The Last Generation of Seniors: Why Schools Must Stop Producing Juniors

The article argues that AI has replaced the routine work of junior programmers while simultaneously heightening the value of seniors' tacit experience — and in doing so has severed the chain by which juniors gradually mature into seniors. The author ties together demographic data, meta-analyses of older workers' performance, and the COBOL paradox (220 billion lines of code maintained by aging programmers) to the thesis that companies are both ceasing to hire juniors and laying off seniors, thereby endangering an entire layer of expertise. In closing, he proposes that schools stop producing hundreds of identical graduates and start distinguishing future engineers from operators, and that the state change how it funds universities.

Imagine two programmers. Both sit at the same editor, both have access to the same AI assistant. The first has thirty years of experience, knows the system's history, remembers three major migrations and two late-night incidents that nearly took down production. The second is two years out of school and can prompt faster than they can write.

The AI assistant generates the same code for both. The senior reads it, frowns, and says: "This is going to blow up on us in a year, because there's a race condition here that we dealt with back in 2019." The junior approves the code because it looks correct — and they have no database of experience by which to recognize that it isn't.

For decades we believed that the tech industry mainly needs young people. Fast, cheap, adaptable. Seniors approaching retirement we regarded as an expensive luxury a company keeps on out of inertia. Artificial intelligence has turned this assumption on its head. AI replaced what the junior could do — and highlighted what only the senior can do. But seniors are dwindling. And the schools that should be training their successors instead churn out, year after year, hundreds of graduates into positions that no longer exist.

## Why right now

Three curves intersect at a single point. First, the demographic one: the EU's workforce will shrink dramatically by 2050, while the share of seniors in the population will grow at a pace unmatched in modern history. Second, the technological one: junior programming positions have dwindled dramatically over the past two years, and large tech firms are halting the hiring of new engineers, citing AI productivity. Third, the knowledge one: the programmers of systems that process a critical share of the world's commerce every day are retiring, and no one is replacing them.

What is at stake is not just corporate efficiency. At stake is the entire knowledge chain by which juniors gradually become seniors. If AI eliminates the junior layer and schools fail to adapt what they produce, then in twenty years we will have a generation of AI operators who can prompt but who have never debugged a memory leak at three in the morning.

## The gold inside the company

They've been everywhere, they know everything, and they've lived through everything at least once. That's how you might describe the typical fifty-something with twenty or more years of experience. Research consistently shows that these people are a strategic asset — not a burden.

How much does losing such a person cost? A study by the Center for American Progress analyzed 30 cases and put the median replacement cost at roughly 20% of annual salary. For highly skilled and leadership positions, the cost climbs to as much as 213% of annual salary. For a lead specialist with a salary in the millions of korunas, these are sums any CFO would rather invest elsewhere. And those are only the measurable items — recruitment, onboarding, lost productivity while a newcomer gets up to speed.

The real loss lies in what Dorothy Leonard of Harvard Business School called deep smarts: profound knowledge built over decades of practice, encompassing pattern recognition, intuitive judgment, and systems thinking. It is inherently tacit — it cannot be written into documentation, because the person who holds it is often not even fully aware of it. Michael Polanyi put it succinctly: we know more than we can tell. And what we cannot tell, we cannot write into a prompt for an AI either.

In his research on expert firefighters, Gary Klein described how experienced professionals make decisions. They do not analytically compare options. They use pattern recognition — comparing the current situation with prototypical past experiences. Across 156 analyzed decision moments, analytical comparison of options appeared in fewer than 12% of cases. In the vast majority of situations, commanders identified a solution immediately, on the basis of experience.

This is exactly what a senior programmer does when they say "this is going to blow up" — they don't know it from a manual, they know it from thousands of situations they have lived through.

The most dramatic illustration is NASA's case. When the agency planned a return to the Moon, it discovered that the key engineers from the Apollo program had retired and that critical knowledge — including the manufacturing procedures for the Saturn V rocket — had been irretrievably lost. The investigation report after the Space Shuttle Columbia disaster concluded that years of layoffs and outsourcing had negatively affected the agency's knowledge base. Between 1993 and 1999, NASA cut its civil-servant workforce by 26%. With them went skills that cannot be replaced by a recruitment campaign.

