One Must-Read Article

The Apprenticeship Problem

If AI does the work that used to build judgment, where will the next generation of judgment come from?

I have spent the last several months making a single argument: when execution becomes cheap, judgment becomes the premium. Machines now do the analysis, the first drafts, the routing, and the reconciliation. What remains important is the human ability to know what is worth doing and to recognize when something is off. I still believe that. But there is a problem buried inside this thesis, and it is time I named it.

Judgment is not issued at hire. It is earned. And it has always been earned the same way — by doing the unglamorous work.

How judgment was always built

Think about how a great staff member actually develops. The junior associate built the model by hand, got the assumption wrong, watched the number break, and learned why that assumption mattered. The associate wrote the bad first draft, had it torn apart, and slowly developed an ear for what good looked like. The new manager sat through the tedious reconciliation and, somewhere in the tedium, observed the pattern that nobody senior had time to see.

None of that work was valuable on its own. The model got rebuilt. The draft got rewritten. The reconciliation got filed and forgotten. The output was disposable. The staff was not. That grunt work was never really about the output — it was the apprenticeship. It was how a person accumulated the thousand small encounters with consequences that eventually compound into the thing we call judgment.

What we are quietly removing

Now look at what AI is taking off the table. It is not taking the senior judgment work — at least not yet. It is taking exactly the entry-level staff who used to build that judgment. The first-pass analysis. The starter draft. The grunt reconciliation. The very tasks we are most willing to automate are the same ones that used to season people.

And the logic is airtight in the moment. Why pay a junior person to spend three days on a model that an agent produces in three minutes? Why staff the tedious first draft when the machine writes a competent one instantly? Every individual decision to automate the bottom rung is correct on its own terms. That is precisely what makes this dangerous. You do not feel that you are doing anything wrong. You feel efficient.

But run it forward five years. The work that used to turn a green hire into an experienced staff member is gone — absorbed by AI agent. The junior people you do keep never get the reps, because the reps now go to the agent. And the senior people whose judgment you are leaning on harder than ever? They are not getting any younger, and there is no one coming up behind them who was forged the way they were.

Here’s my take

This is the part leaders often miss, because it never shows up on this quarter’s P&L. Automating the bottom rung looks like pure margin. The cost is deferred and invisible: it lands years later, as a thinning bench of people who can actually exercise the judgment your whole strategy now depends on.

You are not only buying efficiency when you automate entry-level work. You may be quietly spending down a reserve you did not know you were holding — the reserve of future judgment that your grunt work used to manufacture as a byproduct. And reserves that compound on the way up compound on the way down, too.

I do not have a clean answer to offer this week, and I am suspicious of anyone who does. The companies that figure this out will not be the ones who refuse to automate — that ship has sailed, and it should have. They will be the ones to notice that they have removed the old apprenticeship and deliberately decide to build a new one. What that looks like is the question we will be sitting with.

For now, one thing is worth holding in view. The cheapest work in your company was never merely cheap work. It was where your next generation came from. Before you automate the rest of it away, it is worth asking what you plan to put in its place.