Automating junior work is the most obviously correct decision available to most organizations right now. The work is routine, well-defined, expensive in aggregate, and exactly what current AI does well. Every incentive points at it: the savings are immediate, the quality is often comparable, and nobody has to be laid off if you simply hire fewer people next year. It is the rare decision that looks good on every dimension a quarterly review measures.

It also carries a bill that arrives in about five years, payable in a currency you cannot buy at any price. Because senior people are not hired. They are grown, and the thing that grows them is precisely the junior work now being automated. An organization that removes its bottom rungs is not just saving money on juniors. It is quietly shutting down the factory that produces its own seniors, and the shutdown will not be visible until it needs a senior and finds that the supply has stopped.

The rungs are already disappearing

This is not speculative. Researchers at the Stanford Digital Economy Lab, working with payroll records from ADP covering millions of workers, found that early-career workers aged 22 to 25 in AI-exposed occupations experienced 16% relative employment declines, controlling for firm-level shocks, while employment for experienced workers in the same occupations remained stable. The asymmetry is the story. AI is not reducing employment in these fields generally. It is reducing employment at the entry point specifically, while leaving the senior tier untouched.

Their paper is titled "Canaries in the Coal Mine," and the metaphor is apt in a way that is worth taking seriously. The canary does not die because it is unimportant. It dies first because it is most exposed, and its death is information about what is coming for everyone else. Entry-level employment is the leading indicator here, and what it indicates is a structural change in how organizations source labor at the bottom, which necessarily becomes a change in how they source expertise at the top a few years later.

Expertise is manufactured, and the junior rungs are the factory

The reason this matters is a fact about expertise that organizations rarely make explicit: nobody arrives senior. Every experienced person in your company was manufactured, over years, by a specific process, and that process is doing progressively harder work with progressively less supervision until judgment forms. The analyst who becomes the person who can look at a set of numbers and know something is off did not acquire that by being told. They acquired it by working through hundreds of sets of numbers, most of them routine, some of them wrong, and slowly building a sense of what normal looks like.

That is what the junior rungs are for. They look like cheap labor performing routine tasks, and they are also the only mechanism by which the routine becomes intuitive and the intuitive becomes judgment. The work is the curriculum. When the curriculum gets automated, you keep the output and lose the education, and the loss does not register anywhere because education was never the stated purpose of the role.

There is a version of this argument about individuals losing their own judgment-building reps, which is real and covered elsewhere in what you're actually automating when you automate a role. The version here is organizational and colder: it is about supply. Your company needs a certain number of senior people in 2031. Those people are, right now, either doing junior work somewhere or not existing. If the industry collectively stops running the junior tier, the 2031 senior tier does not get made, and no amount of budget in 2031 conjures it, because the thing you need is time-under-load and time cannot be purchased in a hurry.

Why nobody notices until it is too late

The cruelty of this failure mode is its timing. Cut junior hiring this year and everything looks fine this year, and next year, and the year after. The seniors you already have keep doing senior work. The AI handles the junior work at lower cost. Every metric improves. There is no point during the first several years at which a dashboard shows a problem, because the problem is an absence of something that would only have appeared later.

The signal finally arrives as a hiring problem. A senior person leaves and the replacement search takes twice as long and costs 40% more, and the candidates who do appear are being bid on by everyone else who also stopped growing their own. At that point the organization concludes it has a recruiting problem and hires a better recruiter, which does nothing, because the shortage is not in your ability to find senior people. It is in the total number of senior people the industry produced, and that number was set years earlier by hiring decisions that looked, at the time, obviously correct.

This is also why the problem is difficult to solve competitively. If only your company cut junior roles, you could buy seniors from firms that did not. When the whole market makes the same rational local decision simultaneously, there is no one left to buy from. Everyone optimized their own pipeline away and then discovered they were all drinking from the same well.

The honest part: the evidence is genuinely contested

It would be dishonest to present this as settled, because the causal story is harder to establish than the headline suggests, and there are real objections worth engaging.

The obvious alternative explanation is macroeconomic. Entry-level hiring fell during a period of sharply higher interest rates and post-pandemic correction in exactly the sectors that had over-hired, so the decline might have little to do with AI. The Stanford authors took this seriously enough to publish a follow-up specifically addressing interest rates and timing, and concluded that while rates affect overall employment, they do not explain the disproportionate decline concentrated in AI-exposed occupations. Their results also held after excluding technology firms and remote-capable roles, which addresses the suspicion that this is really a tech-sector story. That is more rigor than most claims in this space receive, though it remains a working paper on a fast-moving question, and reasonable economists still disagree.

There is also counterevidence pointing the other way. Some firms making the largest AI investments have increased entry-level headcount rather than cutting it, and graduate hiring projections in some surveys remain positive. Aggregate employment in AI-exposed occupations has moved only slightly. The effect is concentrated and real at the early-career margin, not a general collapse.

