The Seniority Debt: Who Makes Your Next Senior?

This morning the news told me what I already watch happen inside companies I back. A fresh survey says almost one in five business leaders has cut entry-level recruiting, and of the ones who cut, the largest group blames AI. The org chart is changing shape in public. The wide base of the pyramid, all those first jobs, is being sanded down, and the picture everyone now draws instead is a diamond: thin at the bottom, fat in the middle.

Most of the argument about this is a fair one, and it is not the founder’s argument. People are asking whether it is bad for young people trying to start careers. It is. But when I sit with a founder looking at their headcount plan, that is not the question that decides anything. The question that decides things is quieter and it arrives late: if you stop hiring and training the bottom, where do your seniors come from in three years?

You did not cut a salary line. You unplugged the machine that makes the one kind of worker AI still cannot replace. I call the thing you take on when you do that the seniority debt. It is cheap to borrow and expensive to carry, and almost nobody prices it into the decision.

Here is the whole idea in one sentence, and the rest of this is me defending it: the cost of automating junior work is not paid at the junior level. It is paid years later, at the senior level, where there is suddenly no one to promote and no one to hire, because every company made the same cut at the same time.

Seniority is manufactured, and you just unplugged the machine

A senior is not a thing you buy off a shelf. A senior is manufactured, slowly, out of a beginner. Every senior person on your team was once someone who did a few thousand small, cheap, forgettable tasks under somebody’s eye. They got a lot of them wrong. They got corrected. They compiled all of that into judgment, which is really just a large private library of “I have seen this go badly before.” The junior work was never only output. It was the training set that produced the senior.

AI is very good at exactly that layer. The cheap, structured, correctable first-pass work is the sweet spot: code scaffolding, first-draft copy, ticket triage, data cleanup, research synthesis, meeting notes. That list is not mine, it is the list of what employers say they now hand to agents. And the per-quarter math is brutal because it always points the same way. An agent costs a few thousand dollars a year and works at two in the morning. A junior costs fifty to ninety thousand fully loaded and needs a manager’s time on top. Put those side by side and you will automate the junior every single time. Each cut looks like a clean win, and on this quarter’s numbers it is one.

This is not a forecast, it already happened. SignalFire’s data on hiring shows new graduates fell to about seven percent of big-tech hires, down more than half since 2022; at startups the share went from around thirty percent to under six. Starts by people with less than a year of experience dropped roughly half over five years. Entry-level hiring at the fifteen largest tech firms fell about a quarter from 2023 to 2024. The base of the pyramid is not thinning in theory. It has already thinned.

The stake is simple once you name it. You are draining the tank you plan to drink from later, and you are doing it at the exact moment the tank is hardest to refill from outside. Three quarters of employers worldwide already report they cannot fill open roles. Demand for AI engineers runs near 1.6 million openings against roughly half a million qualified people, better than three to one. And the wage premium climbs with every rung: a few points at entry, close to twelve percent at mid, over fourteen at senior, near nineteen at staff. That is scarcity being priced in real time. A senior engineer fully loaded now runs past a quarter million dollars a year and the number keeps moving up.

The saving is a loan, and the interest is a senior you can’t hire

A debt is money you get now and pay back later with interest. The seniority debt works the same way. When you automate the bottom rung you collect the saving now, and it is real, it lands on this quarter’s profit and loss. You repay it later as the cost of a senior bench you never grew. The interest is the gap between what a homegrown senior would have cost you, which is a junior salary spread over a few years plus some mentoring time, and what a scarce outside senior costs you in a bidding war three years from now, if you can find one at all.

The reason the debt is invisible is that the saving is legible and the debt is not. The saving shows up as a line you cut, with a name and a number. The debt shows up as an absence: a role you cannot fill, a project that stalls because the one person who could have owned it left and there was nobody behind them. Absences do not get line items. Nobody is ever blamed for a senior they failed to grow, the same way nobody is blamed for the thinking muscle that quietly went slack. The gain is booked and the liability is hidden, so the decision feels free when it is not.

What you actually stop producing when you cut the base is two things that look like one. The first is judgment, which I will come back to, because it is the whole game. The second is managers, and that one is worth its own section, because founders miss it completely.

