Cutting juniors is the bet that's hard to undo.
When AI shows up in the budget math, the junior role is the first one that looks expendable. The agent writes the test, tweaks the screen, and documents the endpoint, and paying someone to learn by doing those things starts to look like a luxury.
Cutting that role saves little today and leaves a concentrated risk for later. AI takes the small tasks away from juniors, but the reason to hire them still stands: to train, in-house, the people who will carry the company a few years from now.
In the United States, the cut already shows up in the data. Hosseini and Lichtinger, of Harvard, analyzed résumés from 65 million workers at more than 280,000 companies in a study still in preliminary form. At companies that adopted generative AI, junior employment fell about 9% over six quarters compared with those that didn't, while senior employment showed no break in its trend. The drop came mainly from hiring fewer people, which means the cut happens quietly, in the role that no longer gets opened. The study measures an association, and it looks at a market other than mine, but that's already enough to show where the decision gets made.
The veterans direct and the new generation does the work
On November 5, Nintendo released a remake of The Legend of Zelda: Ocarina of Time, a 1998 game, for the Switch 2. In an interview published by Nintendo itself three days earlier, Eiji Aonuma, the series producer, said that part of the reason for wanting the remake was the team: "Many of the team members who work alongside me on Legend of Zelda games today have only experienced developing the more recent titles." Rebuilding the original in full, while learning its structure and the way it was developed, "may trigger new insights that they can carry forward into future development."
It's the same team that was already making the recent Zelda games, and by Aonuma's own account, training it wasn't the only reason for the project. What interests me about it is the arrangement.
The producers let the team decide what to change. "I basically let the developers make whatever changes they genuinely felt were necessary. I wanted the team to make The Legend of Zelda: Ocarina of Time their own," Aonuma says, then adds: "That said, I did deliver my fair share of lectures along the way." In the second part of the interview, Aya Shida, the remake's art director, recalls playing the Nintendo 64 version for the first time after joining the project.
The veterans set the direction from a distance, and the new generation remakes the game with the freedom to decide and learns the foundations along the way. I think about training at a tech company the same way, and cutting juniors leaves the people who give direction in place. Out go the ones who would have learned by doing.
Luis Garicano and Luis Rayo worked out the economics of this arrangement in a theoretical model of training in the age of AI: in a traditional apprenticeship, juniors pay for their own training by doing the menial work. Garicano writes that if AI eliminates the junior work that used to subsidize learning, it will devalue the currency workers have traditionally used to pay for their own training. With the currency devalued, someone still has to pay the bill.
The junior trades small tasks for the business
My bet is that the junior's work is going to shift. The junior moves out of the small tasks the agent handles in minutes and deep into the company's business, close to the customer, with seniors setting the product's direction, as they always have.
Take a junior who joins a product team today. The agent writes the tests and simple screens that would have filled their first months, so they join customer conversations and write down what keeps coming up in the complaints. They also shadow a senior through a tough negotiation over a deadline, first just listening, then leading part of the conversation. When they review what the agent produced for a screen that customer will use, they need to know what the customer expected from it, because without that there's no way to review it well. Based on what they've heard, they propose small experience improvements around things like a confusing line of copy or an extra step in a flow, and the senior decides what goes in, when it goes in, and why.
It's the Nintendo arrangement at team scale: the new generation makes the changes it feels are necessary, and someone who sees the whole picture delivers their fair share of lectures. Every customer conversation teaches the junior the business and brings the team the perspective of the people who use the product. By my math, the junior keeps paying for their training with work, in a currency that's worth more to the company than the old one.
People who arrive fully trained leave too
Aonuma wanted the team to make the game their own, and I know firsthand what that does. On the project teams I led, I trained the people who joined, both on the product and in project management techniques. The people who went through that training felt like part of what was being built and delivered a lot because of it. On those teams, voluntary turnover was always low.
That's why I'm skeptical of the savings the cut promises. My sense is that a junior costs the company far less than the senior who ends up absorbing the junior's tasks on top of their own, and that the international competition for top professionals is intense. A team made up only of seniors concentrates the risk in a few people, and they can leave the same way they came in.
Gary Becker gave the opposing argument its strongest form in a 1962 paper, and Garicano sums it up this way: if companies have to pay for training, they will invest less than they should in general training, because rivals can hire away the trained professional. Becker writes that income-maximizing firms in competitive labor markets "would not pay the cost of general training." Applied to today's junior role, that logic says cutting is rational for anyone deciding alone, since training the junior ends up financing the competitor.
That risk has always existed, with or without AI, and it applies to anyone at the company. A senior who arrives fully trained leaves too, and when they leave, they take with them years spent building knowledge and internal relationships. In both cases, the answer I know is internal advancement and recognition, which is the practical way to take care of people.
Becker's own paper supplies material for that answer. It calls training specific when it "increases productivity more in firms providing it," and my reading is that much of what I described for the junior fits that definition: the company's business, its customers, and the network of people who make decisions inside it are worth much more inside that company than outside it. Negotiating works anywhere, and the junior takes that part along if they leave. With specific training, in Becker's model, the company keeps part of the return, shares the rest with the employee, and has a way to lower the chance that the person leaves.
Becker also writes that companies can reduce the risk of losing the people they trained by paying them, after training, higher wages than those people could get elsewhere. "In effect, they would offer employees some of the return from training." That's what I call internal advancement and recognition.
Read all the way through, the Becker argument that serves to justify cutting juniors also shows how to keep them.
The shortage shows up when it's already too late
Careful critics of the cut admit it may work out. James Stanier, who writes the newsletter The Engineering Manager, describes the logic of those who cut: AI will improve faster than the company's talent base depreciates, and it's possible to cut junior hiring today and still have the senior engineers needed when they matter. "That may be true," Stanier writes. Hosseini and Lichtinger also read the cut as a bet: the decline may reflect companies adjusting in anticipation, before automation actually happens. If companies overestimate what generative AI can automate in the near term, the drop in junior employment may turn out to be temporary. The authors say they can't test this directly.
I'll admit I don't have a settled view on the future. The AI market is new and may surprise for better or worse, and anything said about it as fact today is a bet. That's why my position is about what to do with the role now. In managing people, I side with caution and the long term, ready to react quickly when a market shift proves safe.
Of the two bets, cutting is the one that's hard to undo. Stanier finishes the thought with the problem: if the logic of the cut is wrong, "you won't know until it's too late to fix." When training gets cut today, Stanier writes, the shortage doesn't show up right away. What builds up in that interval is what I call training debt.
It resembles the technical debt every tech manager knows, which is cheap to take on and stays invisible quarter to quarter until it comes due and gets expensive. The difference is in how it's paid. With effort, technical debt gets paid off in refactoring sprints, while training debt can only be paid with the years it takes for someone to become a senior. No budget can speed that up.
Aonuma wanted the team to learn the foundations by remaking the 1998 game, and the generation that would do that work was already in-house, having worked on the recent Zelda games. A company that accumulates training debt reaches that moment with no one to hand the remake to.
Keeping training in place leaves room to correct course. If AI surprises for the better, the company adjusts the next cohort and is left with people who know the business from the inside; if it surprises for the worse, the company has somewhere to find the seniors it's going to need.
Anyone who cuts juniors for fear of training people for the competitor will end up vying, down the road, for the seniors the competitor trained.