The Missing Middle: Why Cutting Junior Developers Will Cost You Senior Engineers
Entry-level developer hiring has collapsed by 67% since 2022. CS enrollment is falling for the first time in a decade. The industry is treating this as an efficiency gain. It is a time bomb.
I want to tell you about something that happened in 2008, because the industry has apparently forgotten it.
When the financial crisis hit, tech companies froze junior hiring almost overnight. The reasoning was identical to what we hear now: budgets were tight, senior engineers could absorb the workload, and the junior pipeline was a luxury the business could not justify. The freeze lasted roughly two years. By 2012, those same companies were tearing their hair out trying to find engineers with three to five years of experience. The engineers did not exist. They had never been hired. The gap had migrated upward into the mid-level and senior bands, and no amount of recruiter aggression or signing bonuses could fill it.
We are doing the exact same thing again. Except this time, the freeze is not driven by a recession. It is driven by a spreadsheet.
The Spreadsheet That Ate the Pipeline
The calculation is simple, and that is exactly why it is dangerous.
Before AI tooling: one senior developer at $150,000 base, two mid-levels at $120,000 each, three juniors at $90,000 each. Total: $630,000 a year. After AI tooling: one senior, three mid-levels, zero juniors, plus under $2,000 a year in Copilot Business licenses. Total: around $512,000. Savings: $118,000 a year, plus velocity gains the CTO can put on a slide.
That is the math being presented in every budget meeting in the industry right now. I understand its appeal. I have sat in rooms where this math was presented and watched people nod. It is clean. It is legible. It is also missing half the equation.
Salesforce announced it would hire zero new software engineers in 2025. Marc Benioff said it plainly: AI productivity gains meant the company no longer needed entry-level headcount. Thirty-seven percent of business leaders now say they would rather deploy AI than hire a recent graduate. The signal could not be louder.
And the market has responded. ByteIota research shows entry-level developer opportunities down approximately 67% since 2022. In the UK, entry-level technology roles fell 46% in 2024. Here is the detail that should unsettle you: job postings labelled “entry-level software engineer” actually grew 47% in that period, but actual hiring into those roles dropped 73%. The postings exist. The jobs do not.
The Stanford Digital Economy Lab, working with ADP payroll data covering over 25 million workers, found that employment for software developers aged 22 to 25 declined nearly 20% from its late 2022 peak. Computer engineering graduates now face 7.5% unemployment. Higher than fine arts graduates. Let that land for a moment. A computer engineering degree in 2026 has worse employment outcomes than a fine arts degree.
Seniors Are Not Hired. They Are Grown.
Here is the thing about senior engineers that every hiring manager knows but no budget model captures.
Nobody becomes a senior engineer by reading documentation. Nobody develops the judgment to make architectural trade-offs, to diagnose production incidents under pressure, to sense when a specification is dangerously ambiguous, without years of doing those things badly first, in an environment where someone more experienced caught the mistakes and explained why they mattered.
That environment is called a team with junior developers on it. And the industry is dismantling it.
The Harvard study tracking 62 million workers across 285,000 U.S. firms found that junior employment at AI-adopting companies declined 9 to 10% within six quarters of implementation. Senior employment stayed flat. The mechanism is quiet. Companies are not laying juniors off. They are simply not posting the roles. The requisitions vanish and nobody notices because the seniors, augmented by AI, are absorbing the immediate workload.
But absorbing the workload is not the same as replacing the function.
Juniors were never just cheap labour for simple tickets. They were the entry point of a knowledge transfer system that takes a decade to produce a senior engineer. The code they wrote was often rough. The mentoring that corrected it was the entire point. Every senior engineer writing production code today was once a junior who needed someone to take a chance on them. Every good architectural decision they make in 2026 is downstream of a mistake they made in 2016 that somebody had the patience to walk them through.
Cut the input, and the output does not decline immediately. That is the trap. It declines five years from now, when you need mid-level engineers and they do not exist. It declines ten years from now, when you need senior architects and the cohort that should have produced them was never hired.
The Review Assembly Line
The UC Berkeley and Yale study published in Harvard Business Review in February 2026 tracked 200 employees at a tech company for eight months. What they found is the opposite of the productivity story the vendors are selling.
Senior engineers are not spending their freed-up time on architectural thinking. They are spending it reviewing AI-generated code. Task expansion, not task reduction, was the dominant pattern. AI speeds up code production, which raises velocity expectations, which means seniors are now handling the workload that used to be distributed across a larger team while simultaneously reviewing the output that replaced the junior contributions.
One engineer in the study put it in terms I have not been able to get out of my head: the job does not feel like engineering anymore. It feels like being a judge on an assembly line that never stops, stamping pull requests that keep coming.
I wrote about the review bottleneck in detail last week in “The AI Productivity Paradox.” The short version: Faros AI’s telemetry showed review times up 91% across teams with high AI adoption. The bottleneck migrated from code production to code review. But what I did not write about, and what the Berkeley study makes painfully clear, is what this means for knowledge transfer.
Before AI, the review process was also a teaching process. A senior reads a junior’s PR, leaves comments, explains why the approach was wrong, suggests a better pattern. The junior learns. The next PR is better. Over months and years, this accumulates into competence, then expertise, then judgment. It is slow and expensive and occasionally frustrating, and it is the only mechanism the industry has ever had for producing senior engineers.
Now the senior is reviewing AI output instead. The AI does not learn from the feedback. The AI does not get better at understanding the codebase’s specific constraints. The review time is spent, the judgment is applied, and nothing comes back. No growth. No knowledge transfer. Just an empty loop.
The Signal Reaching the Classroom
The damage has breached the education pipeline.