Seniority is compiled from cheap reps, not bought off a shelf

Think about how a person actually becomes good. It is not a course. It is reps with outcomes attached. You try something small, you are wrong in a way that costs almost nothing, someone shows you why, you adjust, and you do it again. Do that a few thousand times across real situations and the pattern-matching gets fast enough that we call it intuition, and then good enough that we call it judgment. This is the ordinary machinery of expertise, the deliberate-practice loop that every field runs on, and I wrote about the individual side of it in the apprenticeship gap.

Now look at where AI sits in that loop. It does the cheap first-pass reps. That is the good news for output and the quiet disaster for formation, because the value of those reps to the learner was never the output. It was the chance to be wrong on something small, where being wrong cost almost nothing. The junior’s rough draft was two things at once: cheap labor, and tuition. AI outbid the junior on the labor. When the labor left, the tuition went with it. What is left for a beginner to learn on are the expensive tasks, the ones where being wrong actually hurts, and no sane manager hands those to someone with no reps. So the beginner does not get worse. The beginner never gets made in the first place.

Two concrete cases make it real. A support rep who spends a year on live tickets learns, without noticing, which customers are about to leave, which complaints hide a real product bug, and which promises the company should never make to anyone. Hand every ticket to an agent and you get faster replies and nobody who learned any of that. A first-year engineer who ships small fixes learns which parts of the codebase bite back, and why the clever change is usually the wrong one. Let the agent write those fixes and you get merged code and nobody who built the scar tissue. In both, the output survives and the education quietly dies.

That matters more now, not less, because the specific thing a senior supplies is the thing AI is worst at. A senior notices when the plausible answer is quietly wrong, decides what is worth doing at all, and owns the outcome when it breaks. That is the initiative and ownership agents do not have. You are removing the pipeline for exactly the input your product will be starved for.

Every quarter you make the correct cut, and the sum is wrong

I want to be careful here, because this is not a story about one dumb decision. Nobody sits down and decides to stop making seniors. The trap is subtler and much harder to fight: every individual decision along the way is correct. Automating this ticket queue is cheaper and faster and you can measure it. Handing that data-cleanup to an agent frees a person for better work. Each call, taken on its own merits, is the right call. The damage lives entirely in the sum.

This is the same shape as most of the quiet AI traps I keep running into. Locally correct choices add up to a slope you never chose to walk down. You are not deciding, once, to hollow out your org. You are making forty small correct decisions that happen to add up to exactly that, spread across two years so no single one feels like the moment it happened. The per-role math, the actual hire-or-automate call on a given task, can be right forty times and still leave you, at the portfolio level, with no one to promote.

The buy column looks fine right up until the supply row bites. Growing a senior is a slow, boring, front-loaded investment, and automating the junior work is a fast, legible, immediate return. Human beings and quarterly reviews both reward the fast legible thing. That is precisely why the debt accumulates: the market’s incentives are all pointed at borrowing, and none of them are pointed at the repayment.

You stop producing managers, not only workers

Here is the part that surprises even good founders. The bottom rung does not just feed you senior individual contributors. It feeds you managers. You learn to manage by managing someone, and the someone you learn on is a junior. Cut the juniors and you have not only stopped making senior engineers and senior operators, you have stopped making the next layer of people who can run teams, because there is no team to run.

The forecasts already point straight at this. Gartner expects that through 2026 about a fifth of organizations will use AI to flatten their structure, cutting more than half of current middle-management positions, and nearly half of the companies deepest into agentic AI already expect middle-management reductions within three years. The org chart everyone is drawing, the diamond, is wide in the middle right now. The problem is that the middle is fed from the bottom and drains into the top. Thin the bottom and the same diamond, five years on, is an hourglass with nothing coming up through it.

The sharpest version of this I have read put it plainly: eliminating middle managers in 2026 dismantles the machinery that produces senior leaders in 2028, and the bench thins in two directions at once. The people who would have grown into the role are gone, and the people who might have stepped sideways into it have looked at the job and decided it is not worth wanting. You do not feel either of those on the day you make the cut. You feel them the week you need to fill a leadership seat and the internal shortlist is empty.

The debt compounds the same way the bench would have

The reason to start now rather than later is that this compounds in both directions. Growing your own is not a one-time purchase, it is a loop that feeds itself. A senior you grew is also the person who trains the next two juniors, who become the seniors who train the next four. Keep the base and the whole pipeline gets cheaper and better every year, because you are building the teachers at the same time as the students. The investment pays a dividend that funds the next investment.

