By This Hour Business Desk
Anthropic has presented an unusually detailed account of an AI company using its own products: more work handled by Claude agents, less entry-level technical hiring and a workforce in which staff increasingly spend time directing automated systems rather than carrying out every task themselves. The company argues that the pattern has lifted output and offers an early indication of how workplaces could change as AI agents spread.
Outside researchers cited in coverage of the report question both the scope of that conclusion and the way the company reaches it. Their central objection is not that AI use has no effect on work, but that one company’s internal experience—particularly a company selling the technology and approaching a possible stock-market debut—cannot by itself show what is happening across industries. They also challenge the implication that AI explains recent changes in hiring and employment.
The disagreement matters because Anthropic’s report puts concrete workplace numbers behind a much broader corporate argument: that increasingly capable AI systems will alter labor demand, especially for junior technical workers. Its account may be a useful description of choices inside Anthropic. It is much less certain whether those choices can be treated as a forecast for software work, sales, contracting or the wider economy.
Anthropic describes agents taking on a larger share of work
In its report, Anthropic said deeper internal use of Claude agents had changed how jobs were organized. It described gains in code produced per employee, sales closed per representative and contracts processed. The company attributed those gains to delegating portions of coding and sales-related work to AI agents, while employees took on more responsibility for directing and reviewing the systems’ activity.
Anthropic said that, from May 2026, its AI agents logged more work hours than human employees. That measure is striking, but the report as described does not settle what an agent “work hour” represents in comparison with a human hour, how those hours were counted, or whether the logged activity translated consistently into completed work of equivalent quality. Those questions matter when the figure is used to suggest that agents have overtaken people in an operational sense rather than simply generated more recorded system activity.
The company also said an analysis of employee chat summaries found that workers devoted about 22% of their time to managing Claude agents, more than twice the reported level a year earlier. It characterized human attention as a principal constraint in an agent-heavy workplace. Under that framing, the scarce resource is not the ability to generate drafts, code or other outputs, but the human capacity needed to set direction, judge results, resolve ambiguity and decide which work deserves attention.
That account points to a consequential shift even if Anthropic’s productivity claims are taken only as company assertions. AI may not simply remove tasks; it may recast jobs around supervision, selection and escalation. A manager or engineer responsible for several agents could produce more, yet also face a more fragmented workload. The report’s figures do not show how employees experience that trade-off, whether management time has displaced other work, or whether the systems reduce mistakes and rework over a longer period.
Fewer junior hires sit at the center of the labor claim
Anthropic linked its increased AI use to a marked change in its technical workforce. It reported that entry-level technical hires fell from nearly 30% in March 2025 to 9% in the period covered by the report. The share of overall headcount in that category, it said, declined from about 21% to 14%. The company attributed the changes to its use of AI agents.
Those figures raise a practical concern beyond Anthropic’s own payroll. Entry-level jobs have traditionally provided a route for people to gain the judgment, institutional knowledge and collaborative habits required for more senior positions. If employers automate a larger share of tasks once assigned to junior staff, they may reduce opportunities to enter the profession. The longer-term issue is not only whether current teams can work with agents, but how organizations would develop the experienced workers they expect to need later.
Still, the reported timing and the company’s attribution do not establish causation. Hiring shares can move for reasons unrelated to AI: a company can change its growth plans, alter the mix of roles it needs, reduce hiring generally, or make a different strategic choice about experience levels. A falling share of junior staff is not on its own evidence that AI made those workers unnecessary. The available account does not describe a comparison group, external audit or other method that would separate AI use from such factors.
Morgan Frank, a researcher who studies how AI and technological change affect workers, offered a broader reason for caution. His work, described as being in peer review, examines job-market outcomes associated with AI exposure and indicates that deterioration began before the major generative-AI releases named in the discussion. On that view, recent labor-market weakness cannot readily be assigned to ChatGPT, Claude or related products merely because the downturn coincided with growing public interest in them.
