The Apprenticeship Problem

In my last two memos, I argued that judgment is what really matters. When time and materials no longer cost anything, judgment is the only thing left to charge for. When everyone has the same information, judgment is what helps you avoid following the crowd. I still believe both points.

But there’s a problem I haven’t addressed and that is that judgment isn’t something you just find. It’s made. And the work that creates it is exactly what we’re now automating.

Every expert started out as a junior, doing basic tasks under someone’s watch. Think of the first-year associate reading contracts late at night, the analyst redoing a model for the fourth time, or the resident holding a retractor. That work wasn’t really about the end result. It was the price of learning. The old, unspoken deal was that beginners traded their time and effort for a chance to learn from experts, and after thousands of repetitions, judgment would develop.

Now, machines do repetitive work while people just observe. But you don’t get fit by watching someone else exercise.

The Bottom Rung is Already Gone

This isn’t just a guess, it's actually backed by payroll data. Stanford’s Digital Economy Lab (Brynjolfsson, Chandar, Chen) analyzed millions of ADP records and found that since generative AI became common, employment for workers aged 22 to 25 in the most AI-affected jobs dropped by about 13 percent. Jobs like software, customer service, and accounting. At the same time, older workers in those roles stayed steady or even increased. Same jobs, same companies, but different levels.

The most important point is that these declines happen where AI replaces work, not just helps with it. In other words, the jobs disappearing are the ones that used to be training grounds. No one set out to end apprenticeships. Each hiring decision made sense on its own, why pay a junior a high salary to do something a model can do better in seconds? This logic makes sense for one hire, but over years and across an industry, it leaves the profession without a way to train new experts.

The Trainee Became Optional

We’ve seen this happen before, but on a smaller scale and with more detailed data. Matt Beane at UC Santa Barbara spent over two years in operating rooms, studying how surgical residents learn robotic surgery. In traditional surgery, trainees were essential because the procedure needed four hands, so residents learned by doing. With robots, one expert could do the job alone from a console. Suddenly, the resident was no longer needed and mostly just watched.

Beane found that the residents who became skilled did so by going outside the official rules. He calls this shadow learning. These residents spent hours on simulators, while the program required only about four hours a year, the successful ones logged around three hundred. They specialized early and looked for chances to practice with little supervision. The official training mostly created observers. Real skill only survived when individuals found their own ways to practice.

This is what’s happening in knowledge work today. The console is now a chat window, and the resident is every 24-year-old employee in your organization.

But... What About Chess?

There’s a strong counterargument worth considering. Chess faced superhuman AI thirty years ago, but chess training didn’t disappear. Now, kids train against computer engines from the start, and the record for youngest grandmaster keeps dropping. Sergey Karjakin’s record of 12 years and 7 months lasted 19 years until Abhimanyu Mishra broke it in 2021 at 12 years and 4 months, and more young players are following. The machine didn’t end the training pipeline at all. In reality, it became the best coach ever.

So the real question isn’t whether AI destroys apprenticeship. It can actually speed it up. The real issue is figuring out when it helps and when it hurts.I think chess works as a training model because practice is free and feedback is immediate. You can lose thousands of games without any real cost, and the engine shows you your mistakes every time. Apprenticeship survives automation when practice is cheap and mistakes don’t matter. It fails when practice is the actual paid work. No company lets a junior handle thousands of client projects without consequences. Firms stop paying for practice because it’s an expense, and now the model makes that expense optional. Chess separated practice from real performance a long time ago. Professional services never did, and that’s the main weakness.

The Truth

Cutting junior roles looks like efficiency in the short term but the real cost doesn’t show up on any balance sheet, yet. It’s a hidden risk that the market ignores.

Meanwhile, the senior generation benefits. Their judgment, built through years of experience paid for by others, is becoming more valuable. When something is scarce, its price goes up. The same people saying juniors aren’t worth hiring are the ones whose value depends on judgment staying rare. I’m not saying it’s a conspiracy. I’m saying no one has a reason to change things, which is even more troubling.

Quick Additional Thoughts

First, the whole chess proxy idea might generalize further than I think. AI tutors, simulation, synthetic reps, etc. Beane himself argues the technology that broke apprenticeship can be redesigned and repurposed to rebuild it. If someone makes deliberate practice as cheap in law or strategy as it is in chess, the pipeline reroutes instead of dying. Second, perhaps the judgment muscle forms differently than we assume? Perhaps reviewing a thousand AI outputs builds it as well as producing a hundred by hand. After all it's kind of how the chess theory works. Works in poker, blackjack, etc. But I doubt it. I hold the doubt at less than certainty.

The Bet

Judgment is still the main product. But the system that creates it is broken, and the market hasn’t noticed because today’s experts took decades to train and will take decades to retire.

I bet that companies that keep paying for ‘unprofitable’ junior work aren’t just being sentimental. They’re making sure they have a steady supply of judgment, while competitors give it up for short-term gains. In ten years, the rare asset won’t be the AI model because everyone will have that. The rare asset will be the 32-year-old who managed to get real experience.

If your company got rid of entry-level roles this year, you didn’t just save money, you gave up part of your future.