Anthropic's own workforce data: more agent hours, fewer junior hires. What it means for your office

What Anthropic published

On Oct. 16, 2026, Anthropic published "How employment is changing at Anthropic," a report on its own workforce by Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory. It covers early 2025 through August 2026 and draws on headcount records, internal logs of how staff use AI agents, manager surveys and employee interviews.

It's unusual for a company to publish this about itself, and more unusual for an AI company. Here are the numbers that matter for an office manager:

• Junior hiring fell fast. The share of junior staff in technical departments slipped from 21% to 14% since early 2025. The share of technical hires who were junior fell even faster, from 27% to 9%.
• Agents outwork people, on paper. By August, AI agent hours outnumbered employee hours six to one.
• Managers hired less. About a third of machine learning managers and a fifth of software engineering managers said they were hiring less because of AI. A majority said they were hiring for different skills.
• People became the bottleneck. Claude sessions spent around 80% of their time waiting for a person. For non-technical staff, reading everything their agents produced would take half their work week.
• Instructions piled up. Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later.

Junior share of technical hires at Anthropic

Anthropic's own HR data, early 2025 vs. 2026. One company, not an industry figure.

1. Early 202527%
2. 20269%

Source: Anthropic, How employment is changing at Anthropic, Oct. 16, 2026.

The part that isn't about engineers

Most coverage will focus on coding. The more useful parts for a manufacturer or distributor are about sales and office work. One account executive told the researchers that Claude automations handled email drafting and connections to customer management software, replacing around 20% of their prior tasks. Another said the same kind of automation freed time for face-to-face work, letting them talk to 30% more customers than before. Those are two people's accounts, not measured results, but they describe the pattern well: the software takes the typing and logging, the person takes the conversations.

Hiring criteria are moving too. Managers mentioned a candidate's track record with AI tools in about 5% of technical assessments and 7% of non-technical ones. By 2026, Anthropic says those rates had tripled for technical roles and quadrupled for non-technical roles.

The agents waited on people about 80% of the time. Review, not software, sets the pace.

Why the bottleneck matters more than the six-to-one

The headline number is agent hours. The more useful number is the 80% waiting time. It says that even at the company that builds the AI, the limit on getting more done was a person reading, checking and deciding.

That matches what the evidence keeps showing about offices. G2's survey this month found that checking AI agents eats much of the time they save. At a 40 to 400 person company, the order desk or AP desk doesn't get faster because software drafts 200 orders. It gets faster when the drafts arrive in a form a person can check in seconds: the PO next to the draft, unsure fields flagged, one click to approve.

The 8,500 written instructions point the same way. Those are, in effect, procedures written for software. A company that has its order entry and AP rules written down, even in a rough list, starts ahead.

The other side

There are good reasons not to read too much into this.

• Anthropic sells the agents. A report saying agents outwork staff six to one is also good marketing for the product. Read it with that in mind.
• It's one unusual company. Anthropic tripled in size in the last year. The report says many of its employment practices can't be copied across the economy.
• The measures are loose. A blog analysis by Gene Dai points out that the six-to-one figure compares agent activity to an assumed 40-hour week, not hours staff actually worked, and that the report doesn't estimate jobs lost to Claude. The report also admits that people who adopted agents early weren't chosen at random, and that more code doesn't necessarily mean better code.
• AI isn't the only thing hitting junior hiring. Economist Jed Kolko called the research on AI and young workers' hiring "collectively inconclusive," as reported by ABC News in May.

Fair enough. Still, the direction lines up with the broader data. Stanford's payroll study finds the gap for young workers in AI-exposed jobs comes mostly from less hiring, and the Fed's October Beige Book reported firms delaying hiring after AI gains. We pulled those together in Will AI replace office staff?

What to do with this in your office

1. Expect hiring to change before headcount does. The first effect is the entry-level job that doesn't get posted, or gets written differently. Decide on purpose, not by default. Our hire-or-automate calculator helps.
2. Budget the reviewer's time. If software drafts orders or invoices, someone has to own checking them. Make that a named part of a job, with a screen built for fast checking.
3. Write the rules down. Which customer means which part number, which vendor always bills freight separately. That list is what an automation runs on, and it protects you when a veteran retires.
4. Train your people on the tools. If Anthropic's managers are now asking about AI experience in non-technical hiring, your current team is worth training before you go looking outside.
5. Keep a person approving. Nothing in this report says the review step goes away. It says the review step is the job.