Frank’s argument leaves open the possibility that AI is affecting particular firms or occupations. Rather, it challenges a simple explanation for layoffs and weaker hiring. He identified other possible drivers, including employers cutting positions added during an earlier hiring period and the use of AI as a rationale for reductions a company might have made anyway. That latter possibility, often described as AI-washing, is difficult to test from public claims because it concerns managerial motivations as well as employment totals.
A company case study is not an industry verdict
Anthropic’s report carries an inherent tension. It is a case study of a company that makes and deploys AI agents at unusual depth, which could make it informative about what intensive adoption looks like. It is also an account produced by a business with a direct interest in demonstrating that its products improve organizational performance. Those two facts do not make the report wrong, but they limit what can be inferred from it without outside examination.
Julián Posada, a Yale social scientist, questioned the company’s incentive to publish the findings as it seeks investment. He argued that the report lacks the independent assessment expected of an economics study and that charts and internal figures leave important gaps. In particular, the numbers do not convey how employees engage with the models or how they assess the changes in their working lives.
That criticism goes to the distinction between measuring activity and measuring durable business value. More code, contracts or logged agent hours might reflect a meaningful increase in productive capacity. They might also create a larger review burden, change the type of work performed, or obscure quality questions that emerge later. The source material provides no independent validation of Anthropic’s internal measurements, no comparable data from other employers and no employee survey that would resolve those possibilities.
Nor is it clear that an AI developer’s internal deployment can be generalized to employers that have different products, workforces, customer obligations and tolerance for error. A company building AI systems may have staff, workflows and incentives especially suited to using agents. The report’s scenario of small teams or even individual founders directing large groups of agents is therefore a possibility advanced by Anthropic, not a demonstrated outcome for the wider business economy.
Software work may not fit a simple attention-bottleneck model
Dylan Baker, a research engineer at the nonprofit Distributed AI Research Institute, focused on the report’s description of human attention as the limiting factor. He argued that this language mischaracterizes software development, which depends on social and interpersonal work as well as technical output. Designing a product, making architectural choices and coordinating among people require judgment that cannot be reduced neatly to a queue of agent tasks awaiting review.
Baker also questioned the combination of fewer entry-level hires, more AI-generated development work and remaining engineers charged with managing agents. He described the approach as potentially shortsighted. His concern is that organizations could save effort in the near term while weakening the training pipeline and placing experienced staff in a role dominated by oversight rather than the collaborative work through which teams develop products and people.
The contrast with Anthropic’s framing is clear. Anthropic portrays human supervision as the crucial scarce input in a system where agents perform a growing volume of tasks. Baker sees the human dimension as more fundamental: software development is not merely execution that can be shifted away from people, but a process of communication, design and collective decision-making. The available information does not establish which description better captures Anthropic’s internal results, much less the sector as a whole.
Investor scrutiny could sharpen the debate
The report arrived one day after Anthropic said it would commit 5% of its shares to a public-benefit fund. The company did not provide details of that commitment in the supplied account. A separate report on the pledge says that the fund’s governance, selection process and timetable for distributions have not been set out publicly.
That timing has contributed to critics’ concern that the employment report also serves a business narrative ahead of a potential public offering. Anthropic did not respond to the request for comment described in the source material. The company’s report may nonetheless give investors and workers a window into how it says it is using Claude internally. The more ambitious claim—that the window reveals the future of work—requires evidence from beyond Anthropic.
For now, the report has not been independently corroborated. Its workforce counts, productivity measures, agent-hour calculations and causal conclusions are reported as Anthropic’s own findings, while the objections from outside researchers identify significant unresolved questions about method, motivation and generalizability.
Reporting notes
What is confirmed: Anthropic reported greater agent activity, more employee time spent managing agents and a lower share of entry-level technical hires. These are company-reported figures.
Why this matters: The report links AI adoption to productivity, supervision demands and reduced junior technical hiring, all central questions for employers and workers.
What remains unclear: Whether AI caused Anthropic’s hiring changes, how its measurements were defined, and whether its experience applies beyond the company remain unresolved. This report is based on one source and has not been independently corroborated.
