AWS CEO on cutting junior hires: “the dumbest thing I’ve ever heard

The studies landed within weeks of each other, and they all point at the same spot on your org chart. Not your senior staff. Your juniors, your project coordinators, the entry-level seat where people used to learn the job. That seat is shrinking, and nobody made a plan for what replaces it. Let me catch you up.

The deep cut

• Measure tasks, not titles. The spec-and-check test from Patrick Neeman scores exposure a job description hides.
• The entry seat is where the apprenticeship lived. Anthropic cut junior technical hires from 27% to 9%, and AWS's Matt Garman called gutting that seat the dumbest idea he's heard.
• Reading is now the job's ceiling. Anthropic's agents logged six hours for every human hour, and nobody has time to review it all.

Score the week, not the badge

A title tells you almost nothing about who is at risk. Neeman's point is blunt: AI predicts an outcome without looking at your LinkedIn. He offers a test you can run in a review. Can you describe a finished version of a task in two sentences? Can someone check it faster than it took to make? Two yeses means that task is exposed, whatever the title on top of it.

Weight the list by hours. GPTs are GPTs, the 2023 OpenAI and Penn study, found 80% of US workers could see a tenth of their tasks hit and 19% could see half. Averages hide the person. A senior designer whose week is status decks and empty-state variations is carrying junior exposure with a bigger paycheck. Run the test on your own team before someone else runs it for you.

The seat where people learned the job

Here is the number that should stop you. At Anthropic, junior technical hires fell from 27% to 9%, and the junior share of technical staff dropped from 21% to 14% since early 2025, per TechCrunch's read of the report. Stanford's "Canaries in the Coal Mine" found a 13% relative drop in employment for 22- to 25-year-olds in the most exposed jobs. The hiring, not firing, is where the cut shows.

The logic is simple and wrong. Entry-level work is the most specified work on a team, handed down with clear instructions, which is exactly where a model does best. So why hire the junior? AWS CEO Matt Garman gave the answer in August 2025, calling it "the dumbest thing I've ever heard." Juniors cost the least and use the tools the most, he said, and cut them and nobody learns the work in ten years. The skill to supervise an agent comes from years of doing the job it now does.

The reading tax nobody budgeted

The second shift is quieter and just as expensive. By August 2026, Anthropic's agents logged six hours for every human hour, and reading everything they produced would eat more than half the workweek, even for people who don't code. That is the reading tax. One engineer compared managing agents to babysitting, where a juice box and a cut finger both arrive at full volume.

This turns your team into managers whether you planned it or not. A fifth of surveyed Anthropic staff said their role shifted from doing the work to managing Claudes. Meeting time rose 20% because people need face-to-face trust to coordinate. The catch from METR's 2025 trial: 16 experienced developers took 19% longer with AI after expecting to save 24%. If people can't see how long review takes, they can't tell when they've stopped doing it. Budget reading time like page load time, and log it.

Hire for the override

What goes in a portfolio changes with all this. A grid of polished final screens shows what a model can now produce on its own. The design director who has reviewed hundreds of portfolios says the single most common reason strong designers don't get interviews is 40 beautiful screens with no decisions behind them. He reads for judgment, not craft. Show the wrong turn at week two, the cut feature, the stakeholder who pushed back.

Anthropic's own interview scorecards back this up. Mentions of candidates' creative AI use roughly tripled for technical roles, while domain expertise shrank as a share of the assessment. In interviews there, candidates get a page of agent output with planted problems and get watched for which they catch. Meanwhile the technical bar is climbing: in a sample of design engineer roles, 86% mentioned React and median base midpoint hit $235,000. The person who understands the problem, not the model, is the one you want.

Who owns the deadline now

One more role worth mapping while you redraw the org chart. Jeff Gothelf went looking for project managers on product teams and couldn't find them, until 69 comments taught him the work never left. It got absorbed by the product manager, often at midnight before the steering committee. His test: where the deadline comes from decides who owns the work. A hardware or regulatory deadline is real and immovable and needs a named owner. An internal date on a slide is a variable.

Run the same exercise the studies push. Write the task, name the person who did it last quarter, mark the deadline real or internal. Do it before you assume the work vanished because the title did.

Three questions for your team

• If we ran the spec-and-check test on every role this quarter, which senior titles are sitting on junior work, and what one real decision can each of them own instead?
• We cut or never opened the junior seat. Where exactly does someone on this team learn the judgment we now say we hire for?
• What is our reading budget per workflow, who defends it, and how will we know when review has stopped happening?