Anthropic Just Spent $100M Proving AI Skills Are the Bottleneck. Here's What That Means for You.

If your company is still debating whether to buy everyone a chatbot license or actually train people to use one well, three numbers from this week settle it: $100 million, 100% versus 37%, and 2.2%.

What happened this week

Anthropic committed $100 million to train 10,000 "Frontier Deployed Engineers" through a new Claude Frontier Academy, built on a medical-residency model — simulated enterprise deployments, graded assessments, first cohorts arriving from Accenture, Bain, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk. The same week, a Mercor study pitted 12 licensed CPAs against frontier models on scoped month-end-close tasks: Claude Opus 5 scored 100% across 20 attempts, each in under 10 minutes. Unassisted accountants averaged 37%. But on the full APEX-Accounting benchmark — the ambiguous, open-ended version of the same close work — the models' score dropped to 56.4%. a16z's State of Markets II report found only 2.2% of U.S. households pay for AI services, even though 49% of adults use chatbots and 41% of workers use AI on the job. And OpenAI's head of ChatGPT, Tibo Sottiaux, told Lenny's Newsletter this week that judgment, taste, and supervision are the professional skills on the rise — rote execution is the one in decline.

The skill implication

That gap between 100% and 56% is the whole story. It's not a model limitation that a future release fixes — it's the line between a task that's been scoped into something checkable and one that hasn't. Anthropic isn't spending $100 million teaching 10,000 people to phrase prompts; a curriculum graded on simulated deployments is testing whether someone can take an ambiguous piece of real work and break it into the parts a model can finish at something close to 100%, versus the parts that need a person's judgment call. That's also the reason 49% of people use a chatbot and only 2.2% pay for one: most usage never leaves the shallow end, so it never needs to clear that bar.

The habit worth building this week isn't a new prompt trick. It's pausing before you hand off an open-ended task to ask which pieces of it are actually scoped — and treating the unscoped remainder as yours to do, not the model's to guess at.

Two people, two close calls

A controller at a 35-person manufacturing company runs AI-assisted reconciliations every month-end. She hands the model the line-by-line matching — invoices against purchase orders, flagging mismatches — because that's the scoped part it gets right nearly every time. The accrual estimates, the judgment calls about which discrepancies are rounding versus real, she still does herself, because that's the ambiguous 56% territory where a confident-sounding answer and a correct one aren't the same thing.

A marketing operations manager at a 40-person SaaS company gave her whole team chatbot access eight months ago. Output quality still varies wildly between people, because access was never the problem — nobody agreed on what "good" looked like before generating. She now opens every AI-assisted deliverable with a three-line quality checklist the team writes together first: what it has to say, who it's for, and what would make it embarrassing to send.

The takeaway

The technology cleared 100% on the scoped half of the job months ago. The $100 million bet, the 56% ceiling on ambiguous work, and the 2.2% who've bothered to pay for any of it are all pointing at the same unfinished half: knowing which half you're handing over.