## Data versus myths

The belief that productivity declines with age is one of the most widespread — and most thoroughly empirically refuted — myths in the labor market. The 2008 meta-analysis by Ng and Feldman, covering 380 studies, demonstrated that age is essentially unrelated to performance on core job tasks or to creativity. On the contrary, age correlated positively with citizenship behavior, safety, and reliability, and negatively with absenteeism and counterproductive behavior.

The key moderator was uncovered by Sturman in 2003: job complexity. For simple tasks, performance declines with age over time. For highly complex tasks — and programming undoubtedly belongs among them — experience predicts performance in a linearly positive way. The more experience, the better the performance. With no ceiling.

How is this possible? Cognitive psychology offers an elegant explanation. Fluid intelligence — processing speed, working memory — does indeed decline from middle age. But crystallized intelligence — vocabulary, accumulated expertise, semantic knowledge — grows into one's sixties. The Seattle Longitudinal Study, which followed roughly 6,000 participants, found that middle-aged adults outperform their younger selves on four of six cognitive tasks. In 2013, Li and colleagues formulated and tested the compensatory-abilities hypothesis: the higher crystallized intelligence of older adults provides an alternative path to good decisions that compensates for lower fluid intelligence.

In a 2012 large-scale meta-analysis of 418 studies covering more than 208,000 workers, Ng and Feldman tested six common stereotypes about older employees — less motivated, unwilling to train, resistant to change, less trusting, less healthy, worse work-life balance. The empirical data refuted five of the six stereotypes.

And loyalty? A 2010 meta-analysis by the same pair of authors, drawing on more than 800 studies, demonstrated that age is positively associated with job satisfaction, intrinsic motivation, organizational commitment, and loyalty. Older workers are more engaged — a study by AARP and Aon Hewitt found that 65% of workers over 55 are fully engaged, compared with 58–60% in younger groups.

Loyalty, however, has its price. When a company offends a senior — leaving them out of training, ceasing to give them interesting tasks, signaling that it regards them as a relic — the reaction is harsher than managers expect. Armstrong-Stassen and Ursel demonstrated that the key factors in retaining older workers are targeted development and the assignment of challenging tasks. In a 2008 meta-analysis, Bal and colleagues found a paradoxical phenomenon: older workers generally regulate their emotions better, but when it is precisely job satisfaction that is violated — respect, recognition of expertise, meaningful work — they react more intensely than younger colleagues. They have a smaller field of alternatives and a larger investment in the relationship with the organization.

Put simply: a senior is loyal as long as they have a reason to be. But once a company breaks the psychological contract, it does not build that loyalty back.

## What AI actually replaces

When Salesforce announced that in 2025 it would hire no new software engineers, it was not saying it doesn't need programmers. It was saying that, thanks to a thirty-percent productivity gain from AI tools, it doesn't need to expand the team. The broader impact, however, falls disproportionately on juniors: according to aggregated labor-market data, the number of openings for junior developer positions fell by roughly 60% between 2022 and 2024.

The reason is simple. An AI assistant — Copilot, Claude, GPT — can generate functional code from a specification. Write a CRUD application, a unit test, a standard API endpoint. That is exactly what the junior used to do. Not because they did it better than the AI, but because the senior didn't want to do it. The junior was the execution layer.

Now the execution layer is AI. And a senior who knows what they want will use it more effectively than a junior who knows neither what they want nor how to recognize a bad output. A Google study confirmed that senior developers record the most significant productivity gains from AI tools. That makes sense: AI is a multiplier, not a generator. It multiplies what you already know.

A junior with AI can write code quickly. But they can't spot a subtle race condition. They can't decide whether the architecture scales. They won't say "we tried this in 2015 and it failed because the client changed the requirements after the third iteration." They don't understand the business context behind the technical specification. And above all: they don't have Klein's database of thousands of situations from which the expert draws in the overwhelming majority of decision moments.

The traditional ratio of senior to junior developers on a team — one senior to three to five juniors — is being reconsidered. Organizations are finding that they need more experienced engineers able to review AI-generated code, maintain architectural quality, and make decisions AI cannot. Not fewer.

The COBOL paradox illustrates this trend with surgical precision. There are still at least 220 billion lines of COBOL code active in the world. Three trillion dollars of commerce flows through these systems every day. 95% of ATM transactions, 80% of in-person banking operations, 43% of banking systems, including JPMorgan and Bank of America. The average age of a COBOL programmer, according to Reuters, is 58. More than 85% of universities have dropped COBOL from their curricula since the 1990s.

