Claude Code Writes 90% of Anthropic’s Code. The CFO Says That Means Hire More People.

Anthropic's Claude writes 90% of its code. The CFO says that means it needs more humans, not fewer.

When Krishna Rao joined Anthropic as CFO in May 2024, the company had an annualised revenue of roughly $250 million and a product that had not yet launched. Two years later, Anthropic is reportedly running at $30 billion in revenue, its Claude models are embedded in millions of workflows, and Rao's finance department is reviewing AI-generated financial statements that arrive 90 to 95 percent ready for human sign-off. Rao's description of what that shift actually feels like from inside: "It's accentuating and accelerating the talent we already have."

The quote comes from an episode of Patrick O'Shaughnessy's "Invest Like the Best" podcast published this week. Rao laid out a picture of a workplace where the execution layer of knowledge work, from software engineering to financial reporting, has been substantially delegated to AI systems, while humans concentrate on judgment, strategy, and what he calls "talent density" — smaller groups of skilled people directing more work than the same headcount could previously manage.

"We've hired a lot more people because of that," Rao said. "And I think that's starting to be true across many companies as well."

Claude Code now generates more than 90 percent of Anthropic's codebase, according to Rao. The finance function, which once required days of preparation for quarterly reporting cycles, now produces drafts that are close to final in a matter of minutes. Rao cited an internal review that consumed hours in prior quarters, which now runs in approximately 30 minutes with AI handling the construction and human reviewers handling the certification.

The framing matters. Rao did not frame this as a story about automation replacing workers. He framed it as a story about the division of labour inside a knowledge company. "Everyone kind of becomes a manager," he said, describing teams that oversee multiple AI agents operating across simultaneous workstreams. Execution has been offloaded. Oversight has not.

This is a specific and verifiable claim from the CFO of one of the most-watched companies in AI. It is also a claim with obvious limits. Anthropic is not a representative firm. It builds the technology it is deploying internally, it has access to capabilities before they reach customers, and it is in a hiring environment shaped by its own fundraising and revenue trajectory. A company at $30 billion in revenue run rate that is also expanding headcount is not evidence that AI adoption is broadly job-neutral. It is evidence that AI adoption at scale in a fast-growing AI company looks different from AI adoption at a manufacturing plant or a mid-size law firm.

The more durable observation may be structural rather than numerical. When the execution layer of knowledge work becomes cheap to automate, the scarce resource is not execution anymore. It is the capacity to set direction, evaluate outputs, and decide what to do next. Rao is describing something that economists of automation have theorised for years: that generalised AI does not eliminate the need for human judgment but concentrates it. The question his account raises is whether that concentration is broadly accessible or whether it requires a type of institutional and educational capital that most organisations do not have.

Anthropic is preparing for a future IPO, with speculation that its valuation could eventually approach $1 trillion. Rao has overseen the financial scaling of what may be the most capital-intensive product cycle in the history of software. He described compute as "the lifeblood of our business — it is the most important thing in the company." The company recently struck a partnership with SpaceX to access the Colossus 1 supercomputer for model training.

That context matters too. Anthropic's AI adoption story is inseparable from its compute procurement story, which is inseparable from its valuation story. Rao has incentives to make AI adoption look productive and orderly. The data he cited is real. The framing is his.

What he described is worth taking seriously precisely because it is a first-party account from a company operating at the frontier of what it is describing. If execution-layer AI adoption at Anthropic is producing a need for more human oversight, more talent density, more judgment calls about what AI outputs to accept or reject — that is a signal about where the labour market pressure points actually are, even if the sample size of one fast-growing AI company cannot settle the broader debate.

The doomsayers and the productivity optimists have been arguing about this for years. Rao is not settling the argument. He is just adding a data point from inside the company that both sides will cite.