What If the Work AI Removes Was Teaching Us Something?

Every conversation about AI and careers eventually arrives at the same reassuring formula: AI takes over execution, people move up to judgment, strategy, and leadership.

It's a comforting story, and often a true one.

But it skips a question worth asking before anyone repeats it again.

Where did the judgment come from in the people who already have it?

For most of them, the honest answer involves the work now being automated. A salesperson learned what actually moves a buyer partly by writing dozens of mediocre proposals herself, and seeing what worked and what didn't land. A junior analyst learned what actually mattered in a set of numbers by building the model from scratch, making the wrong assumption, and watching it break. A young lawyer learned to spot the clause that matters by reading a hundred contracts that mostly didn't.

None of that work was the point. The point was the client meeting, the recommendation, the closed deal. But the work was also, quietly, where the judgment got built.

The formula has a gap in it

The prediction and the expectation don't quite fit together. Organizations are moving toward fewer traditional junior roles, with AI absorbing more of the research, drafting, and first-pass analysis that used to be somebody's entry-level job. At the same time, those same organizations expect the next generation of employees to develop strategic thinking, leadership capability, and judgment faster than ever.¹

Both things can be true and still leave a real problem sitting between them. If AI increasingly performs the work through which people historically developed judgment, something has to explain where the next generation's judgment is supposed to come from. "They'll develop it some other way" isn't an answer. It's a hope, standing in for one.

This isn't an argument that AI makes people less capable. It's a narrower and more useful question: some of the work getting removed wasn't only producing output. It may also have been producing expertise. Those two things happened to live in the same task, and automating the task doesn't obviously preserve both of what it was doing.

Not all removed work is equal

Here's where it gets genuinely difficult, because the honest answer is: some of it clearly was just overhead, and some of it clearly wasn't, and they don't come labeled.

A lot of manual work really was just friction — slow, repetitive, teaching nothing beyond patience. Nobody becomes a better strategist by manually formatting a report for the fortieth time. Removing that is a clean win, no tension, no loss.

But some manual work was where a specific judgment got built, one repetition at a time, and it looked identical from the outside to the work that taught nothing. The analyst building the same kind of model over and over wasn't just producing a document. She was building a felt sense for which numbers behave oddly, which assumptions are load-bearing, which outputs deserve a second look before anyone trusts them. That sense doesn't arrive from reading someone else's finished model. It arrives from building enough of your own that you've been wrong in enough specific ways to recognize the pattern next time.

The point was the client meeting, the recommendation, the closed deal. But the work was also, quietly, where the judgment got built.

Production is not practice

This is the distinction that actually resolves the tension, and it isn't "keep the inefficient process out of nostalgia for how senior people learned it."

Pilots don't learn to handle an engine failure by waiting to experience a real one. The stakes are too high, and the experience too rare. So aviation created something deliberately: the flight simulator. Production and practice became separate activities.

Most organizations never had a flight simulator for judgment. They had something that worked by accident: junior people doing real, lower-stakes versions of the work, learning from real but survivable mistakes, on the way to eventually doing the work that mattered. The manual proposal, the from-scratch model, the hundred ordinary contracts — none of it was designed as training. It was just what the entry-level job happened to consist of, and the training came along for free as a side effect.

AI is very good at removing exactly that kind of work, because from a pure output standpoint, it's genuinely inefficient. The problem is that removing the inefficient production task doesn't automatically replace the training function that came with it. If nobody identifies that function deliberately, both can disappear together.

The question that actually helps

The answer isn't to preserve slow, manual work as a kind of character-building tax on junior employees. That's the nostalgic version of this argument, and it's both unconvincing and a poor way to run an organization — inefficiency doesn't become virtuous just because someone learned something from it once.

Before automating a task, ask two questions instead. What output does this work produce? And what capability does doing the work develop? If AI can produce the first, don't assume you've preserved the second.

If this used to be how people built the judgment we now expect from them, what's the flight simulator for it? Sometimes there's a good answer: structured case reviews, deliberately harder practice problems, mentorship built for the purpose rather than assumed as an accident of the workflow. Sometimes there isn't yet, and that's worth knowing before the training ground disappears quietly enough that nobody notices until the next generation reaches for judgment they were never actually given a chance to build.

AI can remove the work. It can't automatically replace what the work was teaching.