Entry-level work used to run on a bargain nobody wrote down.
The organization accepted that you didn’t know very much yet. You accepted the photocopying, the modest authority, and the specific humiliation of being the person in the room who has to ask what the acronym means. In return, you got proximity — a seat close enough to watch how the stronger people actually worked, room to attempt things under supervision, and enough rope to make decisions serious enough to teach you something without being serious enough to sink the business.
Nobody designed this well. Plenty of junior roles confused development with drudgery, or with finding someone cheap enough to do the work nobody senior wanted to touch. But somewhere inside that badly-run system, people quietly accumulated experience.
AI is rewriting the bargain.
My Substack - Some Assembly Required - has always been where I work things out in public. It’s where the ideas are longest, sharpest, and least edited for palatability.
If you want to stay connected to where the thinking is going, not just the occasional post that survives the algorithm, please subscribe.
Routine work is the first thing automation reaches for. Research, synthesis, drafting, analysis, production - all of it now gets done faster, with fewer people, at a standard that often looks remarkably accomplished. Meanwhile, the junior roles that survive have started asking for distinctly non-junior things: judgment, stakeholder management, strategic thinking, the ability to operate in ambiguity without an adult in the room.
Entry-level talent is expected to arrive already experienced. AI-augmentation looks, at first glance, like the answer.
Hand a junior employee a capable model, and the work that comes back looks senior. The analysis is polished, the options are comprehensive, the language is confident, and it all arrives in a fraction of the time it used to take. The visible gap between junior and senior output is closing fast.
Is the person getting better or is the work just looking better?
Performing like an expert
There’s no single, tidy experience pipeline getting quietly automated away — the truth is messier than that headline. Junior hiring is genuinely down in a handful of AI-exposed and tech occupations, but a hiring chart alone doesn’t prove AI did it. Restructuring, outsourcing and a post-pandemic correction are all tangled up in the same numbers, and untangling them cleanly is harder than any op-ed wants to admit.
Nor does AI simply switch off the opportunity to learn. One workplace study of customer-support agents found the least experienced workers made the largest performance gains from generative AI, the tool handed them the language and instincts of their strongest colleagues, on day one instead of year three.
That’s not nothing.
Expertise has always been rationed unevenly, by which office you landed in, who happened to sit next to you, whether a manager had twenty spare minutes to explain why the plan wouldn’t work. AI can hand strong practice to people who’d otherwise have waited years to bump into it by accident.
It can compress a performance gap. Whether it compresses an experience gap is a different question entirely, and a much harder one.
The history of consequences
Experience was never just a bigger pile of information. AI already holds more information than any one person could accumulate in a career. It can spot patterns, surface precedents, and describe with total confidence what an experienced practitioner would probably do.
What it can’t hand over is the history of consequences behind that confidence.
Experience is watching a sensible plan produce a ridiculous outcome. It’s discovering that the exact same approach which worked brilliantly with one client blows up spectacularly with the next. It’s the itch of recognizing a familiar pattern forming — and the sharper, more useful itch of recognizing when a situation only looks familiar enough to be dangerous.
Mostly, it’s being wrong when being wrong actually costs something.
AI can tell a junior colleague what an expert would do here. It cannot tell them why the expert pauses for three seconds before doing it because that hesitation isn’t a data point. It’s a scar.
The model can lend someone the patterns of experience. It cannot automatically lend them the bruises that produced the patterns in the first place.
That distinction is easy to miss, because output is the thing you can put in a dashboard. The organization can count hours saved, tickets closed, documents produced, quality scores trending the right way.
Judgment doesn’t show up on a dashboard. It shows up months later, in the one meeting where the standard workflow stops working and somebody has to notice before the client does.
Can they tell when the recommendation is wrong for this particular situation? Can they carry what they’ve learned into a mess the model has never seen? Can they explain their reasoning instead of repeating the answer the machine gave them? Can they put their name on the decision when the model was confidently, articulately wrong?
Those are the questions that matter. They’re also the ones nobody asks until the quarter it’s already too late to ask them.