Kill the base and it compounds in reverse, which is the part that traps the founder who assumes they can switch it back on later. Say you cut junior hiring for three years, the outside market tightens the way it is already tightening, and you decide to restart. Who trains the new cohort? The seniors who would have mentored are stretched thin covering the work that the missing layer used to do, and the natural mentors, the people two rungs up who remember being junior recently, are the exact cohort you never hired. You cannot restart a pipeline by pouring water in at the top. The machinery in the middle has to be there, and you removed it.

This is why the bench thins in two directions at once, and it is worth saying slowly because it is the whole trap. The people who would have grown into the senior roles are gone. The people who could have trained a fresh crop are also gone. A gap you could have closed with one good graduate cohort turns into a gap that takes most of a decade, because you did not only stop producing seniors, you stopped producing the ability to produce them. Debt that compounds is the most expensive kind there is, and this one compounds.

The senior market dries up because everyone read the same memo

Every founder who cuts the base has the same unspoken plan: we will just hire seniors when we need them. That plan works beautifully if you are the only one running it. You are not. The whole industry read the same memo in the same eighteen months, and everybody is quietly betting on buying seniority from a market that all of them stopped restocking.

Play it forward. If most companies thin their junior intake at once, the pool of people who would have become available seniors three to five years later simply does not fill. The survivors get bid up, and the price of senior talent goes close to vertical, which is what the early readings already show: the three-to-one demand gap, the three quarters of employers who cannot fill roles, the widening premium at every level above entry. Those are not signs of a healthy market correcting. They are the first innings of a shortage that a lot of individually rational cuts manufactured together.

This is the piece founders underweight, because it is somebody else’s balance sheet until it is suddenly yours. When you model your own hiring, the external senior market looks like a given, an endless supply you can tap on demand. It is not a given. It is a commons, and it is being drawn down by everyone who assumes someone else will keep it full. The company that quietly kept growing its own is the one that will not be standing at that auction in 2029.

The bench you don’t see is what ships your hard quarters

Ask what a senior bench actually does day to day, because it is not filing tickets. It is the group you throw the ambiguous, high-stakes, no-clear-answer problems at: the outage nobody understands yet, the customer about to churn for a reason no dashboard shows, the architecture call that will either save you a year or cost you one. That work does not parallelize onto agents, and it does not parallelize onto beginners either. It sits with the handful of people who have compiled enough judgment to own it and be trusted with it.

So when the bench is thin, every one of those problems lines up behind the same few people. The obvious cost is speed, because hard things ship slower when there is a queue for the only people who can do them. The quieter cost is the people themselves. The clearest early signal in the under-hiring data is burnout among the seniors who stay, since they absorb a load that used to be spread across a layer that no longer exists. You do not just end up with fewer seniors. You grind down the ones you kept, and some of them leave, which thins the bench again.

This is why the debt is worse than a plain headcount gap. A missing junior is a missing junior. A thin senior bench is a bottleneck on everything in the company that needs judgment, and in an AI-heavy shop that is a growing share of what is left for humans to do at all. The work your customers actually pay you to take off their hands, once they cross the point where they stop testing your product and start depending on it, is the judgment-heavy work. Starve the bench that does it and you have capped the part of the business that was supposed to compound.

The work AI can’t do is exactly what you stopped training people for

Sit with the cruel symmetry for a second. The entire reason seniors are worth so much right now is that they hold what AI still cannot: the judgment to catch a confident wrong answer, the taste to decide what is worth building, the ownership to stay on the hook when it fails. The market is paying more than it ever has for that senior judgment. And it is paying that premium at the precise moment companies are unplugging the only pipeline that produces it. Demand for the thing is spiking while the factory for the thing is being switched off.

This is the same incompressible core I keep coming back to. When I wrote about the one-person company, the point was that the human part that does not compress under automation is exactly this judgment and ownership. It is also the part behind the recovery problem, the skill you cannot delegate because you only have it if you built it yourself. A company that stops growing seniors is choosing to have less and less of its own incompressible core over time.

There is an uncomfortable version of this for the smallest founders, the solo builders leaning entirely on agents. If you are running a company of one on top of a stack of AI, you are the senior. You are also the only junior. Your own judgment is the whole bench, and when your judgment is the bottleneck there is nobody behind you at all. That can work, and it works right up until the day the thing that breaks is a thing you never built the reps to fix.

