Here are two numbers that should not be able to coexist, and yet here we are.

US unemployment: sitting around 4%. Professional and business services — the white-collar jobs everyone said AI would gut first — actually adding jobs this year. By the aggregate numbers, the AI jobs apocalypse simply failed to show up.

And then the second number. Payroll research covering millions of workers finds that employment for 22-to-25-year-olds in the jobs most exposed to AI is shrinking by roughly 4% a year. Entry-level postings are down double digits. In the UK, graduate job postings have fallen by more than half since 2022 — and yes, some of that is the economy and offshoring and post-COVID correction, which is why the cleaner signal is the gap between AI-exposed and non-exposed roles for the same age group. That gap is real, it's growing, and it points one direction.

So the machine didn't eat the jobs. It ate the entrance.

If you've been reading the last three issues, you already know why the damage hides from the headlines. Layoff trackers count announcements, and there is no announcement for a graduate scheme that quietly doesn't run this year. But I want to spend this issue on why this particular version of the problem is the one that should actually worry you — even if you don't know a single 23-year-old — because the bill lands on all of us, on a delay, around 2032.

The bundle nobody knew they were buying

Here's the thing about entry-level work that was true for a century and that we all stopped noticing: it was never just work. It was a bundle.

The firm bought output — the research memo, the first-draft contract, the tier-1 ticket resolution, the QA pass. And riding invisibly along with that output came training. The junior lawyer reading a hundred contracts to write one clause was becoming a lawyer. The support rep taking two thousand calls was building the pattern library that one day makes someone a great operations director. The firm thought it was buying cheap labor. It was actually running an apprenticeship and getting the tuition paid in output.

Generative AI is best at exactly the codifiable, repeatable tasks that made up the output half of that bundle. Which is a polite way of saying: the cover charge for the apprenticeship just went to zero.

And so the bundle unravels. No board voted to stop training the next generation. No CEO announced it. Every firm just, individually and rationally, declined to keep paying for output it no longer needed — and the training that had been smuggled inside that output stopped being funded as a side effect. A hundred thousand reasonable decisions, one unreasonable outcome. Economists have a name for this shape of problem, but you don't need the name. You need the date it comes due.

We've run this experiment before

Because we have, three times, and the results are in.

Zuckerberg's essay — the one I took apart last issue — contains an accidental confession on this exact point. He notes, correctly, that there's a shortage of skilled tradespeople: carpenters, electricians, construction workers, so scarce that Meta is building its own academy to train them for data center construction. He presents this as good news for blue-collar work. But ask why there's a shortage. Because a generation ago, the trades apprenticeship pipeline collapsed — firms stopped taking on apprentices, everyone was told to go to university — and twenty years later, the people who should have spent those decades becoming master electricians simply don't exist. What Zuckerberg describes as an opportunity is the invoice from a previous pipeline failure, arriving two decades late. And his essay proposes no reason the same invoice won't arrive for knowledge work.

Second run: law firms after 2008. They cut first-year associate classes to the bone — sensible, times were hard, the doc review could be outsourced anyway. By around 2014 the same firms were in a bidding war for mid-level associates who barely existed, because the class of people who should have been fifth-years had never been hired as first-years. Seniors, it turns out, are not hired. They are manufactured, and the production run takes five to eight years, and you cannot rush a batch when you discover you're short.

Third run, and the human cost: the research on people who graduated into the 2008 recession shows earnings damage that persisted for a decade or more. Scarring, the economists call it. A cohort doesn't just wait out a bad entry market and resume normal service. It carries the dent.

Put those three together and the 2032 forecast writes itself. Senior scarcity. Senior compensation inflating exactly when companies expected AI to be cutting labor costs. People promoted on half the reps their predecessors got — "experience compression," which sounds efficient until one of them is running your incident response. And a scarred cohort of people who did everything right and arrived to find the door removed.

The 2032 senior market is being priced right now, by 2026 hiring decisions, mostly by people who think they're just closing a req.