AI can read and translate COBOL. But it doesn't understand why that code looks the way it does — what business rule from 1987 lies behind it, what edge case it handles, why that seemingly pointless check calculation is there. Only that fifty-something programmer knows. When they leave, the answer leaves with them.

When, in April 2020, the unemployment systems in New Jersey, Kansas, Connecticut, and other states collapsed under the onslaught of a 1,600% surge in claims during the pandemic, Governor Phil Murphy had to publicly appeal for volunteers with COBOL knowledge — mostly people of retirement age — to come back and help fix the systems. Modernizing one such system costs hundreds of millions of dollars. The Commonwealth Bank of Australia completed its migration off COBOL in 5 years, at a total cost of nearly 750 million Australian dollars.

## The broken pipeline

Here we arrive at the core of the problem, which is graver than anything above. Where will the future seniors come from?

Tacit knowledge is not built by reading or studying. It is built by lived experience. Klein's 88% of intuitive decisions arise from thousands of real situations — production incidents, architectural dead ends, communication failures with clients, late-night deployments that went sideways. From all the things no school will give you and no documentation will capture.

The existing model was simple: a school produces a junior, a company lets them work alongside a senior, the junior gradually absorbs tacit knowledge, and in 10 to 15 years they become a senior. Slow, expensive, but it worked.

AI breaks this model from two directions. From below: junior positions are dwindling, so young programmers have nowhere to gather experience. From above: companies lay off seniors because they think AI will replace them — and in doing so they lose the mentors who would have raised the juniors.

The result? A generation of people who spent their entire careers prompting AI but never debugged a production system under pressure. They will know how to generate code, but they won't know how to make decisions. And decision-making — not coding — is what companies pay seniors for.

It's the same problem as COBOL, just on a different scale. Today we have seniors who know. But we are systematically failing to create the conditions for more of them to emerge. Except this time it isn't one programming language — it's an entire layer of expertise.

## The school must change its product

The solution does not begin in companies. It begins in schools. And it requires an uncomfortable admission: schools today produce graduates for positions that won't exist in five years.

The numbers first. The whole of Czechia produces roughly 3,600 university-educated ICT specialists a year — and that ceiling has held practically unchanged since 2008, despite a rising number of students. Where do those students disappear to? The dropout rate in ICT programs at Czech universities reaches 41% — the highest of all fields, along with agriculture and services. To FIT ČVUT, the most sought-after computer-science faculty in the country, more than 3,000 applicants apply each year, about 900 enroll, and over the faculty's entire thirteen-year existence it has produced fewer than 3,600 graduates — that is, roughly 270 bachelor's and master's graduates a year. Only a fraction of those go on to a master's degree.

Take that master's cohort — say 200 students who actually reach the degree. They all get the same curriculum, the same diploma, the same assumptions. The school pretends that 200 equals 200. But the market knows it isn't so.

Of those 200, how many have genuine intrinsic motivation, the capacity for independent work, and intellectual curiosity? The teachers know exactly — they are the same people who come to office hours, write interesting theses, seek out projects on their own. Say 20. Maybe 30. Certainly not 200.

And now the key question: how many quality research internships and corporate internships are available? A quality research internship requires a mentor who has time and motivation. A real research problem, not make-work. Infrastructure. Funding. There are dozens of such places across all of ČVUT. Abroad, a few dozen more for Czech students. Not hundreds. And this is the heart of the capacity argument: the question is not how many engineers the market needs — it's how many the school can produce well. The limiting factor is not demand, but the supply of quality internships and quality students.

The school has no choice. Either it pretends all students are the same, gives them the same diploma, and lets the market sort them — painfully, expensively, with a two-year lag. Or it admits reality: it has the capacity for 20 engineers and 180 people with a different profile. And it gives both groups an education matching what they will actually do.

What would a graduate of the first group look like? A master's holder with two stays abroad — one academic, one in research — and two half-year internships in companies. Two years of real practice in total, before the degree. Not a senior, but a junior with a senior's foundation. Someone who arrives at their first job with experience from three different environments, who has already read someone else's code, taken part in a code review, seen a production deploy, worked on an international team.

That person becomes a senior in 3 to 5 years instead of 10 to 15. Because they don't arrive with an empty database of patterns — they arrive with their first hundred experiences, onto which the company stacks more.

But here is what is essential: the school must deliver what the company cannot. That's why the emphasis on research internships, not just corporate ones. A company teaches corporate processes. But scientific thinking — formulating hypotheses, systematic experimentation, critical reading of the literature, working with uncertainty — that a company will not supply. And this is precisely what distinguishes an engineer from an operator. An engineer can solve problems that don't yet have an answer on Stack Overflow.

