Anthropic's Own Data Says AI Removes the Skilled Part of Your Job First. That's the Half of the Market You Should Be Building For.
Anthropic's Own Data Says AI Removes the Skilled Part of Your Job First. That's the Half of the Market You Should Be Building For.
Anthropic's Economic Index (the running research where they measure how Claude is actually used across the economy) landed on a finding that runs against the standard story. The tasks Claude handles skew toward requiring around 14.4 years of education, versus roughly 13.2 years for the average task in the economy. In plain terms: AI is eating the higher-skill slices of a job first. Not the grunt work. The analysis, the drafting, the planning: the parts you needed a degree for.
And there's a second number sitting next to the first that matters just as much. When Anthropic re-ran their productivity estimate accounting for reliability (whether the output is actually good enough to ship without a human checking it), the projected boost to US labor productivity dropped from around 1.8 percentage points a year to closer to 1.2. Both numbers, read together, draw a map. And the map points straight at where a solo operator still has pricing power.
The story everyone told was backwards
The comfortable narrative for years was: AI takes the boring, repetitive tasks and frees humans up for the creative, high-value work. Relax, it only comes for the drudgery.
The data says the opposite is happening. AI is disproportionately good at exactly the cognitively demanding parts: synthesizing a research summary, drafting the analysis, structuring the plan, writing the first version of the thing that used to require expertise. What it leaves behind is a lot of the routine glue: the coordination, the follow-up, the judgment calls, the "is this actually right and can I stake my name on it" part.
That's uncomfortable if your professional identity is built on being the person who does the skilled drafting. It's also, if you're building a business rather than defending a job title, the single most useful piece of market information you'll get this month. Because it tells you precisely which part of the value chain is getting commoditized and which part isn't.
The reliability gap is the whole opportunity
Look again at that productivity revision. The estimate fell when they factored in reliability, and that word is doing enormous work.
Here's what "1.8 down to 1.2" actually describes: AI can do the skilled task fast, but it can't do it reliably enough to ship unattended. There's a gap between "the model produced something that looks right" and "this is correct and I'll put it in front of a customer, a regulator, or a paying client." Somebody has to stand in that gap. Somebody has to own the last mile: verify it, catch the plausible-but-wrong output, take responsibility when it ships.
That gap is not a temporary bug that the next model release closes. It's structural for anything where being wrong has a cost. A model that's right 95% of the time is genuinely miraculous and also completely unshippable for work where the 5% includes a wrong contract clause, a broken migration, or a number that's off by a decimal. The value didn't disappear when the model got good at the task. It relocated: from doing the task to guaranteeing the task.
The solo-operator read
Put the two findings together and the business logic falls out.
The thing that's getting commoditized is access to the skilled capability. Everyone has it now. "I can draft the analysis" or "I can write the code" or "I can produce the design" is no longer scarce, because the model does the first 90% of it for anyone with an API key. If that's what your business sells, the floor is dropping out, and no amount of being good at the drafting saves you, because the model is now also good at the drafting and it works for pennies.
The thing that stays scarce is everything the model can't guarantee: reliability, accountability, and the domain judgment to know when the confident output is quietly wrong. That's the durable business. Not "I have access to the model" (everyone does). But "I own the reliability layer for this specific, consequential problem, and I'll stake my name on the result." You're not selling the capability. You're selling the guarantee wrapped around it.
Concretely, that means selling verified outcomes in a narrow domain where being wrong is expensive enough that customers will pay someone to own the correctness. It means building the eval harness and the domain corrections that catch the model's failures before the customer does, and treating that harness as your actual product, because it is. It means positioning as the accountable human in a loop that's otherwise automated, in exactly the places where "the AI said so" isn't a good enough answer.
What I'd actually do
If you're starting or repositioning a one-person business right now, pick a domain where wrong answers cost real money, and build your whole offer around being the reliability layer on top of cheap, capable models. Let the model do the skilled 90%. Sell the 10% it can't guarantee, plus the accountability for the whole thing. Price on the verified outcome, not on your time and not on the model's output, because time and output are both racing toward commodity and the guarantee isn't.
Own your evals like they're the crown jewels, because in this business they are. The set of tests and domain corrections that prove your output is correct for this specific job is the thing a competitor with the same model access can't copy, and it compounds every time you catch a new failure mode.
Here's the honest counter-take, and it's not a soft one. This cuts against a lot of people, and it might cut against you. If your service was the skilled drafting layer (you were the one who wrote the analysis, produced the design, drafted the code, and charged for that skill), then AI is coming for the exact thing you sold, and "reposition to the accountability layer" is a real, hard move, not a reassuring platitude. It means changing what you're expert at, from producing the work to guaranteeing it, and those are different muscles. Some people won't want to make that move, and that's a legitimate choice. But the data is pretty clear about which way the ground is tilting, and it's better to read the map now than to find out at the bottom of the hill.
Author
Lukas
@lukcombinator