And there is a legitimate objection on the merits: not all junior work builds expertise. Plenty of entry-level work is genuine drudgery that teaches nothing, retained because it was cheaper to hire someone than to fix a process. Automating that work is unambiguously good, and defending it on developmental grounds would be sentimentality dressed as strategy. The argument is not that all junior work is sacred. It is that some junior work is load-bearing for expertise formation and most organizations have never distinguished which.

The lever: automate versus augment

The most useful finding in the Stanford work is not the headline number. It is buried in the abstract: employment changes were concentrated in occupations where AI automates rather than augments labor. Where AI replaced the work, entry-level employment fell. Where AI assisted the worker, it held or grew.

That distinction is the whole lever, because it is a choice organizations make rather than a condition they suffer. You can deploy the same technology in two configurations. In one, the AI does the junior task and you need fewer juniors. In the other, the junior does the task with AI assistance, producing more and better work than they could alone, and in the process still doing the work, still seeing the outputs, still building the pattern recognition that turns them into a senior in six years. The first configuration is cheaper this year. The second one is the only one that keeps the factory running.

Choosing augmentation is not free and should not be presented as costless. It means accepting a higher cost per unit of output today, since you are paying both for the AI and for the person. What you are buying with that premium is your own future senior tier, which is either worth it or not depending on whether you expect to need experienced people in five years. Most companies do, and almost none of them have priced it.

What a leader should actually do

The practical discipline has two parts, and the first is an inventory nobody currently runs. For each piece of junior work you are considering automating, ask whether it is drudgery or curriculum. Does doing this task repeatedly build a judgment your senior people rely on, or is it just tedious? If it is drudgery, automate it fully and without guilt. If it is curriculum, the default should be augmentation rather than replacement, which keeps the human in contact with the work while raising their output.

The second is to make the succession cost explicit in the business case. When someone proposes automating a tier of junior work, the case should state how many senior people that tier was producing per year and where those people will now come from. If the honest answer is "we will hire them," ask from whom, given that every comparable firm is making the same decision at the same time. The point is not to block the automation. It is to stop pretending the decision has no cost simply because the cost is deferred past the horizon of the analysis.

Organizations flattening the middle of the org chart are already discovering a version of this, since the layer being removed was also where a lot of judgment got developed, an effect explored in what happens to the middle of your org chart. Between the two, a pattern emerges that should concern anyone planning past the next fiscal year: we are efficiently removing both the rung people climb from and the rung they climb to, and calling it productivity.

FAQs

Q1. What does the Stanford research actually show?
Using ADP payroll data covering millions of workers, Brynjolfsson, Chandar, and Chen found that workers aged 22 to 25 in AI-exposed occupations experienced 16% relative employment declines after controlling for firm-level shocks, while employment for experienced workers in the same occupations stayed stable. The decline was concentrated in occupations where AI automates rather than augments, and held after excluding technology firms and remote-capable roles.

Q2. Couldn't this just be interest rates and the post-pandemic correction?
That is the strongest alternative explanation and the authors addressed it directly in a follow-up note. Their conclusion was that while interest rates affect overall employment, they do not account for the decline being disproportionately concentrated in AI-exposed occupations specifically. It remains a working paper on a contested question, so treat it as strong evidence rather than settled fact.

Q3. Why does automating junior work threaten senior talent?
Because senior expertise is produced by doing junior work over time, not acquired through training or hiring. The routine tasks that look like cheap labor are also the mechanism by which pattern recognition and judgment form. Remove the tasks and you keep the output while losing the education, which means the organization stops manufacturing its own future senior tier.

Q4. Why is this so hard to notice?
Because the cost is deferred past the horizon of any normal review. Cutting junior hiring improves every near-term metric and shows no downside for several years, since existing seniors continue doing senior work. The absence only becomes visible when you need to replace a senior person and discover the supply has thinned, at which point it registers as a recruiting problem rather than a consequence of decisions made years earlier.

Q5. Can't we just hire experienced people when we need them?
Only if someone else grew them. That works when your firm alone cuts its junior tier, and fails when the whole market does the same thing simultaneously, which is what is happening. Everyone optimizing their own pipeline away leaves no external pool to draw from, and experience is one of the few inputs that cannot be accelerated with money.

Q6. Is all junior work worth preserving?
No, and pretending otherwise would be sentimentality. A great deal of entry-level work is genuine drudgery that builds nothing and exists because hiring someone was cheaper than fixing a process. That work should be automated without hesitation. The discipline is distinguishing the drudgery from the work that functions as curriculum, which most organizations have never done.

Q7. What is the difference between automating and augmenting, in practice?
Automating means the AI performs the junior task and fewer juniors are needed. Augmenting means the junior performs the task with AI assistance, producing more than they could alone while still doing the work and building the judgment it develops. The Stanford data found entry-level employment declines concentrated in automation rather than augmentation, which makes the configuration a deliberate choice rather than an inevitability.

Q8. What should go in the business case for automating junior work?
An explicit statement of how many senior people that tier was producing annually and where the replacements will come from. If the answer is external hiring, the case should name the source, given that comparable firms are making identical decisions on identical timelines. The purpose is not to prevent automation but to stop the succession cost from being invisible simply because it falls outside the analysis window.