Is the Robot Really Cheaper Than the 23-Year-Old?

A junior analyst gives notice, and somebody in the budget meeting says the thing half the room is thinking. Do we even need to backfill that seat? The team has agents now. The slide practically writes itself: one salary versus one subscription, and the subscription wins by a mile.

I spent close to fifteen years at McKinsey and Deloitte building slides exactly like that one, so I mean this professionally. It prices the wrong product.

The replacement is not hypothetical

This stopped being a thought experiment somewhere in 2023. Stanford’s Digital Economy Lab pulled payroll records from ADP, the biggest payroll processor in the country, and found that employment for workers aged 22 to 25 in the most AI-exposed occupations fell 13 percent relative to their peers since late 2022. Employment for experienced workers in those same occupations held steady or grew. The researchers checked back a year later and the pattern had not gone away.

Same jobs, different ages

The entry rung is being sawed off specifically.

Source: Brynjolfsson, Chandar, and Chen, “Canaries in the Coal Mine?” Stanford Digital Economy Lab, 2025. ADP payroll data. Change is relative to less-exposed peers, since late 2022.

Anthropic’s own CEO told Axios that AI could eliminate half of entry-level white-collar jobs within five years. When the person selling the technology is issuing the warning, take the warning. And hiring already moved: SignalFire’s talent data shows new-grad hiring at big tech down more than 50 percent from pre-pandemic levels, with new graduates now 7 percent of hires.

So companies are running the math, and the math keeps saying robot. The problem is what the math thinks it’s pricing.

What the salary was actually buying

Nobody ever hired a 23-year-old for their year-one output. For most of that first year a junior is worth less than they cost, and everyone who has ever managed one knows it. Companies made the trade anyway, for a hundred years, at every bank and law firm and agency and consultancy on earth. They weren’t bad at arithmetic.

They were buying a different product. The entry-level job has always been two things bundled together: mediocre work you have to correct, and the person that correcting produces. The midnight deck formatting, the model reconciliation, the first drafts that come back bleeding red ink. Nobody kept juniors around because those drafts were good. The work was tuition, and it was also the curriculum. You cannot develop judgment about work you have never done, and the grunt work was where the doing happened, supervised, at low stakes, while somebody senior caught the damage.

Economists have argued for sixty years about who should pay for general training. The apprenticeship was the industrial answer: the junior pays through low output, the firm pays through supervision, and the trade balances because the firm captures the senior that comes out the other end.

Then AI broke the deal, in a more specific place than the layoff headlines suggest. Juniors are as capable as they’ve ever been. What collapsed is their tuition currency: the cheap correctable work they used to trade for training is now nearly free from a machine, so the trade stops penciling, so the req gets deleted. The spreadsheet concludes the junior is overpriced. But the junior was never priced on output. The slide is comparing an agent to a deck, when what was for sale was the person who makes decks this year and judgment calls in five.

The analyst was the product all along.

That’s the mistake, and almost every AI-versus-headcount analysis I’ve seen this year makes it.

The math nobody runs

Since the slide loves math, give it the whole model instead of the first row. My napkin, stated as a napkin.

Year one, the junior costs you call it a hundred grand loaded and produces maybe sixty of supervised value. The agent costs fifteen all-in and produces the same deliverables faster. The robot wins year one by a landslide, and if year one were the whole game, this article would be one sentence long.

Year three, the junior runs the workflow, supervises the agent’s output, and catches the error the agent invented, because three years ago she made the same error herself and someone caught hers. The agent still costs what it costs and still needs the checking. Two of the numbers on this row are the same as year one. The junior’s number moved.

Year seven, she manages the function, carries the institutional memory, and knows why the biggest client almost left in 2029. There is no agent row on this line of the model at any price.

And if you skipped the hire, year seven has a different line: recruiting a senior operator on the open market. Standard recruiter fee runs a quarter of first-year comp, ramp takes months, and you’re betting culture fit on interviews, which is roulette with better catering. You’re renting someone else’s year-seven, at a premium, precisely because you didn’t grow your own. Two costs from earlier in this story also belong on the slide: the senior hours spent checking AI output that looks finished and isn’t, which BetterUp and Stanford clocked at about two hours per incident, and the liability when nobody checks, which you own in court, as Air Canada learned for $650.88. Meanwhile the felt experience of AI speed is a bad instrument: in METR’s randomized trial, experienced developers using AI were 19 percent slower while believing they were 20 percent faster. Enthusiasm is real. It is not a measurement.

Cheaper is a unit-cost word. Hiring was never a unit-cost decision. The agent is cheaper per task, this quarter. The 23-year-old is cheaper per decade. Both statements survive contact with the data, and only one of them fits on the slide.

Some reqs deserve the agent

Before this reads like a defense of every entry-level job on earth: some of them should absolutely go to the machine.

