The Apprenticeship Problem

AI is eating the work that used to train people. The efficiency gain shows up this quarter. The missing generation of judgment shows up in ten years.

Every technology eats the bottom rung first. Not because junior work is unimportant, but because it is the most structured, the most repeatable, and the easiest to specify. First drafts. Research memos. Basic analysis. Document review. The unglamorous work that fills the first few years of almost every professional career.

That work is exactly what current AI is best at. So firms are doing the obvious thing: automating it. The business case writes itself. Why pay a first-year analyst to build a model an AI can draft in minutes? Why staff three juniors on diligence when one senior with good tools covers it?

As a first-order decision, it is correct. Almost every firm making it is making it rationally.

And that is precisely what makes the second-order effect interesting: a decision that is right for every individual firm can quietly break something the whole industry depends on.

Junior work was never just output. It was tuition.

Here is the thing the automation math leaves out: the deliverable was only half the point.

Expertise does not come from reading about a field. It comes from reps — thousands of small, low-stakes attempts with fast feedback. The first-year associate marking up a contract is producing mediocre output, and everyone knows it. What she is actually doing is calibrating. Learning what a bad clause smells like. Learning which numbers are worth checking twice. Learning, slowly, what good looks like.

The firm was never really buying her output. It was selling her reps and calling it a salary. Junior work was an apprenticeship wearing the costume of a job.

Automate the work and you automate the apprenticeship out of existence with it. Nobody decided to stop training people. It happened as a side effect of a spreadsheet.

Seniors are made of juniors

The problem compounds because of where seniors come from. They are not hired from a separate senior-person factory. They are juniors, plus ten years of accumulated judgment.

So the pipeline works like a fishery. Harvest the mature fish and skip the spawning season, and the catch looks fantastic — for a while. The collapse arrives on a lag, long after the decisions that caused it, disconnected enough from the cause that everyone gets to act surprised.

This is what makes the apprenticeship problem a genuinely second-order effect rather than just a downside. The cost is invisible at the moment of decision. Current seniors are unaffected — if anything, AI makes them more productive and more valuable. The books look great. The org chart looks lean. The damage is entirely concentrated in people who do not work there yet.

We have seen milder versions before. Aviation spent decades automating the routine parts of flying, then discovered that pilots' hand-flying skills had atrophied exactly when rare emergencies demanded them. The industry's answer was not to remove the automation. It was to deliberately manufacture practice — simulators, mandated hand-flying, training designed to preserve a skill the job no longer exercised naturally.

That is the template worth paying attention to: when the work stops teaching, someone has to build the teaching on purpose.

The new bottom rung is reviewing the machine

The optimistic story says juniors will simply move up a level: instead of producing the work, they will supervise AI that produces it. Managing a fleet of models sounds like a promotion.

But there is a catch, and it is the crux of the whole problem: it is very hard to grade work you have never done. Reviewing a contract well depends on the pattern library you built by marking up hundreds of them badly. Catching the subtle error in an AI-drafted model requires the instincts of someone who has built models by hand. Verification is a senior skill. We are about to hand it to the most junior people in the building, pointed at output that is fluent, confident, and wrong in ways designed to be hard to notice.

A junior reviewing AI output without reps is not an apprentice. He is a rubber stamp with a login.

So the second-order effect creates its own labor market response, the way these things usually do. Watch where the new work forms:

The firms that figure out synthetic reps win a quiet advantage. Somebody has to design the equivalent of the flight simulator for analysts, lawyers, and engineers — deliberately manufactured versions of the grunt work AI absorbed, with feedback loops attached. That is a new profession hiding in plain sight.

Training becomes a market instead of a byproduct. When employers stop providing reps for free, people will buy them elsewhere — residencies, studios, bootcamps that sell supervised practice rather than credentials. Expect the return of the guild in modern clothing.

Judgment becomes legible as an asset. Firms that still grow seniors will increasingly be farmed by firms that don't — free-riders who let someone else pay the tuition and then poach the graduates. That works until too many firms try it, which is exactly how commons collapse. The equilibrium probably involves contracts, credentials, or compensation structures that price judgment formation explicitly, because the implicit version stopped working.

The question that actually matters

The first-order question is the one every executive is already asking: can AI do our junior work? The answer is increasingly yes, and pretending otherwise is not a strategy.

The second-order question is the one that will sort winners from losers a decade out: if the market no longer produces judgment as a free byproduct of cheap labor, who pays to produce it on purpose?

Every previous era got apprenticeship for free, bundled into the price of getting work done. AI unbundles it. The work gets cheaper. The judgment gets orphaned.

Firms that notice will build the new rungs deliberately — and will spend a few years looking inefficient while they do it. Firms that don't will look brilliant right up until their bench is empty.

The bottom of the ladder is gone either way. The only open question is who builds its replacement, and whether they start before or after the shortage has a name.