grow 10.1% between 2024 and 2034 — more than three times the all-occupation average. Healthcare and social assistance should account for roughly one-third of all new US jobs through 2034, per BLS — physical, human-intensive work that isn't automatable at scale. Citadel Securities reportedly saw software engineer hiring tick upward recently, suggesting demand for AI oversight roles is growing even as some tasks automate.
All true. But the mix of growth matters more than the total.
A labor market that adds 5 million healthcare and trades jobs while high-exposure knowledge-work roles stagnate is structurally different from uniform growth — even if headline unemployment rates look identical. That structural shift shows up first in wage growth deceleration among exposed occupations, then in consumer spending patterns among high-income earners, then eventually in housing markets in high-cost metros where knowledge workers are concentrated.
During the 2007–2009 financial crisis, US unemployment doubled from 5% to 10%. A comparable doubling in the top quartile of AI-exposed occupations — from roughly 3% to 6% — wouldn't register as a national emergency in aggregate data. But that cohort carries disproportionate weight in mortgage markets, consumer spending, and equity market participation. The transmission speed into financial markets would be faster than a comparable shock among lower-wage workers historically has been.
The One Number to Watch
Here's what actually tells you when this story is accelerating: the monthly job-finding rate for workers aged 22–25 in high-AI-exposure occupations. It's currently down 14% from 2022 baseline. If that number moves to 20% or beyond — confirmed by two consecutive BLS occupation-level reports — the repricing of human capital in knowledge-work roles has left the early-warning stage and entered the measurable economic damage stage. That's the canary. The unemployment rate won't tell you in time. This one might.
This analysis is for informational purposes only and does not constitute financial advice. Always consult a qualified financial advisor before making investment decisions.