The firms winning kept a training ground on purpose

None of this is destiny, and the proof is that some companies are running the opposite play on purpose. Brainlabs grew its entry-level cohort from nineteen hires in late 2023 to sixty-four in early 2026, a jump of well over two hundred percent, by retooling its two-year internal academy around AI instead of shutting it down. Today roughly a third of its global team is homegrown. IBM is tripling its US entry-level hiring in 2026, shifting the roles away from task execution toward analysis, problem-solving, and AI oversight. These are not charities. They decided that the pipeline is an asset worth protecting and pointed it at higher work sooner.

The honest counter-data cuts the same way. When Ramp and Revelio looked across tens of thousands of companies, the most intensive AI adopters actually showed entry-level headcount up around twelve percent, not down. A careful read of the raw hiring numbers by others in 2026 argued that the death of the entry-level job is overstated. Put those next to the cutting statistics and the real pattern is not “AI kills the junior.” It is “AI changes what a junior is for.” The companies pulling ahead did not delete the base. They kept it and rebuilt the job around directing and checking machines, which, not by accident, is the exact skill employers now say they hire for.

The move has a name I will give it plainly, because it is the one thing to take away: keep a training ground on purpose. Find the functions where the junior work can be automated and where that same work is how your future seniors get made, and protect at least one of them from the automation you would otherwise be right to do.

Automation should raise the floor of the work, not remove the bottom of the org

There is a healthy way to point AI at your junior functions, and it is the opposite of deleting them. Use the tool to raise the floor of the work rather than to remove the bottom of the org. The same agent that could replace a first-year can instead clear the drudge out of the first-year’s day, so they spend month three on work that used to belong to month thirty. That is a strictly better junior, formed faster, and it is the play the companies keeping their training grounds are actually running.

The distinction is between a floor-raiser and a rung-remover, and you make it function by function. A floor-raiser hands the beginner harder, more real problems sooner because the machine took the parts that taught nothing anyway. A rung-remover just takes the whole role. Both use the identical technology, bought from the same vendor, deployed the same week. The only difference is whether a human is still in the seat being stretched, or whether the seat is gone.

There is a failure mode hiding inside the good version, and it is an easy one to slip into. If you let the agent do the work and the junior only nods at the output, you have kept the seat and lost the training, which is the worst of both. The person has to stay above the machine, not beside it, or their sense of what good looks like quietly slides down to whatever the model hands them. Raising the floor works only when the human is still the one deciding the floor is not high enough. Keep them judging rather than approving, and a junior on top of AI becomes a senior faster than a junior ever did without it.

The junior salary was never the real cost

Most founders model this decision as a cost line and nothing else. Junior at seventy thousand, agent at four, automate, done. That framing is the whole mistake, because it prices the wrong thing. This was never a cost decision. It is a supply decision for the one input AI makes scarcer rather than cheaper, which is senior judgment. When you tally only the salary you save, you book a profit and quietly write yourself a liability that does not appear anywhere, and you will not see it until the bench is empty.

The frame that actually fits is a supply chain. Some of your functions are mines that produce future judgment. The question is not whether you can automate the work in them, you almost always can, it is whether you are strip-mining a source of seniors for one quarter of cleaner margin. Cost thinking cannot see that, because the loss is not a cost, it is a source you turned off.

I want to give the other side its due, because it is real. A lot of the old pyramid was bloated. Plenty of junior work genuinely was makework that taught almost nothing and existed mostly to justify a headcount plan. And AI-native generalists do ramp faster than the fresh graduate of five years ago, which is the honest case for the hourglass some analysts prefer to the diamond. All true. The answer is not to freeze every junior seat in amber and call it apprenticeship. The answer is narrower and harder: stop pretending the training-ground function is free to delete. Cut the makework. Protect the mine.

How to keep one training ground alive

The practical version of all this is short, and you can start it this quarter. First, run the two-by-two on your own functions. For each one ask the two questions: can the junior work here be automated, and does this function grow the judgment I will need a senior for. The box that is automatable and judgment-building is the trap. Do not automate all of it. Pick one function in that box and protect it on purpose.

Second, and this is the whole reframe, do not hire a junior instead of AI. Hire a junior on top of AI. Let the agent do the first pass, and make the human’s job to direct it, check it, and own the result. That is not a downgrade of the junior role, it is the reps that build a senior, and it happens to be exactly what employers now say they screen for. The person who can take an agent’s plausible draft and find the one wrong thing in it is worth more in three years than the person who never had to look.

