AI Is Hollowing Out the Bottom Rung, and That Breaks the Whole Ladder

AI isn't taking the jobs at the top. It's quietly removing the bottom rung: the boilerplate code, the first-draft decks, the document review that juniors learned on. Companies aren't firing young workers. They're just not backfilling them. That looks like efficiency now. In five to seven years it looks like a missing layer of senior talent nobody can buy back.

That's the part almost everyone is getting wrong. The headline panic is "will AI replace me." The real threat is slower and worse: AI stops your replacement from ever getting trained. The career ladder was never just a metaphor. It's shared infrastructure. Juniors do the reps, seniors supply the judgment, and the reps are how juniors become the next seniors. Pull out the reps and the whole thing stops manufacturing experts. You can't 10x your way out of that in a quarter.

What does the data actually show about AI and entry-level jobs?

The numbers aren't noisy. They're being read wrong. New role starts by workers with under a year of post-graduate experience fell 50% between 2019 and 2024 at the largest public tech firms and maturing startups, with a 25% drop from 2023 to 2024 alone. Stanford's Digital Economy Lab found that early-career workers aged 22 to 25 in the most AI-exposed occupations have experienced a 16% relative decline in employment, even after controlling for firm-level shocks. US computer programmer employment dropped 27.5% between 2023 and 2025, per BLS data.

Here's the read that matters. Companies aren't running mass layoffs of the young. They're just not refilling entry-level seats when someone leaves. No spike, no announcement, no news cycle. The ladder loses a rung and nobody notices until they reach for it.

Was junior work ever really just busywork?

No. That's the costly misread baked into every "we automated the grunt work" press release. The grunt work was the apprenticeship.

Weak version of the story: "AI is replacing junior devs." Strong version: a 2023 new grad at a Series B startup used to own the ticket queue. Bug triage, boilerplate React components, unit tests for legacy code. By 2025 that queue runs on Copilot and an internal agent. The company fired no one. It just stopped backfilling when the person left. The role wasn't eliminated. It dissolved.

Notice what dissolved with it. That ticket queue was low-stakes, high-volume reps. Writing a hundred mediocre components is how you earn the taste to know a good one. First-draft decks taught structure. Document review taught what actually matters in a contract. The output looked like cost. The function was deliberate practice. Automate the practice and you keep the day-one productivity while quietly deleting the year-five engineer.

Why is this a senior problem, not a junior problem?

Because the pipeline has a five-to-seven-year lag, and you can't hire your way out of a lag.

When a Big 4 firm cuts graduate intake today, the gap doesn't show up today. It shows up around 2030, when that cohort should have become your mid-level managers and the cohort simply isn't there. The hiring manager congratulating themselves on a leaner team right now is the same person who, in a few years, will be writing a desperate req for "senior talent shortage" they personally caused.

This is the Praxy read on it. Compounding beats tenure, and the bottom rung is where compounding starts. Skip it and the math never recovers. Companies optimizing entry-level headcount this quarter are borrowing against senior talent they'll need later, at an interest rate they haven't priced. The bill comes due on a delay, which is exactly why nobody's panicking and exactly why they should be.

Which sectors are actually getting hit, and how?

It's uneven, and the unevenness is the most useful part of the picture.

Sector	What's happening	The tell
Tech	New-grad role starts down 50% since 2019; programmers down 27.5%	New grads are now just 7% of Big Tech hires
Accounting	UK Big 4 grad listings down 44% year-on-year	KPMG cut intake from ~1,399 to 942; EY delayed start dates a third straight year
Consulting	Accenture cut over 11,000 staff in a quarter while AI/data headcount grew from 40k to 77k	An explicit swap of generalist bodies for AI-capable talent
Law	Graduate employment hit a record high; firm headcount grew	The counterpoint that proves the rule (below)

Strong example, accounting: KPMG UK cut its graduate intake roughly 29 to 33% in a single year, and EY delayed 2025 graduate start dates to March 2026, the third year running. These aren't layoffs. They're deferred on-ramps. From the outside the pipeline looks fine until the year you need mid-level managers and the cohort isn't there.

