When I started in sales, one of the most annoying parts of my job was figuring out how to get something from one country to another.
Every quote meant finding the route, checking the transit time, finding the cost, and painstakingly pulling information from a bunch of different documents.
I kept opening the same files, looking for the same answers, over and over again.
Eventually, I got tired of it and built a spreadsheet that pulled everything together.
It probably cut 80 or 90% of the time out of the process.
The spreadsheet made me faster.
But building the spreadsheet did something even better.
To create it, I had to go through everything. Find the source information. Understand how the pieces connected. Notice the exceptions. After enough repetition, a big chunk of that knowledge was stuck in my head, whether I liked it or not.
The tool saved me time.
The work I had to do to build the tool made me better.
I think about that a lot now.
Because today, a junior salesperson could probably get much of the same information from an AI model in seconds.
That’s obviously useful. But it may also be a problem.
We keep hearing that AI can’t replace human judgment.
I wholeheartedly believe that.
But it leaves out the more interesting question:
Where does human judgment come from?
Where human judgment comes from
Robert Greene has spent years writing about mastery and apprenticeship (in between power, seduction, and human nature).
In one passage I highlighted in his book The Daily Laws, he describes learning as something that happens through repetition and active, hands-on involvement. Elsewhere, he argues that you should treat your whole life as a kind of apprenticeship.
You watch, practice, try something out.
Someone corrects you.
You do it again. And again. And again.
Eventually, something changes.
You stop merely knowing more.
You start seeing more.
Greene was describing the conditions that produce mastery.
Fourteen years after Mastery was published, we’re beginning to remove those very same conditions from work.
The hidden apprenticeship inside junior work
Earlier this year, USC economist Miao Ben Zhang published a working paper called The Broken Job Ladder.
It’s an interesting take that gives me a bit of clarity on something that has been hard to articulate.
Most skilled professions have historically had two stages to their careers.
Education gave you knowledge. Work gave you judgment.
But companies rarely paid separately for the second part.
Doing the job and learning the job tended to happen at the same time.
According to Zhang, production and judgment formation were bundled. The activity that produced useful output for the company also developed the worker’s expertise.
That might seem like a small thing until you realize that we’ve spent decades treating the two outputs as one. The junior work wasn’t just there to get work done but to help a junior person to grow into a senior one.
Junior work always produced two things.
The obvious one was whatever the company needed (it’s your job after all).
The contract got reviewed. The spreadsheet got built. The research got done. The customer got called. The presentation got created (most importantly, of course).
But a second product was being created as all of those monotonous actions were being taken.
The junior was becoming less junior.
They were seeing examples.
Making mistakes.
Watching what their manager changed.
Learning which number mattered and which one looked important but wasn’t.
Discovering the exceptions that didn’t appear in the onboarding manual.
All the while, they were building an internal library of situations, decisions, mistakes, and corrections.
Eventually, we call that library experience.
And when experience becomes useful enough, we call it judgment.
The company pays for the output, and you get to develop the judgment as a byproduct. Win-win.
Until now.
How AI separates output from learning
AI unbundles the two.
If a model can create the first draft, summarize the documents, build the analysis, find the route, or produce the research faster and cheaper, the rational (and dare I say, capitalistic) decision is pretty obvious.
Use the model.
That’s what I do too.
But when we take away the work, we’re also removing the place where learning happens.
I think of this as the apprenticeship severance.
It’s not that we consciously decided to stop developing people. Instead, we’ve separated the output from the process that used to develop them. Once those two things are separated, the economics change. If you still want someone to get the reps in, somebody has to deliberately create them.
A manager has to supervise. The company might even have to tolerate work being done more slowly than AI could do it. As with psychological safety, someone has to create the conditions for people to develop.
Someone has to assign a task that the machine can complete in eleven seconds, because the purpose of the task is no longer just the task itself.
For the first time, judgment starts showing up as a line item.
What the early AI jobs evidence shows
There are early signs that this shift may already be showing up in the labor market.
Stanford’s Digital Economy Lab updated its research in August 2026. It did not find evidence of widespread economy-wide AI job destruction.
But among workers aged 22 to 25 in highly AI-exposed occupations, employment was about 19% below the level you would expect if it had kept pace with less-exposed occupations. The researchers also found that much of the adjustment appeared to be happening through reduced hiring rather than mass firing.
Another preliminary study examining tens of millions of U.S. workers found something similar inside firms adopting generative AI: junior employment fell relative to comparison firms, largely because companies hired fewer juniors.
Then PwC found something fascinating, or disturbing, depending on your perspective. It turns out that in highly AI-exposed entry-level jobs, employers were increasingly asking for abilities that used to appear in more senior roles: skills like judgment, leadership, and decision-making.
In other words, the jobs weren’t only becoming scarcer, but also more demanding.
