Replacing Junior Jobs With AI Is a Bad Business Strategy

A friend who runs engineering at a Series C company told me last quarter that he had cut his new-grad hiring plan from twelve to two. The board liked it. The CFO liked it. The remaining seniors were quietly furious, because they could see what the spreadsheet could not: in three years, half of them would be gone, and there would be nobody who had grown up inside the codebase to replace them.

This is the dumbest cost-cutting move happening in tech right now, and it is happening everywhere. The logic sounds clean. AI writes code, drafts emails, summarizes meetings, and triages tickets. So why pay a 23 year old to do work an LLM can do for the price of a few tokens? The answer is that you are not just buying output when you hire a junior. You are buying the only person on the team who has nothing to unlearn, and you are funding the senior engineer who will run a critical system in 2031. Cut that line and you save a few hundred thousand dollars this year in exchange for a hiring crisis you cannot recruit your way out of later.

The Data Is Real, and It Is Being Misread

The numbers behind the panic are not made up. The Stanford Digital Economy Lab's Canaries in the Coal Mine study found that workers aged 22 to 25 in the most AI-exposed occupations saw a 16 percent relative employment decline after generative AI adoption took hold. The New York Fed's recent-grad tracker put Q1 2026 unemployment for recent college graduates near 5.7 percent with underemployment hovering around 41.5 percent. The Harvard Business Review's coverage of AI and entry-level roles flagged the same pattern. And Bernard Marr's recent Forbes column made the obvious point that anyone running a business should already see: if the bottom rung breaks, the ladder collapses.

What the headlines miss is that these are not autonomous effects of the technology. They are choices. Companies decided to slow junior hiring because AI made the line item easier to defend. The technology did not pick up the phone and decline interview requests.

That distinction matters because it means the trend is reversible, and the companies that reverse it first will own a labor advantage in five years that competitors cannot copy.

The Math On Juniors Is Better Than It Looks

Let us put the actual numbers on the table.

The new grad is the cheapest full-time builder you can put on a team, and they are the only person in the room who learned to code with Copilot, Cursor, and Claude already in the editor. They do not have a workflow to defend. They are not slowed down by a 12 year habit of reading documentation linearly before writing a line of code. They open Claude Code, write a spec, and ship.

AWS CEO Matt Garman made this point bluntly in 2025 when he called replacing junior employees with AI "one of the dumbest things I've ever heard." Tim O'Reilly was even more direct in his essay Firing Junior Developers Is Indeed the Dumbest Thing, arguing that the cost calculus only works if you assume the company will not need senior engineers in 2030.

That is not an assumption you should make.

What juniors actually give you, on top of the work

• AI fluency at zero retraining cost. They use Claude and Cursor the way you used Google. You do not have to run a six-month enablement program.
• No legacy workflow attachment. They will rewrite your runbook because they have not memorized it yet.
• Fewer fixed costs. They tend to have fewer family obligations and can take on the irregular hours of an incident, a customer escalation, or a launch week.
• Learning velocity. A new grad in their first 18 months absorbs more about your systems than a contractor will in five years.
• Senior bench in 2031. This is the one nobody priced in.

A senior engineer is not a free resource the labor market spawns on demand. They are made, slowly, by a junior engineer who joined your company, broke production once, learned why the system is the way it is, mentored the next class, and stayed.

What AI Actually Does To Entry-Level Work

The mistake is treating "junior work" and "AI-replaceable work" as the same thing. They overlap, but they are not identical.

A junior engineer's job in 2018 was: read the ticket, find the file, write the code, write the test, open the PR. Generative AI is genuinely good at the middle three steps. It is bad at the first and last in any system more complex than a tutorial app, because reading a ticket correctly and getting a PR through review both require context the model does not have.

So the real question is not "can AI do what a junior does?" The question is "what is the new junior job description in a company that uses AI well?"

The answer, in the teams I have seen ship effectively:

• Eval writing. Juniors write the test cases and golden datasets that catch regressions in agent behavior. This is the most undervalued skill on an AI team and the highest-leverage place a new hire can spend their first six months.
• Output review and labeling. Someone needs to look at thousands of model outputs and decide which are good. This is hard, repetitive, and exactly the kind of work a sharp, motivated junior will do better than a distracted senior.
• Workflow mapping. Before you automate a process, somebody has to document what the process actually is. Juniors are good at this because they ask the dumb questions seniors stopped asking years ago.
• Test harness construction. Building the scaffolding that lets the team run an agent against real data, capture traces, and replay failures is greenfield work and a fast path to systems thinking.
• Customer context capture. Sitting on sales calls, reading support tickets, and writing down what users actually do is hugely valuable and almost nobody senior wants to do it.
• Tool-assisted execution. Pairing with an AI to ship small, well-scoped features under a senior's review. This is the modern apprenticeship.