And the 180? They don't need a five-year master's. They need a practical program focused on AI-augmented development, testing, DevOps, data analysis — roles where AI is a tool, not a replacement, but where you don't need research thinking. Not a worse education. A different one.

## The model exists — it's just aimed in the wrong direction

The paradox is that the university system already has a format for raising future seniors. It's called the PhD. Three to four years of deep work under a mentor's guidance. Solving problems without a known answer. Publishing results. A defense before a committee that verifies the candidate truly understands what they did. These are exactly the competencies that make a person a senior: the ability to work with uncertainty, to think rigorously, and to defend one's decisions.

But the PhD is aimed at producing academics, not senior engineers. A doctoral candidate spends four years optimizing one algorithm or one method. Industry needs a person who spends those four years solving different problems in different environments — but with the same depth and rigor. A master's holder with two stays abroad and two internships is essentially an industrial PhD without a dissertation. The same intensity, a different output. Instead of a paper — the ability to make decisions under uncertainty. Instead of a defense before a committee — the ability to defend an architectural decision before a team.

Why don't schools do this? First, funding: the PhD is paid from grants; the doctoral candidate is cheap labor for research. The industrial equivalent has no one to pay for it — the company wants a finished employee, the school has no money for an individualized program. Second, metrics: the PhD has a clear output — publications, a dissertation, a title. An industrial master's with internships is "just" a master's. How do you measure "is ready to be a senior in three years"? Third, prestige: in the academic system the PhD is a prestigious track; an industrial master's with internships is invisible. No one gets a grant for it, no one writes a habilitation on it.

And yet it is the same principle: prolonged, intensive preparation under the guidance of a more experienced person, building a repertoire of recognized situations before the graduate enters practice. Only the goal is different — not an academic career, but the ability to manage complex technical systems.

## The Czech context

The Czech Republic stands in an interesting starting position. The employment rate among those aged 50 to 59 reaches 91.4% — the highest in the entire OECD, where the average is 75.7%. Czechia has the lowest unemployment in the EU, at 2.6%. People in their fifties are working, they are productive, and companies need them.

But at sixty comes a sharp cliff. Employment among those aged 60 to 64 drops to 58.4% — a fall of 33 percentage points. One of the most dramatic in Europe. Unlike Sweden or the Netherlands, Czechia does not offer older workers flexible arrangements: 70% of workers aged 55 to 64 work full-time, 40 to 44 hours a week.

The pension reform of December 2024 sets a gradual increase in the retirement age to 67. That means we will need older workers for longer — but we don't have the infrastructure to let them work differently. Reduced hours, mentoring, knowledge transfer — none of this is standard in the Czech work environment.

In a recent report specifically on the Czech Republic, the OECD pointed to a problem of career mobility: at age 26, 10% of workers change jobs annually; at age 60, just 1.8%. Older workers in the CR don't move — they either stay in the same place or leave entirely. The room for intergenerational knowledge transfer is minimal.

And the Czech schools? ČVUT FIT, Matfyz, FI MUNI — all produce graduates with the same diploma, as we showed above. None of the faculties systematically distinguishes between future engineers and future operators. Stays abroad are an optional Erasmus, not a mandatory part of the curriculum. Corporate internships are up to the student. Research internships are for those who find them on their own.

Yet Czechia has a specific advantage: a smaller system, where change is easier than in France or Germany. One faculty, one dean, one decision — and 20 students a year can have a fundamentally different program from the other 180.

## Where the truth is more complicated

The strongest counterargument goes: AI plus a junior handles 80% of a senior's work at a fraction of the cost. And in some domains that holds. A routine CRUD application, a standard data pipeline, a simple web service — here a junior developer with a good AI assistant really does deliver a comparable result. Not every company is solving architecture for a million users.

The second serious counterargument: a school cannot predict what the market will need in five years. How many companies in 2020 had any idea that within two years junior positions would be vanishing? The "20 engineers and 180 operators" model assumes we know what the ratio will be. We don't.

The third: it's elitism. To tell a twenty-year-old student "you are a future operator" is socially toxic and possibly factually wrong as well — talent sometimes shows itself late.

