Why eliminating entry-level roles is the most expensive saving in software
A 120-engineer infrastructure company ran the numbers on its 2026 graduate intake and decided not to make one. The logic was clean. Junior engineers cost roughly $140,000 a year in loaded compensation. For the first six to nine months they are net-negative: they consume senior review time, ship little of consequence, and occasionally break things that someone more experienced has to fix. AI coding tools, meanwhile, now produce the boilerplate, the test scaffolding, and the straightforward features that used to be a junior's first year of work. The company had been hiring eight graduates annually. Cutting that intake to zero saved about $1.1 million a year against a tooling budget a tenth the size. The board approved it in a single slide.
The saving is real. It is also a loan against a balance the company will not see for six years. Senior engineers are not recruited into existence. They are produced, slowly, by junior engineers doing junior work until they are no longer junior. The tasks AI now absorbs are not incidental to that process. They are the process. A company that stops hiring juniors has not removed a cost. It has stopped manufacturing its own future seniors and outsourced the problem to a labour market that every other company is depleting at the same time.
The Work That Made Seniors
An engineer becomes senior by building judgment, and judgment comes from a specific sequence of low-stakes failures. The graduate who writes a clumsy database query, watches it lock a table under load, and spends a tense afternoon learning why, gains something that cannot be written down. The one who ships a feature without handling an edge case, sees it surface in production, and traces it back to their own assumption, builds the instinct that later lets them spot the same flaw in someone else's design at a glance. This is the apprenticeship, and it runs entirely through the kind of bounded, forgiving work that an organisation can afford to let a beginner get wrong.
That work is precisely what AI coding tools now do faster and, for the narrow case, more reliably. The boilerplate, the unit tests, the small CRUD endpoints, the routine refactors: these were never valuable in themselves. They were the gym. A company that hands them to a machine has not eliminated busywork. It has demolished the training equipment and kept the building it was housed in. The senior engineers who run that company today were forged in a gym their employer is now closing to the next cohort.
The swap looks like efficiency because the output is identical. A test suite written by a graduate and a test suite written by an AI agent are, on the day they merge, worth the same. The difference is entirely invisible on the balance sheet: one of them also produced an engineer.
The Hiring Math That Looks Irresistible
The short-term economics are not close. A junior engineer in their first year delivers perhaps a third of a mid-level engineer's output while costing two-thirds as much, and takes an extra ten to fifteen hours a month of senior time in review and mentorship. Priced at a senior's loaded rate, that supervision is worth $30,000 to $45,000 a year on top of the salary. Against that, an AI subscription costs a few hundred dollars a seat and asks nothing of anyone's calendar. For a CFO modelling the next four quarters, there is no contest.
The market has already moved. SignalFire's 2025 State of Tech Talent report found that new-graduate hiring at big technology firms is down roughly a quarter from 2023 and more than half from pre-pandemic 2019, leaving new grads at just 7% of all hires. The cuts are concentrated at exactly the entry level, while hiring for experienced engineers has held far steadier. The industry is, in aggregate, buying finished seniors and refusing to fund the raw material that becomes them. This works for any single firm only so long as some other firm is still doing the training. It cannot work for all of them at once, and the share still doing it is shrinking every quarter.
This is the structure of a classic deferred cost. The saving is immediate, visible, and easy to claim credit for. The bill is distant, spread thin, and lands on a different executive's watch. No incentive inside an annual planning cycle rewards the manager who spends $1.1 million today to avoid a labour shortage in 2032. The manager who books the saving is promoted long before the shortage arrives.
The Shortage You Are Manufacturing
The reflexive answer is that the company will simply hire seniors instead. But wanting a senior is a statement about demand, and the shortage is a problem of supply. There is no way to want a senior into existence; the only production line that makes one runs through a junior, and it takes six to eight years to run. A firm that resolves to buy seniors rather than grow them is not escaping the cost of training. It is relying on other firms to keep paying it, and those firms are shutting the same line for the same reason.
Consider the maths on the other side of the loan. An engineer hired as a graduate in 2020 is, in 2026, a capable mid-level engineer approaching senior. The supply of genuine seniors in 2032 is being set right now, by the graduate hiring of 2026, and that hiring is being cut. The pipeline has a six-to-eight-year latency and no shortcut. There is no amount of money a company can spend in 2032 to have invested in apprenticeship in 2026.
When a scarce input has a fixed multi-year lead time and demand for it is rising, the price does not rise gently. Senior engineering compensation, already the largest line in most engineering budgets, becomes the contested resource. The firms that stopped hiring juniors discover that the seniors they planned to buy instead are now being bid up by every other firm that made the same decision. The $1.1 million saved per year compounds into a structural premium on every senior hire, paid indefinitely, in a market the company helped create.
