The standard defense of AI at work is that it handles the boring parts so people can focus on the interesting parts. It sounds unarguable.
It is also a fair description of dismantling the only training system most companies have.
The boring parts were the curriculum. Nobody learned to read a deal by being told how. They learned by building the list, sitting through the calls that went nowhere, writing the account summary a manager tore apart, and being wrong early and often enough that the pattern eventually landed. The grunt work wasn't the toll you paid to reach the real job. It was the teaching mechanism, and it worked because it was tedious and nothing much rode on it. You got to be bad at something that didn't matter yet.
Now a machine does it in nine seconds, and does it better than the twenty three year old would have.
The labor data is pointing at the mechanism, not the headline
Most people read the entry level employment numbers as a story about jobs. The more interesting part is what they say about knowledge.
The Stanford Digital Economy Lab's revised Canaries in the Coal Mine?, released August 12, found that employment for workers aged 22 to 25 in the occupations most exposed to AI now sits about 19% below their less exposed peers, widening from 15% a year earlier. The researchers are careful about what they are not seeing: no widespread displacement across the economy. The adjustment runs almost entirely through reduced hiring rather than firing.
But look at where the line falls. Employment declined in roles that rely on codified knowledge, the formal, documented, teachable kind. It rose for experienced workers in roles that lean on tacit knowledge, the sort acquired through practice and mentorship.
That is not a story about young people being unlucky. It's a story about which kind of knowing still commands a price. And it should worry the people on the winning side of that line more than it currently does, because the two categories are not independent. Tacit knowledge is what you get when someone does codified work for several years under supervision. Cut the intake and you have not made the senior bench more valuable forever. You have made it the last of its kind.
Tacit knowledge has never had a vendor
Any company can now buy competent output. Research, drafting, synthesis, a competent first pass at the analysis. The things that used to separate a good analyst from a mediocre one are now available to everyone at roughly the cost of electricity. That was the moat for a lot of professional work, and it's gone.
What is left is unbuyable. Nobody sells the instinct that a champion went quiet for a political reason rather than a scheduling one. Nobody sells the read that the room agreed too easily. That knowledge has exactly one production method: exposure, repetition, and someone experienced telling you that you got it wrong and why.
So the scarce resource inside a company is no longer people who can produce plausible work. That's abundant and free.
The scarce resource is people who can tell when plausible work is wrong.
The confidence trap
Here's the part that makes this compounding rather than merely sad.
Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real AI use cases at CHI 2025 and found a clean relationship: the more confidence someone had in the AI, the less critical thinking they reported doing. The more confidence they had in themselves, the more they did. The work shifted from producing an answer to supervising one.
Supervising an answer is a much smaller cognitive act than generating one, and it depends entirely on having a model of the domain to check against. Once you have read a fluent answer, you can't produce an independent one anymore. You can only agree or disagree with the one you were handed.
Now consider who is doing the most supervising and the least generating. It's the people who have no prior model to check against, so they can only evaluate the one signal these systems never fail to produce, which is fluency. We have built a system that is most persuasive to the people least equipped to catch it, and then routed it straight at them.
Buy the friction back
Three things I'd put my name to.
One: the strongest teams will start deliberately reintroducing friction. Commit first, reveal second: the person writes their own read of the account, the risk, the number, before the system shows its version. It costs ten minutes, it's unpopular, and it's the only reliable way to keep someone's judgment loaded rather than idling. Friction was never the enemy. Pointless friction was.
Two: "years of experience" is about to become a worthless proxy. What matters is how many times a person committed to a call and then found out whether they were right. Call them reps of judgment. AI inflates the first number and quietly deflates the second, so the two are now measuring opposite things.
Three: someone is going to reverse a junior hiring freeze for supply reasons, publicly, within three years. Not out of generosity. Because they'll do the math on their senior bench, count the retirements, look behind them, and find nobody there.
None of this is an argument against using the tools. The productivity is real and it isn't going back in the box. It's an argument against a specific piece of sloppiness: treating the work as pure cost, when the work was also the mechanism that turned people into the ones you trust.
Every hour you automate is an hour nobody learns from. That's a fine trade, right up until the moment you need somebody who learned.