Let’s talk about the rung you deleted.
Picture a first-year article clerk at an accounting firm. Not the glossy version. The version where you spend nine months on bank reconciliations, matching line after line against a statement, getting it wrong, having a manager with no patience walk you through what a reconciling item is for the fourth time, and then doing it again on Monday.
Nobody who did that work would call it a career. It was the price of entry. And the thing everyone missed — including the firm — is that the reconciliation was never the product. The reconciliation was the curriculum.
The actual qualification was becoming a person who could smell a fabricated ledger entry from across the room. That doesn’t get taught in a course. It gets absorbed over years, by doing boring work badly, in front of someone who has been doing it for twenty.
That transmission mechanism is now being quietly dismantled, and almost nobody is counting what it costs.
Here’s the data. Stanford’s Digital Economy Lab pulled payroll records covering millions of American workers. For workers aged twenty-two to twenty-five in the most AI-exposed occupations, employment is down around thirteen percent relative to other groups since generative AI landed in late 2022 — while older, more experienced workers in those same occupations stayed flat or grew. In the most exposed roles specifically, entry-level employment fell about six percent between late 2022 and mid-2025, while their seniors grew six to nine percent.
Same industry. Same work. Opposite outcomes, split cleanly by how many years you had already put in.
And here’s the finding buried under the headline, the one that should bother you more. The researchers separated automation from augmentation. Where AI substituted for what juniors did — the routine, codified, repeatable work — junior hiring fell. Where AI augmented the work instead of replacing it, junior hiring held.
That’s not an observation. It’s an instruction. The firms that wired the machine in to do the junior’s job cut the junior. The firms that wired the machine in to sit next to the junior kept hiring them. One of those decisions is cheaper this quarter. The other one is the only one that still produces seniors.
HBR made the plain argument last September: eliminating the bottom rungs of a career ladder doesn’t remove a cost. It removes the mechanism that produced the people at the top — and the bill arrives years later, after the people who made the cut have moved on.
Now the part people are actually missing.
Everyone is framing this as a jobs story. It isn’t. It’s a sovereignty story, and the version that ends with humans holding far less authority than they think they do.
The capability of these machines rises every year. The ability to check them doesn’t come from anywhere except people who have spent real time inside the domain. Judgment is the thing that authorises or vetoes the machine — it’s the human seat at the table. And judgment is exactly what a junior role used to manufacture, not by teaching theory, but by making you live inside the details until the details started telling you things.
You cannot put ten years of that into a system prompt. You can put the manual in. You can put the precedent in. You can put the last five hundred exceptions in and call it context. The machine can be trained on everything you know and still never be your apprentice, because apprenticeship is not information transfer. It’s a relationship with three ingredients the machine doesn’t have: someone with authority over you, time slow enough for the lesson to land, and a mistake that costs you something you actually feel.
Strip those three out and you don’t get a faster junior. You get no junior.
Which brings the second-order effect nobody will say out loud. If your organisation runs for a decade with the machine doing the junior work and no humans coming up through it, you don’t end up with a leaner team. You end up with a team where the only senior judgment available is the model’s — because the humans who could have overruled it were never hired.
That is how machine authority gets handed over. Not by a dramatic decision to delegate. By attrition. Slowly, sensibly, one unfilled graduate position at a time, until there is nobody left in the room who can tell whether the answer is any good.
The machine doesn’t take the wheel. We just stop producing people who can reach for it.
So what do you do about it. Three practical things.
One. If you control a budget, protect one rung. Put a single junior role outside the headcount model and defend it on purpose — not as charity, as your only working supply line for the people who will be checking these machines in 2033.
Two. Put the machine in the teaching seat, not the junior’s seat. Let the apprentice use the model on the boring work, then make them review it, challenge it, and find where it lied. A junior who spends two years catching a confident model being wrong builds judgment faster than any generation before them — provided you let them do that instead of asking them to beat the model on speed, which they will lose.
Three. Let them carry weight early. Real work, their name on it, consequences they feel. That is the whole curriculum. Everything else is slides.
And if you’re the junior and the rung has already been pulled out from under you — build the apprenticeship yourself. Find one person with scars. Do the boring work for them and make them explain their calls, especially the calls that went badly. The machine can teach you what’s true. Only a human can show you how they decided when the truth wasn’t clear yet.
The principle worth defending: keeping the human sovereign is not only about what you’re allowed to refuse. It’s about whether there are still humans around who are capable of deciding at all. Authority you can’t hand down dies with you — and then it flows to whatever is still standing in the room.
So teach somebody. Slowly, in person, with your full attention and no shortcut. It’s the one thing the machine can’t do for you. And it’s the only thing that stops it getting the last word.