Third, keep the reps expensive enough to teach. If the agent does ninety percent and the human rubber-stamps the last ten, you have automated the learning along with the labor, and you get output today and no senior tomorrow. Structure it so the person predicts before they peek at the agent’s answer, gets corrected when they are wrong, and carries something end to end that is theirs when it breaks. Struggle is where the reps land.

Fourth, put a number on your internal seniority supply and watch it like a real metric. How many people on your team could you promote into a senior seat in the next eighteen months. If that number is zero and drifting down while your automation is going up, you are carrying seniority debt no matter how good this quarter’s P&L looks. The whole ninety-day move fits in one line: pick one function, keep one person at the bottom of it, put them on top of the agents, and give them one thing to own alone.

None of this asks you to fight the tools. I use agents across everything I build now, and I would not give them back. The ask is narrower than that. When you point one at a function, decide on purpose whether you are raising that function’s floor or removing its bottom rung, and do not let the second one happen by accident forty times in a row. The technology is not the thing hurting you. Deleting the pipeline while you were looking at this quarter is.

If you want a single check to run this week, look at your last three senior hires and ask where they came from. If all three came from outside, and nobody on your current team is within eighteen months of that level, you are already renting seniority instead of growing it. That is not a crisis on its own. It becomes one the year the rental market has nothing left to rent, and the whole point of the seniority debt is that this year is easier to reach than it looks.

Questions founders ask about hiring under AI

What is the seniority debt?

It is the hidden liability you take on when you automate junior work and stop growing your own seniors. You collect a real saving now, on this quarter’s numbers, and you repay it later as the cost of a senior bench you never built: roles you cannot fill and a market you have to buy from at a rising price. The saving is legible and the debt is invisible, which is exactly why it accumulates.

Isn’t it cheaper to just hire seniors when I actually need them?

Only if you are the only company doing it. When most companies cut their junior intake at the same time, the pool of people who would have become available seniors a few years later does not fill, and the survivors get bid up. The early readings are already here: better than three qualified openings per person, three quarters of employers unable to fill roles, and a wage premium that widens at every level above entry. You are planning to buy from a commons everyone is drawing down.

Which roles should I still hire juniors for?

Run the training-ground test. Keep juniors in the functions where the work can be automated and where that same work is how your future seniors get made, like customer support, sales development, and core engineering. Automate freely in the functions that teach nobody anything, like pure data cleanup. Protect at least one mine even when automating it would be locally correct.

Does this apply to a solo founder or a tiny team?

Yes, in a sharper form. If you run a company of one on a stack of agents, you are both the senior and the only junior, so your own judgment is the entire bench. That works until the thing that breaks is something you never built the reps to fix. The move is the same at your scale: keep doing some of the hard first-pass work yourself on purpose, so your own judgment keeps compounding rather than atrophying.

How is this different from the apprenticeship gap?

The apprenticeship gap is about the individual: how a person forms expertise when the entry-level reps disappear. The seniority debt is about the firm: what a company owes itself when it stops running the pipeline that turns beginners into seniors. One is a personal-development problem, the other is an org-design and supply problem you can see on a headcount plan. They are two views of the same cut.

How do I train a junior when AI already does the entry-level work?

Put the junior on top of the AI instead of replacing them with it. The agent produces the first pass, and the human’s job becomes directing, checking, and owning the outcome, which is a harder and better rep than typing the first draft ever was. The one rule that keeps it a real training ground is to make the person engage before they see the answer, so they are building judgment and not just approving output.

How do I know if I am already carrying seniority debt?

Count the people you could promote into a senior or lead role in the next eighteen months. If that number is low or zero, and it has been falling while your automation has been rising, the debt is already on your books. Another tell is that every senior opening turns into an external search because there is never an internal candidate. That is the absence showing up.

What is the one metric to watch?

Internal seniority supply: the number of people you could realistically move up a level within eighteen months. Track it the way you track runway. If it trends toward zero, no amount of short-term margin from automation is a good trade, because you are financing this year with a bench you will not have next.

I’m a four-time founder building AI-native companies. I write about the founder mindset, AI opportunities, and the mechanics of building with AI at vikasmalpani.com.