This is the signature of a collective-action failure. Training carries a positive externality: the firm that pays to grow an engineer often watches a rival that paid nothing hire the finished product. Each firm, acting sensibly on its own books, under-invests, and the market corrects only through price and only with a lag. Toward the end of the decade, senior pay will climb far enough that growing one's own juniors is obviously profitable again, and firms will pile back in. The cohort skipped in between cannot be recovered. The correction arrives years after the decision that caused it, and it falls hardest on the firms that cut earliest.
The damage is not only to the labour market. It reaches into the work itself. The previous article in this thread argued that AI generates code without generating understanding, and that the calibrated judgment of when to trust a machine's output and when to check it is exactly what separates a productive adopter from an over-trusting one. That judgment is the senior's core function. An organisation thinning its senior ranks while increasing the volume of machine-generated code it must vet is removing the brakes at the moment it presses the accelerator. The AI raises the rate at which unreviewed code enters the system. The shrinking pipeline lowers the rate at which anyone capable of reviewing it is produced. The two trends are not independent. They multiply.
What the Few Are Doing Differently
A minority of companies have read the same hiring data and drawn the opposite conclusion. They are not nostalgic about junior roles. They have recognised that the apprenticeship is now cheaper to run than at any point in the industry's history, and that the firms abandoning it are leaving an arbitrage on the table.
Their starting point is that the irresistible math priced the wrong asset. The junior the CFO modelled, the one who spends a first year producing boilerplate, is genuinely obsolete, and the role deserves no mourning. The junior these firms hire is a different proposition. From the first week the work is to direct, read, and verify the machine that does the building, and to own a small surface end to end. That is the senior's own skill, judgment about what is correct and what fits, practised years earlier because AI removed the mechanical floor beneath it. Pointed at training rather than replacement, the same tools that make the old junior redundant become the most powerful apprenticeship accelerator the industry has built. Time to senior compresses, and the junior turns net-positive in months rather than years. The math does not soften. It inverts.
They retooled the apprenticeship rather than cancelling it. An earlier article on this site documented a company that used AI to compress a junior engineer's training across three technical domains from three years to fifteen months, dropping the senior's mentorship cost from $50,000 a year to under $5,000. The same tools that make a junior look redundant make training one far more efficient. The junior's job is no longer to produce the boilerplate; it is to direct, read, and verify the machine that produces it, which is the senior skill, learned earlier. The gym did not close. It modernised.
They moved juniors up the value chain immediately. The judgment-heavy work that once waited until year two, reading unfamiliar systems, evaluating whether a generated solution actually fits, deciding what not to build, now starts in month one. The learning curve steepens rather than flattens. A graduate two years in at one of these firms reasons about systems at a level their counterpart at a juniors-banned competitor will not reach for four.
They counted the asset, not just the cost. These firms model a junior hire as the production of a future senior, with a six-year payback and a market value at maturity that is rising precisely because their competitors stopped producing them. The timing sharpens the return: graduate labour is cheap and plentiful exactly because rivals have stopped buying it, so the firms still training take the pick of a neglected pool at a discount and hold it through the years the premium climbs. Against a senior premium that rises every year the shortage deepens, growing one's own is no longer the expensive option. It is a counter-cyclical investment dressed as a staffing decision.
What "Senior" Will Come to Mean
The instinct to hoard seniors rests on a definition that is quietly expiring. Where AI is a daily tool for writing code, seniority stops being a count of years and becomes a measure of two things: calibrated judgment about what a system should do, and fluency in directing the machines that now do the building. The two do not always travel together. An engineer three years into a career who grew up orchestrating agents and checking their output can out-deliver a fifteen-year veteran who never adapted. Tenure is becoming a weak proxy for the quality firms are actually paying for.
That is why hoarding is the wrong frame. The senior is a depreciating, contested asset that every firm is now bidding for, and some share of those seniors are coasting on experience the tools have devalued. The durable advantage does not go to the company that accumulates the most seniority. It goes to the company that manufactures judgment the fastest. Seniority is the output a firm wants; a junior paired with AI and pointed at real decisions early is the factory that produces it.
The company that cut its graduate intake to zero will report a better margin this year, and the year after. It will not connect the line item it removed in 2026 to the recruiter it cannot satisfy in 2032, or to the seven-figure premium it pays for every senior it failed to grow. The junior engineer it stopped hiring was never the cost. The junior engineer was the only thing the company was building that it could not buy back later at any price.