She joined the team seven months ago, right out of university. Smart, motivated, technically capable. Her PRs merged on the first or second try. Her velocity numbers were solid. Her manager called her “a natural.”
Last month, there was a production incident. A payment processing service started timing out under load. The team needed someone to trace the failure through three services, understand why the circuit breaker configuration was set the way it was, and determine whether the root cause was a recent change or a latent design flaw.
She could not do it. Not because she lacked intelligence or motivation, but because she had never built the mental model. For seven months, she had generated code with AI assistance, merged it after review, and moved on. The code worked. But she had never written it wrong, debugged it by hand, traced a system failure through layers she built herself, or sat with a senior engineer while they reasoned through an architecture decision out loud. The tasks that would have forced her to develop that understanding were the first ones the AI handled for her.
Her manager told me: “She is productive. She ships. But she does not understand the system she is building.”
That sentence is the talent crisis nobody is budgeting for.
The numbers are worse than most people think
A Harvard working paper published in May 2026 analyzed 66 million workers across 280,000 US firms between 2015 and 2025. At companies that actively integrated generative AI into their workflows, entry-level hiring fell roughly 80% per quarter since 2023. The researchers, Hosseini and Lichtinger, call it “seniority-biased technological change.” Entry-level employment at these firms dropped approximately 9% within six quarters of AI adoption. Senior employment at the same firms continued to grow.
The mechanism is straightforward. Generative AI handles codified, checkable tasks best, exactly the kind of work learned from textbooks and training data rather than accumulated on the job. Those tasks define early careers. When companies automate them, they stop hiring the people who used to do them.
PwC’s 2026 Global AI Jobs Barometer, analyzing over one billion job advertisements worldwide, found something even more revealing. They call it “seniorization”: AI-exposed entry-level roles are now seven times more likely to require traditionally senior skills like judgment, leadership, and stakeholder management than they were in 2019. Those transformed roles grew 35% since 2019. Traditional entry-level openings shrank 10% in the same period.
The rung did not disappear. It moved three rungs up the ladder, out of reach for someone just starting out.
The counterargument worth taking seriously
SignalFire’s 2026 State of Talent Report, published in June 2026, tracked careers across 80 million companies and found engineering roles fell only 11% from 2019 levels while total tech hiring dropped 25%. Engineering is the most resilient job function. Engineers comprised 55% of all new hires at the twelve Tech Majors in 2025, up from 46% in 2019.
This data is real, and it matters. Companies are still hiring engineers. The question is which engineers, and what happens to the pipeline that produces them.
You can have a healthy senior engineering market today and a catastrophic talent shortage in five to eight years simultaneously. One does not prevent the other. It takes five to eight years to develop the systems-level judgment of a senior engineer. Disrupt that pipeline now, and you create a gap that becomes visible around 2030, when the organizations that stopped investing in junior development discover they cannot fill the senior roles they desperately need.
Gartner predicted exactly this in May 2026: 75% of organizations that pause entry-level hiring for AI will pay premiums upward of 15% for early-career professionals by 2030. You save on headcount now; you pay a premium later for the talent you did not develop.
The irony of automation, applied to engineering careers
Lisanne Bainbridge wrote “The Ironies of Automation” in 1983. The central insight: the more you automate a system, the less opportunity operators have to practice direct intervention, and the more difficult the remaining tasks become when they require human judgment. Gregor Hohpe, in 37 Things One Architect Knows About IT Transformation, summarizes five effects: automation handles the easy parts, leaving operators with the hard ones; autopilot reduces attention to system state; manual takeover is needed precisely when the system is already in an abnormal state; automation black-boxes the system, reducing learning opportunities; and small problems get masked until they become big ones.
Every one of those effects is now playing out in engineering careers.
AI handles the easy tasks that used to train junior engineers. The tasks that remain require judgment that only develops through years of doing the easy tasks first. When a junior engineer needs to intervene, the system is already in an abnormal state because the AI-generated code is opaque to the person being asked to debug it. And because AI masks small errors with functionally correct output, the problems accumulate silently until something breaks in production at 3 AM.
Microsoft’s Mark Russinovich and Scott Hanselman named the asymmetry in a 2026 ACM paper: the “Senior Boost” versus the “AI Drag.” For experienced engineers, AI tools act as massive productivity multipliers because they have the judgment to steer, verify, and integrate AI output. For early-in-career developers, the same tools create a drag, producing output that works but teaches nothing, bypassing the struggle that builds expertise.
The Anthropic Economic Index quantified the task-level impact: AI replaces work outright in more than 50% of interactions in entry-level occupations. For senior occupations, that figure drops to 40%. The tasks that train juniors are precisely the tasks AI absorbs first.
What apprenticeship used to look like
I have been building and managing technical teams for 28 years. In every one of those years, the junior engineers who became the strongest seniors shared something: they had done the work badly first. They wrote code that failed in production, traced the failure themselves, understood why their mental model was wrong, and rebuilt it. That cycle, repeated hundreds of times across different problem domains, is what produces a senior engineer.