The three candidate fixes, honestly odds-checked

Whenever I lay this out, I get three rebuttals, and each is a real candidate for the fix, so let's take them straight.

"Firms will just rebuild apprenticeships deliberately." Some will. IBM is tripling entry-level hiring for AI roles even while cutting elsewhere — that's a firm consciously deciding to pay for training as training, with no output cover story. But notice what it requires: paying for something whose return arrives in five years and might walk out the door to a competitor in three. Most firms won't do that until senior scarcity is already biting their own P&L, which is precisely too late. This fixes the problem for the firms that move early and leaves the shortage for everyone else.

"AI itself becomes the tutor." This is Zuckerberg's answer — everyone gets a personal teacher with a PhD in everything and unlimited patience. And it's not nothing; the explaining half of learning genuinely is getting better and cheaper by the month. But here's what thirty years of watching people become good has taught me: judgment doesn't come from explanations. It comes from reps taken under consequence — from being the person whose name is on the thing when it's wrong. A tool that will happily do the task for you is a strange teacher, because it removes the productive struggle that was the whole mechanism. Flight simulators work because they make you fly. Most AI tools, as actually used, fly the plane while you watch. Deliberately designed practice environments could close that gap. Almost nobody is building them, because "tool that makes the work harder on purpose" is a rough pitch to a buyer.

"Juniors will arrive pre-trained — they'll build portfolios with AI before anyone hires them." Partly true, and the most hopeful of the three. The 22-year-old who has shipped real things with AI leverage is genuinely more capable on day one than I was at that age, and it isn't close. But a portfolio proves you can build. It doesn't give you the thing the old bundle secretly provided: reps inside a real operation, with real stakes, real customers, and someone senior watching. That part still has to happen somewhere, and "somewhere" is exactly what's being defunded.

Probably we get some of all three, unevenly, with a gap in the middle. The gap is the story.

What I'm doing about it, since I'm implicated

I'll end where this newsletter always tries to end — on the floor, with a declared interest. My company automates entry-level-shaped work. I am part of the mechanism this essay describes. So what does someone in my seat actually do?

Here's mine. The best apprenticeship in my company right now wasn't designed as one. It's the person who reviews AI agent transcripts — reads the calls our voice agents handled, catches the edge cases, decides what escalates, notices the pattern before it becomes a problem. Six months of that and you understand customer operations better than most people learn in three years of taking the calls yourself, because you're seeing a thousand conversations a week instead of forty, with the judgment work left in and the repetition taken out.

Which suggests the honest reframe for anyone who employs anyone: the output cover story is gone, so stop pretending. A junior hire in 2026 is a capital investment in your 2032 senior bench, and it should be budgeted, designed, and defended as one. Design the reps on purpose — the escalations, the reviews, the small things with their name on them — because the reps no longer happen by accident.

Four issues ago I started with a question about which room the AI jobs debate actually gets decided in. Here's the answer this series has been circling the whole time: it's decided in the meeting where you choose what to do with the freed hour, in the org review where you decide which measurers were secretly coaches, and in the hiring decision where you choose whether your firm still manufactures seniors or just consumes them.

So, last challenge of the run. Count the people in your company who are getting reps right now that will make them formidable in five years. If the number is zero, you haven't automated your junior work. You've automated your future.

Tell me the number. I mean it.

Sandeep Bansal is the founder of GoZupees, which builds the AI agents that automate entry-level-shaped work — and is trying to run its own apprenticeship anyway.

Sources worth your time:

  • Payroll research on 22–25-year-old employment in AI-exposed occupations, as reported by Yahoo Finance, July 2026 (verify primary study attribution before publish)
  • Handshake data on entry-level postings; UK/US graduate posting declines (via AIMultiple roundup)
  • IBM entry-level hiring plans (via Bloomberg / TechCrunch)
  • Research on labor-market scarring for recession-era graduates
  • Mark Zuckerberg, "The Future is for Everyone" (Aug 10, 2026) — the tradespeople passage