The experience trap
Here’s the trap.
AI lets junior staff produce senior-looking work. The organization reads better output as better independent judgment - a reasonable mistake, and a costly one. Expectations climb. Supervision quietly recedes. Entry-level roles start clustering around judgment and coordination, on the assumption that the production in the middle is already handled.
Which means junior employees get fewer reps at making decisions that matter, under the eye of someone who’s already made every version of that mistake.
Eventually the organization looks up and discovers it has a warehouse full of senior-looking output and a genuine shortage of people who can independently judge whether any of it is any good. The productivity numbers look terrific right up until the day something unusual happens, and there’s nobody in the building who’s seen it before.
I’ve argued before that the real story isn’t an aging workforce, it’s an experience shortage. Run the AI version of that argument forward, and the maths gets uncomfortable: as production gets cheap, contextual judgment and the willingness to be accountable for a call get more valuable, not less. Which raises the obvious, unglamorous question nobody wants on the agenda:
Where does the next generation of experienced people actually come from?
Companies may be nudging existing experience out the door while automating the exact work through which new experience used to get built. AI-assisted output is good enough to hide that trade for a while.
Not forever.
The old apprenticeship was not sacred
None of this is a case for preserving junior drudgery.
The old system was never a finishing school. A lot of entry-level work was repetitive because nobody had bothered to design it properly. Mentorship access was a lottery. Plenty of people spent years doing low-value tasks and came out the other side with tenure, not judgment, which are not the same thing, whatever the org chart implies.
Younger workers shouldn’t have to wait patiently for permission to do work that matters, either. They often bring sharper AI fluency, a different cultural register, and less loyalty to habits that only look like wisdom because nobody’s challenged them recently.
Done well, AI could be a better apprenticeship than the one it’s replacing. It can put strong practice in front of a novice sooner. Give instant feedback. Simulate the difficult conversation before the real one happens. Let someone compare two decisions before either has consequences. Make tacit knowledge visible instead of leaving it to be absorbed by osmosis, assuming you’re lucky enough to sit near the right person in the first place.
But none of that happens just because everyone gets a license and the usage dashboard goes up and to the right.
It has to be designed for building capability, not just accelerating output. Those are different design briefs, and most organizations are currently only writing one of them.
Substitution or scaffolding?
The useful distinction is between a substitution and scaffolding.
A substitute performs the task for you. Take it away, and the performance goes with it. Scaffolding helps you build a capability you can increasingly exercise for yourself. It changes shape as you get stronger: support is withdrawn, redirected, or made more demanding.
Most organizations are currently running AI as a substitution while describing it, in every town hall, as scaffolding.
The model keeps producing the analysis, suggesting the language, recommending the next move. The employee gets remarkably good at operating the system.
Is their judgment getting stronger? Or is their dependence just getting more productive?
A real AI apprenticeship would ask more of both sides. It might make the employee commit to a judgment before the model reveals its own. It might leave the uncertainty visible instead of smoothing it into false confidence. It might explain, specifically, why the familiar pattern doesn’t apply this time. It might compare reasoning, flag what’s missing, and slowly walk someone into situations no template covers.
Above everything, it would need to keep leaving room for people to make real decisions, live with the consequences, and answer for what happens next.
The goal was never to remove the machine.
It’s making sure the person working alongside it is getting more capable, not just more productive.
The question organizations should ask
The easy question is: what can this person produce with AI?
The harder, more consequential ones: What are they learning through the act of producing it? What can they see that the model can’t? What can they carry into a situation nobody prepared them for? What are they still willing to question? And when the answer is wrong whose name is on it?
AI may let junior workers perform like experts sooner than any generation before them.
That could be one of the more remarkable expansions of human capability in a working lifetime.
Or it could produce a workforce expected to arrive experienced, carrying all the outputs of expertise, and very little of the practice, consequence, and responsibility that used to be how judgment actually became theirs.
The model won’t decide which one happens. Whether organizations use AI only to speed the work up, or also rebuild how people acquire the judgment to do it — that will.
