Anthropic says AI is reshaping work, but its evidence has limits

What Anthropic says changed

Last year, CEO Dario Amodei warned that AI could eliminate half of entry-level office jobs and write 90% of code within six months. Neither prediction came true.

Anthropic now says it has been hiring fewer junior technical workers as it relies more heavily on Claude agents. Since May 2026, agents have logged more working hours than employees, according to the report. Staff now spend about 22% of their time managing them, more than twice the share a year earlier. Anthropic calculated that figure from summaries of employees’ conversations with its models.

The hiring numbers point in the same direction. Since March 2025, shortly after the release of Claude Code, the share of junior technical specialists among new hires fell from almost 30% to 9%. Technical specialists’ share of the overall workforce declined from about 21% to 14%.

Anthropic says the shift has increased output per employee: more code, more sales per salesperson and more contracts processed. It describes employees as “a thin layer over a growing number of AI agents.” The report also imagines a future in which one-person startups use groups of agents to generate billions of dollars in revenue.

A company report is not an industry forecast

The report offers a rare glimpse into the operations of a major AI company. But researchers quoted by the Guardian question whether one company’s experience can predict what will happen across industries. Some also suspect that Anthropic has an incentive to present AI as transformative as it prepares for a potential public listing.

Morgan Frank, a Case Western Reserve University researcher studying AI and technological change in the labor market, recently published a paper now under review. It finds that labor-market deterioration in areas with widespread AI use began well before ChatGPT launched in November 2022 and Claude followed a few months later. Unemployment-benefit claims point to a downturn beginning in early 2022, months before ChatGPT’s release.

Frank said ChatGPT’s launch did not appear to coincide with a sharp rise in job losses or a sudden worsening for workers. Recent layoffs, he argued, may have other explanations: companies are cutting pandemic-era hiring, and some executives attribute layoffs to AI even when they would have happened anyway. That practice is known as “AI washing.”

Yale sociologist Julian Posada also questions the report’s perspective. He said it reads like the account of a company seeking to sell shares and attract investors. Unlike an economic study, Posada noted, it was not independently checked by external auditors. The charts and figures also say little about how Anthropic employees feel about their work or how they actually interact with the models.

Dylan Baker, a lead research engineer at the nonprofit Distributed AI Research Institute, wants to know what Anthropic’s developers think of the changes. He called it short-sighted to hand some software development to AI, hire almost no junior staff and ask the remaining engineers to manage agents. He also disputed the description of human attention as a “bottleneck,” arguing that software development is substantially collaborative and interpersonal.

The report quotes an unnamed engineer describing work with “a whole army of Claudes, each asking for my attention.” Baker said that does not sound like a productive environment, or one likely to produce thoughtful, well-designed products.

The unanswered question before an IPO

The report recommends that technology companies give employees an ownership stake. Anthropic argues it is hard to imagine AI being adopted in workers’ interests if they do not share financially in its success. That proposal sits uneasily beside its own figures on declining junior hiring: an ownership stake does not answer how new workers will gain the experience the report says may become harder to acquire.

The timing also matters. Anthropic published the employment report the day after promising to direct 5% of its shares toward “the public good,” without detailing what that means. The company may go public as soon as November. Critics have long argued that philanthropy can help technology billionaires, including Bill Gates and Mark Zuckerberg, strengthen their reputations.

A draft S-1 filing disclosed a net loss of $80 billion in the first half of this year, and the company received a low credit rating in June. Investors now value Anthropic at $1.5 trillion, below earlier forecasts of $2 trillion and above, as revenue growth slows. Its cloud-computing partner CoreWeave is burdened by AI-related debt, and its shares are falling sharply, adding to concerns that the AI bubble could deflate.

I think the report’s most consequential omission is not another productivity metric but the workers’ account of what this system feels like from the inside. As an IPO approaches, claims that AI is improving how work gets done will face closer scrutiny. Baker’s concern reaches beyond whether the bubble bursts: engineers will feel the consequences either way, and the industry may have to rethink how much AI progress depends on human labor.