The Entry Point Is Closing
Anthropic published a labor market study last week. Their CEO has been saying for months that AI will eliminate 50% of entry-level white-collar jobs within five years. Their own researchers just found limited evidence this is happening.
Both things are true. And the gap between them is what actually matters.
Maxim Massenkoff and Peter McCrory, two Anthropic economists, built a new metric called "observed exposure" — not what AI can theoretically do to a job, but what it's actually doing, measured from real Claude usage data. The full paper is on Anthropic's site.
The top-line finding: no systematic increase in unemployment for highly exposed workers since late 2022. The "Great Recession for white-collar workers" the paper names as a scenario hasn't happened.
But buried in the data is a number that landed differently for me.
Entry-level hiring for workers aged 22–25 in AI-exposed occupations dropped 14% since 2022. That effect was statistically significant — barely, but it was there. And it was absent for workers over 25.
Companies aren't laying off their senior people. They're quietly not replacing the junior pipeline.
The Adoption Gap Is a Countdown
For software and technical workers, AI can theoretically handle 94% of tasks. Anthropic's actual usage data shows it covering 33% in practice today.
Most people read that 3x gap as reassuring. I read it as a timeline.
The gap isn't stable. It was probably 5% two years ago. HBR's companion research found that job postings in the most exposed roles are already down 17%, while roles requiring judgment and human-AI collaboration are up 22%. Companies aren't waiting for AI to reach 94% before they restructure headcount. They're doing it now, at 33%.
The entry-level hiring signal is what that looks like in practice. Not layoffs — just a smaller door in.
The Ladder Is Compressing, Not Disappearing
The dominant narrative around AI and jobs splits into two camps: AI is going to eliminate everything, or AI is not a big deal because workers have always adapted. Both miss what's actually happening.
The better framing, from Ethan James's analysis of the paper: AI is eliminating the predictable junior-to-senior career ladder that historically trained expertise. Not the job at the top of the ladder. The rungs at the bottom.
In law, consulting, finance, and engineering, the traditional path looked like this: take a junior role that involves a lot of structured, repetitive cognitive work — research, data processing, document review, analysis — and use it as the foundation to develop judgment over time. The apprenticeship model.
That model only works if there are enough junior roles to support it. When AI handles the bulk of the structured work, companies stop needing as many juniors to do it. Senior practitioners gain leverage. Entry-level pipelines shrink. The ones who do get hired need to operate above the automation line faster.
The paper captures something real when it notes that the most exposed workers are "nearly four times as likely to hold a graduate degree." This isn't AI disrupting low-wage work. It's disrupting the credentialed knowledge economy that most career transitioners are trying to enter.
What This Means If You're Making the Move
I transitioned from management consulting into building technical products. I did it by learning to code to solve specific problems, not by working through a series of increasingly senior analyst roles. That path looks better, not worse, in the current environment.
Here's the practical read on the data for anyone making a similar move right now:
The traditional entry-level path is higher risk. Junior analyst, junior associate, entry-level PM — roles where the primary value you're delivering is throughput on structured cognitive tasks. That's the part of the job that's already automatable. Hiring into that pipeline is what's slowing down.
The builder path is lower risk. Not because AI can't touch it — it clearly can — but because it's above the automation line. A solo operator using AI as leverage to produce work that used to require a team is operating at a different level than someone competing with AI on task completion speed. You're not doing the 33% AI already handles. You're doing the judgment layer that determines what the 33% gets pointed at.
The gap is your window. That 3x adoption gap (94% theoretically possible, 33% actually deployed) is real and it's closing. But it means there's still time to position above the automation line before the entry-level door closes further. The transition that made sense five years ago (get in the door at a junior level, learn on the job) requires more deliberate positioning now. Get in above the rote work, or plan to move above it very fast.
The Amodei Tension
I want to sit with the contradiction for a moment, because it's worth thinking about.
Dario Amodei said in a May 2025 interview that AI would eliminate half of entry-level white-collar jobs in one to five years. His own company's researchers just published data showing limited evidence this has started.
The Register ran a sardonic headline. Tech Twitter used the gap to dismiss AI job concerns entirely.
I think both responses miss something. Amodei is in the prediction business — he's talking about where this goes, not where it is today. His researchers are in the measurement business — they're measuring what's happened, not what's coming. The 14% hiring slowdown for 22–25-year-olds is consistent with both. Companies believe Amodei's prediction enough to quietly stop filling the junior pipeline, even while the aggregate employment data hasn't moved.
That's the version of this story that actually matters for someone planning a career move. The lag between "companies stop hiring juniors" and "unemployment numbers change" can be years. By the time it shows up in aggregate statistics, the window for positioning above it has probably narrowed considerably.
The entry point into white-collar technical work isn't closed. The 14% hiring decline is modest. The adoption gap is real and will close slowly. None of this is cause for panic.
But "slowly" is doing a lot of work in that sentence. The time to position above the automation line is before the gap closes, not after.
Build something. Ship it. Own the outcomes layer. That's the move.