For the first time since the early 2010s, CS enrollment is falling. The National Student Clearinghouse reported a 15% drop at graduate institutions and nearly 6% at undergraduate two-year institutions in Fall 2025. The CERP Pulse Survey found 62% of computing programmes reporting declining enrollment in 2025-26. The decade-long growth trend that took CS bachelor’s degrees from roughly 51,700 in 2013-14 to 112,700 in 2022-23 has reversed. Forrester predicts a further 20% decline.
I have been thinking about what it feels like to be a CS student right now. You chose this degree because it was supposed to be the safe bet. The lucrative career. The field that was always hiring. And then you watched Salesforce say “zero engineers.” You watched entry-level postings evaporate while LinkedIn filled up with senior engineers posting about their layoffs. Handshake data shows 64% of pessimistic CS majors cite generative AI specifically as a factor. Georgia Tech reported a 34% drop in on-campus tech recruiting visits between 2023 and 2025. Tech internships dropped 30% since 2023 while applications rose 7%.
These are not abstract numbers. These are 22-year-olds making rational decisions with the information available to them. They are leaving. Some to cybersecurity, some to data science, some out of tech entirely. The BLS still projects 17% growth for software developer roles through 2033. But the students cannot see that projection from where they are standing. What they can see is a job market that does not want them.
The feedback loop is now self-reinforcing. Companies cut junior hiring. Students see the signal. Students leave CS. Fewer graduates enter the pipeline. The talent shortage intensifies five years from now. Companies compete aggressively for scarce senior talent, driving costs up. The CFO’s $118,000 annual saving transforms into a $200,000 annual premium for every senior hire the company cannot fill.
What AI Cannot Reconstruct
I need to make this concrete because I think the abstraction lets people off the hook.
I build pharmaceutical compliance software for American hospitals. In my domain, the systems carry regulatory weight that compounds over decades. A drug tracking rule implemented in 2019 interacts with a DEA reporting requirement from 2021, which intersects with a state-level dispensing regulation that changed in 2023. The engineer who understands why the system behaves the way it does is not necessarily the one who wrote the code. It is the one who was in the room when the decisions were made. Who watched the requirements evolve. Who debugged the edge cases when the regulations conflicted with each other.
AI can read the code. It can even explain what the code does. But it cannot reconstruct the reasoning behind why the code does it that way instead of the three simpler ways that would have been wrong for reasons that are not written down anywhere. That knowledge lives in the heads of engineers who were present for the original conversations, and it transfers to the next generation through the kind of daily proximity that junior developers have with their senior colleagues. Not through documentation. Not through onboarding decks. Through being in the room when something breaks and watching how someone with twenty years of context approaches the problem.
Every complex software system accumulates this kind of institutional memory. Enterprise platforms, financial systems, healthcare infrastructure, payment rails. I have worked across enough of these domains, from fintech in Lagos to health-tech in Edinburgh, to know that the pattern is universal. The knowledge is always more fragile than people think, and the transfer mechanism is always slower than people want it to be. There are no shortcuts. There is only time and proximity and patience.
The industry is optimising for a world where AI can absorb junior tasks. It is ignoring the fact that junior tasks were never the point. The learning was the point. The knowledge transfer was the point. The slow, sometimes painful process of turning a confused graduate into a competent engineer was the point.
The 2031 Cliff
Let me lay out the timeline plainly.
2022 to 2026: Entry-level hiring collapses 67%. CS enrollment begins declining. The current cohort of would-be juniors either pivots to other fields, takes non-engineering roles, or leaves tech entirely. Klarna freezes developer hiring entirely in late 2023, citing AI productivity. Within a year, they quietly start hiring humans again because the strategy did not survive contact with production reality.
2028 to 2030: Companies that froze junior hiring begin searching for mid-level engineers with three to five years of experience. The candidates do not exist in sufficient numbers. The ones who do exist command premium salaries because the supply is structurally thin. This is the 2012 talent desert, replaying with better data and worse fundamentals.
2031 to 2036: The senior talent gap becomes acute. The engineers who should have been growing into architectural and leadership roles were never hired as juniors. The industry faces a shortage of exactly the people it needs most: engineers who can make system design decisions, review AI-generated code for correctness, maintain complex distributed systems, and hold the institutional knowledge that keeps production running.
Korn Ferry projects $8.5 trillion in unrealized revenue globally by 2030 due to talent shortages. The junior developer crisis is not the only cause. But it is the one the industry is actively choosing.
The Investment Frame
The answer is not to hire juniors out of sentimentality. It is to recognise that junior hiring is infrastructure investment, not headcount cost. The same way you budget for CI/CD tooling or observability platforms or security audits, you budget for the pipeline that produces the people who will run your systems in five years.
The organisations getting this right are reframing what junior roles look like. They are not hiring juniors to write boilerplate, because AI does handle that now. They are hiring juniors to review AI output, to write verification tests, to build judgment through structured mentorship. To learn by being in the room. To absorb the knowledge that only transfers through proximity and time.
This costs more per junior than the old model. The onboarding is more structured. The mentoring commitment is heavier. The fundamentals bar at hiring is higher. But the alternative is no investment at all, followed by a talent crisis that makes every AI productivity gain in the ledger look trivial by comparison.
I wrote in “The Architect Shift” that engineers are moving from writing code to overseeing AI agents. I wrote in “The Spec is the New Code” that specification quality is now the highest-leverage skill. I wrote in “The AI Productivity Paradox” that output is going up while outcomes are flat. But none of that matters if there is nobody in the pipeline learning to become the engineers who do those things well.
The spec is the new code. The architect is the new engineer. And the junior you choose not to hire today is the senior you will not be able to find in 2031.