The Missing README by Riccomini captures the traditional model well: shadow experienced engineers, pair program, front-load learning, experiment with side projects, ask questions, get feedback. The Center for Creative Leadership’s 70-20-10 model tells the same story: 70% of development happens through work experience, 20% through developmental relationships, 10% through formal training.
AI disrupted the 70% most. The work experiences that built competence were exactly the tasks an AI tool handles faster and better. Bug fixes, boilerplate code, simple feature implementations, debugging straightforward issues. Those are not glamorous. They are how you learn what the system does when you change something, how you develop the instinct for what will break before it breaks.
Peter Norvig put it precisely: “The key to developing expertise is deliberative practice: not just doing it again and again, but challenging yourself with a task that is just beyond your current ability, trying it, analyzing your performance while and after doing it, and correcting any mistakes.”
AI short-circuits that cycle. The output is correct on the first pass, which means the junior never experiences the failure that builds understanding.
What some organizations are doing
Not everyone is sleepwalking into this gap. The ones getting it right share a principle: intentionally preserve the learning mechanism even when the task can be automated.
Hohpe named this “predetermined breaking points” in his automation writing. Intentionally reduce the automation grade below what is technically possible, so that operators must solve less-than-critical issues regularly and remain prepared for the rare critical ones. Applied to engineering, this means deliberately assigning work to junior engineers that AI could handle faster, not for efficiency, but for development. The tool handles the task better. The human needs the struggle.
Deloitte built a full program around this principle. Their 2026 apprenticeship pilot focused on eight everyday activities: connecting, making expectations explicit, sharing knowledge, working side by side, making expert practice observable, using questions to surface thinking, creating safe practice repetitions, and calibrating through feedback. In a four-week pilot, the large majority of participants moved from “Developing” to “Proficient.” The program was explicitly designed for hybrid and AI-disrupted environments where proximity-based learning no longer works.
Others are finding their own paths. Revelry documented their approach to running a software apprenticeship in 2026: structured mentorship with deliberate exposure to full-system complexity, pairing juniors with seniors on real production work rather than AI-generated exercises. SAP and Wakefield Research found that 88% of CHROs say AI is making early-career talent role-ready faster, which frames the opportunity correctly: redesign entry-level roles around AI rather than eliminating them.
The common thread: the tool changes, but the need for people to develop judgment through practice does not.
The CTO question
This is a workforce planning problem, and it belongs on the CTO’s desk. Not HR’s, not the engineering manager’s, not the recruiter’s. HR optimizes for cost-per-hire and time-to-fill, metrics that actively reward eliminating entry-level roles. The CTO is the only person who sees the five-year talent supply chain.
The Four Things AI Cannot Replace in the People You Hire argued there are four qualities that matter more now, not less: excellence, teamwork, empowerment, and curiosity. Those qualities are not innate. They develop through years of practice within a team, under mentorship, through the experience of building something, watching it fail, and learning from the failure. AI can accelerate that development if structured correctly. AI can also eliminate the conditions that make it possible.
The CTO’s Real Job When AI Changes How Software Gets Built made the case that protecting institutional knowledge during AI transitions is a CTO responsibility. This is the most concrete version of that argument. The institutional knowledge of 2030 is being built, or not built, in the decisions you make about entry-level hiring this quarter.
The data from the Harvard paper showed that much of the decline is not driven by layoffs. Workers who left simply were not replaced, separation rates fell at adopting firms, and the pipeline erosion is so quiet it looks like efficiency on a quarterly report.
The question I cannot answer yet
I have built teams and helped customers build theirs across dozens of organizations. I have hired hundreds of junior engineers and watched them become the senior engineers that carry organizations. The cycle always worked because the entry point existed: messy, unglamorous tasks that forced new engineers to build the understanding that eventually made them indispensable.
That entry point is narrowing. Not vanishing entirely, but narrowing enough that the pipeline produces fewer senior engineers per cohort than it used to. The organizations that recognize this and invest in structured apprenticeship will have a meaningful talent advantage by 2030. The ones that optimize purely for this quarter’s productivity metrics will discover the gap when it is too late to close it.
I do not have a clean framework for what apprenticeship looks like when AI handles the tasks that used to be the training ground. Deloitte’s model points in a direction. Hohpe’s predetermined breaking points offer a principle. In my own teams, I apply it directly: there are tasks AI could handle faster that I deliberately assign to junior engineers because the struggle is the point, not the output. The practitioners running apprenticeships in 2026 are figuring it out in real time.
What I do know: every senior engineer you have today was once a junior who learned by doing the work badly first. If you automate that work away without replacing the learning mechanism, you are eating your seed corn. The harvest looks fine this year. Five years from now, you will be bidding against every other organization that made the same choice, competing for a shrinking pool of people with the judgment that only develops through years of practice.
Ricardo