Doesn't law disprove the whole thesis?

It looks like it does, which is why it's worth sitting with. Law school graduate employment hit a record 93.4% for the Class of 2024, the highest NALP has recorded since it began tracking the data in 1982, and the share in jobs requiring bar admission grew to an all-time-high 84.3%. The jobs exist. So the hollowing thesis doesn't hold uniformly, and anyone telling you all white-collar work is collapsing is overselling.

But look one level down and law makes the deeper point sharper. The first-pass document review, contract analysis, and legal research that used to train associates is being compressed by AI. Associates now edit machine-generated drafts instead of writing from scratch. The roles survived. The reps didn't. You can keep the title and still gut the training that made the title mean something in five years. Law is the warning that "are people getting hired" is the wrong question. The right one is "are the people getting hired learning the thing that makes them senior."

Same nuance on the tech numbers. The 27.5% collapse is in commodity coding. Employment for software developers, a distinct and more design-oriented role in the government data, barely moved over the same period, falling just 0.3%. Task substitution is not role elimination. The post is only accurate when you hold that line.

Is anyone fixing this, or is it all bleak?

At least one major employer is doing it on purpose, and it's the most instructive move in the whole story. In February 2026, IBM announced it would triple US entry-level hiring. The lone big tech employer openly reinvesting in the pipeline.

Read the fine print and it confirms the thesis instead of denying it. IBM didn't restore the old apprenticeship. It redesigned those jobs away from coding and toward customer-facing work, explicitly because the coding reps got automated. That's healthy pipeline management: if AI ate the old reps, build new ones. The open question is whether orchestrating AI, engaging clients, and specifying features build the same deep expertise the old grind did. Nobody knows yet. What's clear is that IBM is doing the deliberate work almost everyone else is skipping while calling the skip "efficiency."

The trade-off nobody is naming

Automating entry-level work is genuinely cheaper this year. That's real, not spin. A leaner team ships the same output at lower cost, and the savings land on this quarter's books where everyone can see them.

The cost lands on a delay and on someone else's books. You're trading a visible gain now for an invisible liability in five to seven years: a missing mid-level cohort, a judgment layer you can't backfill on demand, and a more expensive scramble when the shortage finally shows up. For the individual it's blunter. The on-ramp is narrowing at the exact moment it used to be widest, which is how the need-experience-to-get-experience trap gets engineered into the whole system. AI is one cause among several here. Grad oversupply was visible before ChatGPT, with 52% of the Class of 2023 underemployed a year out, and there's a separate case that hiring inflation, not AI, killed the entry level. But AI is the accelerant, and it's the one squeezing the training, not just the headcount.

What should an early-career professional do right now?

Stop competing for the rung that's being removed. Build the judgment layer that can't be automated, and get there faster than the old track would ever allow.

• Aim for the work AI does badly. The reps that still compound are the judgment reps: deciding what to build, catching where the model is confidently wrong, owning the client relationship. Volunteer for the cases AI gets wrong. That's where expertise still accrues.
• Use AI as an apprenticeship multiplier, not a crutch. Let it draft, then you direct, edit, and pressure-test. A junior who reviews and corrects ten AI outputs a day is building taste faster than the 2015 junior who wrote one deck by hand. Same skill, more reps, if you stay in the loop instead of rubber-stamping. This is the real version of the tired line that the person using AI takes the job, not the AI itself.
• Pick roles and moves that still build leverage. Some "entry-level" titles are dissolving; others are genuinely redesigned, like IBM's. Tell them apart before you accept. Consistency on the right reps beats intensity on the wrong ones.
• Get one story going. You need one real story for a career to go big. Hunting for the cases AI can't close, and being the person who closes them, is how you start writing it.

This is the agency read on a bleak chart. The ladder is being hollowed from below. That's a circumstance. The choice it leaves you is to climb it differently: become the person who directs the AI and supplies the judgment, sooner than anyone expected you to.

Want to know which moves in your field still build real leverage instead of training your own replacement? Message Praxy on WhatsApp. We'll look at your actual situation, the data on where your roles are heading, and the next move that compounds.