The test is whether the job was a classroom or a conveyor belt. Ask when someone was last promoted out of the role. If the answer is never, if the work was pure throughput and no senior person at your company ever came up through it, then that seat was production that happened to be done by young people. Automating it is just automation. Take the savings without guilt.

But if you can name three of your best people who started in that seat, that req is the nursery, and you’re about to demolish it for a subscription. Stanford’s data backs this split: the entry-level declines concentrate in occupations where AI automates the work rather than augments it. The market is deleting exactly the rungs that were closest to pure production, without asking which of them were also load-bearing.

Ask a law firm partner how associate development went after e-discovery software ate document review. The profession has spent a decade asking where trained fifth-years are supposed to come from. That’s the preview.

Growing judgment when the machine does the reps

The standard objection lands here: fine, but I’m still not paying someone to format decks the agent can format. Correct. The reps that built judgment used to come free with production work. Now they have to be designed, the way aviation designed them when autopilot took over the flying and airlines had to decide, on purpose, to keep pilots hand-flying approaches so the skill would exist when the automation quit.

The version I’m building with clients looks like this. The junior does the work first, then the agent does it, and the two outputs get compared line by line. The diff is the curriculum: every place the agent was better teaches technique, and every place the agent was wrong teaches the more valuable thing, which is that the agent gets things wrong while sounding right. Then flip it. The junior becomes the named reviewer of the agent’s production work, with real accountability, which happens to be the job Microsoft’s data says is coming anyway. You keep a deliberate ration of raw reps, because a reviewer who has never produced the work can’t tell polished from correct. What you’ve built is an apprenticeship in judgment that runs on the same tasks the agent automated. The task got cheap. The learning didn’t have to die with it.

Is anyone walking it back?

If the trade really collapsed the way I’ve described, you’d expect a correction by now, some visible re-hiring once the review bills came due. What the data shows instead: intentions are warming, payroll isn’t. Employers keep telling surveys they’ll hire more, and NACE’s latest projection has new-grad hiring up about six percent for 2026. That sounds like a thaw until you notice it’s a bounce off a crater. The hires rate for young graduates remains depressed, and the New York Fed has unemployment for recent college graduates at 5.6 percent, above the national rate. That reverses a pattern older than everyone reading this. The degree used to buy better odds than the market, and for its newest holders it currently buys worse. Stanford’s team re-checked their canaries a year on and the decline had held. The market is not correcting this on its own. Which is exactly what makes the next part true.

The arbitrage nobody is naming

Now the part I’d actually move on if I ran a mid-market company.

Big tech cut new-grad hiring by more than half. The Stanford numbers say the whole market is following. Which means the strongest, most AI-fluent entry-level class in a generation is standing in the market, largely unhired, at the exact moment their tools make one good junior more productive than three of me in 2010.

For twenty years, mid-market companies lost every talent auction that mattered, because Google could outbid anyone for anyone. Google just left the auction. A mid-market firm that hires four of those graduates this year, runs the apprenticeship above, and holds them four years, walks into 2030 owning a senior bench its competitors will be renting at recruiter prices. The window exists precisely because everyone else’s spreadsheet stops at year one. Asymmetries like this show up in talent about once a decade, and the last few went to whoever moved before the consensus did.

Before you delete the req

Run the classroom test. When did someone last get promoted out of this role? Never: automate it without guilt. If three of your best people started there, you’re pricing the nursery.
Extend the model past year one. Put year three and year seven on the same slide as the subscription, including the rented-senior premium if you skip the hire.
Price the supervision. Senior review hours at senior loaded rates, next to the agent’s invoice. The ten-to-one gap shrinks fast.
Design the reps. Junior first, agent second, compare the diff. Then make the junior the agent’s named reviewer, with a deliberate ration of raw production kept on purpose.
Count backwards from 2031. Seniors you’ll need, minus juniors you’re training, equals the premium you’ll pay to rent someone else’s. Make someone look at that number before any req dies.

I was the expensive junior once. My first year in consulting I was worth less than my salary while people with better judgment fixed my work, and every skill I sell today, including the ones in this article, started in that year. The firms that paid for it captured the next fifteen.

The robot is cheaper the way skipping the planting season is cheaper. True all spring. Expensive at harvest, and the harvest price gets set by whoever planted. That deal, cheap now for valuable later, is still on the table. It just stopped being the default, which means for the first time in decades, the companies that take it will get paid for taking it.

How RLK Can Help

The AI Business Case engagement exists for exactly this: the model past year one, supervision and liability included, before a headcount decision gets made on sticker price. My free AI Diagnostic shows where AI is actually earning its keep in your organization in about eight minutes. And if the question is bigger than one req, the Operating Teardown maps the work itself before you decide who, or what, should do it. Start the conversation.