We are essentially not just taking rungs out of the ladder, but also raising the height of the first rung at the same time.
It’s a warning sign we should pay attention to, because the causal story isn’t settled yet.
Research from Denmark found that early-career employment was falling there too, but workplaces that had actively introduced AI chatbots did not experience a significantly larger decline than other workplaces. There are plenty of other things happening in the economy at the same time, so we should be careful about placing blame only on AI.
We can’t blame AI for every missing junior job. But it’s a problem that we have to take seriously.
The junior job that we get rid of today may be the experienced person who we’re looking for five years from now.
It also brings to light the question that we should be asking ourselves right now:
What happens when the work that used to train people stops being given to people?
Why experts see what beginners miss
Watching my son play football has made this feel less abstract to me.
Nobody believes you become a great footballer because someone explains football to you really well. You don’t go down the street where an ancient Cristiano Ronaldo sits on the corner and gives you such vivid storytelling that you immediately get better.
Instead, you watch. You imitate. You hit the ball thousands of times.
Someone tells you to turn your foot slightly.
You try again.
One degree changes.
Then another.
And eventually, you see things that a beginner doesn’t see.
In his book The Mamba Mentality, Kobe Bryant described how his film study changed as he became more experienced.
At first he watched what happened.
Later, he began watching for what didn’t happen. What should have happened. What alternative might have been available. Which counter could have worked.
That’s a beautiful description of expertise.
Judgment isn’t just knowing the answer. It’s seeing the alternatives.
A beginner sees what is there. An expert also sees what is missing.
I don’t believe that you can shortcut your way to that.
What copywork taught me about learning
Writing has taught me the same lesson.
When I was a kid, I used to write out stories from books I liked by hand.
I didn’t know there was anything special about it.
Years later, when I became a marketer, I discovered copywriters had been doing versions of this forever.
Copywork.
Take good writing and copy it.
Sentence by sentence.
It sounds almost absurdly inefficient now.
Why rewrite something that has already been written? Because doing it by hand forces you to slow down enough to notice what the writer is actually doing.
You have to remember that the finished page isn’t the only output. The writer is another output.
In The Adweek Copywriting Handbook, Joe Sugarman wrote about copywriting as a mental process built from the sum of your experiences and specific knowledge. The more you’ve experienced, the more you have to draw upon when a new problem pops up.
That’s why five days of copywork can change your writing.
When you put that pen to paper, you’re not just collecting sentences. You’re training your ear and your eye. You begin to feel where a sentence drags and where it should stop.
You start to feel why one word feels cleaner than another. Nobody has explained every single rule to you because that action and feeling simply become part of you.
The trade-off I make with AI
Which puts me in a slightly uncomfortable position.
I use AI for research constantly now.
And I love it.
Research used to send me down rabbit holes for hours.
One source led to another. Then another.
I’d find some strange study or historical story that had nothing to do with the question I originally had in mind, and occasionally it would change the entire piece I was writing.
It was inefficient as hell. But it was also educational.
Today I can compress hours of that work into minutes.
And the truth is that I don’t have unlimited hours.
So I make the trade.
But it is still a trade.
The thing I ask AI to find may be exactly what I need.
It may also prevent me from finding the thing I didn’t know I needed.
The same is true of writing.
If I outsource every first draft, I may produce more words.
But what happens to the writer?
None of this means we should stop using AI.
That would be ridiculous. Downright stupid, even.
Efficiency is real.
So is formation.
The mistake is assuming they are the same objective.
For most of our careers, useful work and getting better at it often happened in the same place. One often relied on the other.
AI is forcing us to separate them.
Maybe the apprenticeship isn’t over.
Maybe the automatic apprenticeship is.
And if judgment is no longer created as a byproduct of doing real work, someone has to decide that producing it is worth paying the price for.
Which leaves me with three questions I’m asking myself too:
What work made you good at what you do?
Does that work still exist?
And what are you handing over now that would have taught you something?
FAQs
AI can preserve the output of junior work while reducing the repetition, feedback, and exposure to edge cases that helped people build judgment. If those tasks disappear, organizations may need to create those learning conditions deliberately.
Professional judgment grows through repeated decisions, observation, feedback, mistakes, correction, and exposure to exceptions. Knowledge matters, but experience teaches people what to notice and when the standard answer does not fit.
Early evidence suggests some AI-exposed employers are hiring fewer junior workers while asking remaining entry-level roles for more senior skills such as judgment and decision-making. The causal picture is still developing, so AI should not be treated as the only explanation.
Apprenticeship severance is Brian Tomlinson’s term for what happens when AI separates useful output from the work that used to train the person producing it. The task gets automated, but the learning that once came with the task may disappear too.
Yes. Productivity and skill formation are different objectives. AI can make a task faster while also reducing the practice a person would have gained by doing it. The answer is not to reject AI, but to notice which learning loops still matter.