There are answers to these objections, but incomplete ones. To the first: yes, 80% of the work is routine. But those remaining 20% are the moments where millions are decided — architectural decisions, crisis situations, the recognition of failure patterns. A company that has no seniors for that 20% has no one who would recognize that the 80% was done wrong.

To the second: the school doesn't have to predict the exact ratio. It only has to admit that one size does not fit all, and offer differentiation. A two-tier model — a common foundation, then selection after the second year based on results and motivation — reduces the risk of a bad estimate.

To the third: differentiation is not discrimination. It is a recognition of the reality that different people need different things. And it is fairer than today's system, which gives everyone the same diploma and lets them discover in the job market that half of them are unsellable.

But a deeper problem remains, one the counterarguments correctly identify: funding. Czech universities are paid per head. Eighty students with an expensive program — stays abroad, research internships, individual mentoring — versus eight hundred with a cheap one. The rectorate will not back this model until the funding system for universities changes. And that is a political decision that transcends any single faculty.

## Before it's too late

That senior from the start of the article — thirty years of experience, three migrations, two late-night incidents — will retire in ten years. And with them will go a database of thousands of decision situations that no AI model has in its training data. Because they never happened on the internet. They happened in a production environment, at three in the morning, with no documentation.

The demographic curve is irreversible. The old-age dependency ratio in the EU will nearly double by 2050. In the Czech Republic the pension age will be 67, but the infrastructure for seniors to work does not exist.

Companies must stop discarding seniors and start placing them in the role of mentors. Schools must stop producing hundreds of identical graduates and start distinguishing between future engineers and future operators. The state must change university funding so that a high-quality, individualized education is not economically untenable.

And above all: everyone must stop pretending that AI solves the problem of older workers. AI deepens it. It raises the value of experience — and, at the same time, destroys the path by which experience comes into being. If we don't start acting now, the last generation of seniors will be the very one sitting today in open-plan offices, waiting to see whether the company lays them off or values them.

In twenty years we will know what we chose.

The article combines three types of sources. Meta-analyses from organizational psychology (Ng and Feldman 2008, 2010, 2012; Bal et al. 2008; Armstrong-Stassen and Ursel 2009) provide the strongest evidence on the work attitudes and performance of older employees. Cognitive-psychology research (Li et al. 2013; Klein 1985/2010; Schaie 1996, 2005) clarifies the mechanisms of expert decision-making and the compensation of cognitive decline. Macroeconomic and demographic data (OECD Employment Outlook 2025; EU Ageing Report 2024; Eurostat) place the problem in the context of an aging European population.

The research data draw on meta-analyses published in peer-reviewed journals (Personnel Psychology, Journal of Applied Psychology, Journal of Vocational Behavior, Psychology and Aging), reports from the OECD Employment Outlook 2025, Eurostat, the AARP/Aon Hewitt study (2015), and other sources cited in the text.

Limitations: The data on the decline in junior positions (~60%) come from aggregated labor-market surveys (job postings), not from longitudinal employment studies — the actual decline in employment among young developers (aged 22–25) is, according to the Stanford Digital Economy Lab, roughly 20% from the 2022 peak. The 20/180 ratio in the discussion of education is an analytical estimate based on the capacity argument, not an empirical measurement. The number of ICT graduates (~3,600/year) comes from MŠMT data (SIMS) cited by BusinessInfo.cz (2021); more recent figures from the ČSÚ for 2023/2024 had not been verified against the primary table at the time of writing. The 41% dropout rate in ICT fields comes from the EU Education and Training Monitor 2025 (Czechia country report). The data on FIT ČVUT come from the faculty's annual report for 2024. The Czech context is reconstructed from the OECD report specifically on the CR (2025) and Eurostat data; primary Czech research on age management in the IT sector is scarce. The replacement costs from the CAP study (2012) represent a median of 20% of annual salary for ordinary positions; the upper bound of 213% applies exclusively to executive and highly specialized roles. The Center for American Progress study analyzed the U.S. market — direct transfer to the Czech setting requires caution.

The article's concept, structure, and editorial line are the work of the author, who prepared the content outline, set the key theses, and directed the entire creative process. Generative AI (Claude, Anthropic) was used as a tool for research, fact-checking, and elaborating the author's draft.

The author edited the outputs throughout, verified the key findings, and approved the final version. No part of the text was published without human review. All factual data were verified against the publicly available sources cited in the text.

The process complies with the transparency requirements of Art. 50 of EU Regulation 2024/1689 (AI Act) for AI-generated content. #poweredByAI