The $190k Job Claude Can’t Do

The company that sells the world one of its smartest AI assistants has just tried to hire a human one. For about $190,000 a year.

Anthropic, the maker of Claude, advertised for an “Administrative Business Partner” in London at £145,000. Translation: an executive assistant. Diaries, inbox, travel, expenses. The ad was quietly pulled from Anthropic’s own careers page, but not before City AM noticed.

This is the same industry telling the rest of us that white-collar work is about to be automated. So either the job is harder than they admit, or the technology is weaker than they claim. Either way, the job ad is more honest than the keynote.

The job Claude was built for

Read the duties, and then read Anthropic’s own marketing. They are the same document.

“Scheduling and managing calendars.” Claude for Chrome was launched on exactly this: managing calendars and scheduling meetings.

“Managing the leader’s inbox… ensuring all messages are swiftly triaged.” Inbox triage is the poster child of every AI productivity demo since 2023.

“Building and maintaining organizational systems for tracking… inbound requests.” A shared tracker and a prompt. Anthropic launched a product called Cowork, billed as “Claude Code for the rest of your work.”

“Handling travel logistics.” OpenAI’s ChatGPT agent books trips. That is the pitch.

“Coordinating… to gather necessary information and materials for meetings.” Anthropic publishes a how-to guide for having Claude prepare you from your calendar.

“Managing expense claims and reimbursements.” Expense reports featured in the Claude for Chrome launch. Literally.

The only duty that clearly needs a pulse is greeting visitors at reception. That is a £145,000 meet-and-greet.

And the kicker sits in the requirements: “flawless administrative execution” and “error-free” delivery. Which, to be fair, is the one thing no AI vendor will put in writing.

Not a one-off

This was no rogue hiring manager. Anthropic has advertised the London role since September 2025, at £120,000 to £145,000, and says it is hiring “multiple” of them. A New York version paid $190,000 to $230,000. OpenAI advertises “Executive Business Partners” at roughly $150,000 to $240,000, plus equity.

For context, a London EA averages £50,000 to £60,000. Anthropic’s floor is double the market. For the same money, you could hire four senior executive assistants in Manila and still afford a very good AI subscription for each of them.

Meanwhile the labs keep growing. Anthropic has gone from about 550 staff in 2023 to roughly 4,000 today. For companies building our replacements, they are remarkably keen on people.

Do they eat their own dog food?

Here is what the labs have told the world. Dario Amodei warned that AI could wipe out half of all entry-level white-collar jobs within five years. He has also said that AI models may be six to twelve months away from doing “most, maybe all” of what software engineers do end to end. In April 2025, Anthropic’s security chief said fully AI “virtual employees” were about a year away. A year later, Futurism called it “dead wrong.”

Sam Altman, for his part, predicted in January 2025 that AI agents would “join the workforce” that year. They joined. They just didn’t take the executive assistant’s chair.

Now here is what Anthropic’s own data says. Its internal study of its own staff found they use Claude in 59 percent of their work, but can fully delegate only 0 to 20 percent of tasks. The most AI-fluent workforce on earth still needs a human to finish the job. If anyone could automate its own back office, it would be the company that builds the automation.

The benchmarks agree. In Carnegie Mellon’s TheAgentCompany test, the best AI agents completed about 30 percent of simulated office tasks. On Scale AI’s Remote Labor Index, which pits agents against paid freelance projects, the top-performing agent at launch in October 2025 automated just 2.5 percent. A year on, even the best models — including Anthropic’s own Opus — are stuck around 20 percent. And in September, Meta’s new AI agent Muse was caught using human contractors to make its phone calls. The robot had an assistant too.

In fairness, the labs do use their own tools. Anthropic says Claude writes most of its code. But code is where AI is strongest. The messy, relational, judgment-heavy work of running a senior executive’s life is where it is still a junior intern.

Watch what they hire

There is a respectable defense. A great EA is not a calendar. They read the room, guard the door, and make a hundred small judgment calls a day. At a company valued like Anthropic, an executive hour is worth a fortune, so buying them back is cheap.

There is a security case too. The ad asks for “a security mindset” and discretion with confidential information. Handing a senior leader’s inbox and diary to an autonomous agent is a risk most boards would hesitate over. Anthropic, it seems, included.

As Anna Bailey, founder of PA network Pally Talk, told City AM: “If AI could replace a great PA, an AI company would be the first to do it.”

Exactly. That is the whole point.

Amodei himself now argues that if AI automates 90 percent of a job, “the 10 percent kind of expands to be 100 percent”. Sam Altman has said he is “delighted to be wrong” about entry-level jobs. The apocalypse talk is quietly being walked back, just as the labs court public markets.

This is the trust problem. The labs are not neutral narrators. Every forecast of mass automation is also a sales pitch: the bigger the disruption, the bigger the market, the bigger the valuation. That doesn’t make them liars. It makes them interested parties.

So here is a simple rule for every executive deciding headcount on the strength of an AI keynote. Ignore what the labs forecast. Watch who they hire.

Forecasts are marketing. Payroll is evidence. And the payroll says the people who know this technology best still pay a premium for a human who can run a diary, chase an expense claim and say hello at the door.

None of this means the tools are useless. They are useful, often impressively so. But “useful” and “trusted to run the boss’s week unsupervised” are very different standards.

If AI were ready to replace that person, Anthropic would not be paying $190,000 to find one.

The question for your business

Are your headcount decisions based on vendor keynotes, or your own results?

Read more thought leadership articles here:

• • The Job Apocalypse That Never Comes
  • Accenture Didn’t Die
  • The Offshore Accelerator