Hamel Husain has been writing for years that error analysis is the most important activity in LLM development. Most teams skip it because it is unglamorous. A junior with a notebook, a CSV of model outputs, and a week of focused attention will find more real product bugs than a senior architect will in a month.

That is not entry-level work in the old sense. It is critical-path work, and it is exactly the work AI cannot do for you.

The Three to Five Year Trap

Here is the timeline most companies cutting juniors are not modeling.

Year 1. You save roughly $1.5M in fully-loaded comp by skipping a class of ten new grads. Margins look better. Board is happy.

Year 2. A few mid-level engineers leave for competitors who are hiring. You backfill with contractors. Velocity dips but is masked by AI productivity gains.

Year 3. Senior attrition accelerates because the people who would have been promoted into senior roles do not exist. You start paying recruiters 30 percent placement fees on mid-level hires. The contractors do not know your codebase, so onboarding takes longer than it used to. Documentation rots because nobody is incentivized to keep it current.

Year 4. You realize your remaining seniors are spending most of their time reviewing AI-generated code from contractors who do not understand the system. They are unhappy. Two more leave. You try to hire senior engineers from the open market and discover everyone else made the same cut and there are no seniors with five to seven years of relevant experience because nobody hired them as juniors.

Year 5. You restart junior hiring at twice the original budget because you have to compete with the companies that never stopped. The institutional knowledge gap takes another three years to close.

This is not a hypothetical. The AP's coverage of the post-pandemic hiring whiplash showed companies running exactly this pattern in adjacent labor markets. The AI version will be worse because the gap will be wider and the institutional knowledge harder to reconstruct.

Goldman Sachs' analysis of how AI will affect the US labor market makes a similar point in more measured language: the productivity gains are real, but the firms that capture them will be the ones who redesign work, not the ones who simply remove headcount.

What To Cut Instead

If the goal is cost reduction, AI does give you legitimate places to cut. They are just not where most companies are looking.

• Duplicated coordination roles. Project managers whose entire job is moving information between two Slack channels. AI is genuinely good at this and humans hate doing it.
• Mid-tier middle management. Layers of people whose primary output is status reports. Compress this aggressively.
• Contractor spend on undifferentiated work. If you are paying offshore agencies for tasks an AI agent can complete with review, that line item should shrink fast.
• Process overhead. Meetings, recurring reports, and committee reviews that exist because nobody had time to automate them. Build the agent. Cancel the meeting.

Notice what is missing from that list. New grads. Customer-facing roles that build context. Anyone whose job involves learning the system they will eventually own.

The pattern is consistent across the better research on this: Nexford's labor market analysis and Johns Hopkins' Hub coverage on whether AI will make human workers obsolete both land in the same place. AI compresses the middle of the org chart more than the bottom or the top. Companies that cut the bottom are reading the curve wrong.

A Better Playbook For 2026 Hiring

If you are running a team right now, the practical version of this looks like:

1. Hire fewer juniors, but hire them. If you used to hire ten new grads a year, hire five. Do not hire zero. The bench math matters more than the savings.
2. Rewrite the junior job description. Eval writing, output review, workflow documentation, customer context, and tool-assisted execution should be the first 18 months. Not ticket grinding.
3. Pair every junior with a senior on a real system. The senior reviews the work, the AI accelerates the work, and the junior absorbs the context. This is the modern apprenticeship and it is the only way to build a 2031 senior bench.
4. Measure junior productivity differently. Lines of code is the wrong metric. Eval coverage, documented workflows, customer insights captured, regressions caught: these are the right ones.
5. Cut the actual fat. Coordination overhead, redundant middle-management layers, undifferentiated contractor work. That is where AI savings live without breaking your pipeline.

Companies that do this will have a measurable labor cost advantage in 2030. Companies that do not will be paying $500k for engineers who used to cost $250k, and they will spend the next decade explaining to investors why their margins keep slipping.

How OpenNash Can Help

Most of the conversations we have with operations leaders start with "where can AI cut headcount?" and end with "where can AI absorb the work nobody on the team has time for?" That is usually the better question. We help teams audit their actual workflows, identify the coordination and execution work that genuinely belongs to an agent, and design the human review layers that keep the system honest. The deliverable is a production system the client owns, plus a clear map of which roles to grow into AI oversight work and which roles to leave alone.

If you are about to make a junior hiring cut, book a call and walk us through the numbers first. We can usually find the same savings in process overhead instead, and you keep your 2031 bench intact.

The companies that win the next decade will not be the ones that automated the most jobs. They will be the ones that figured out which jobs to keep humans in, and built the AI infrastructure that made those humans 5x more valuable.