Preparing for the post AI order
Executive Summary
The software development workforce is approaching a structural crisis that is not yet visible in today’s hiring data, but is clearly legible in the demographic and educational trends accumulating beneath it. This paper argues that the current wave of junior developer redundancies, driven largely by anticipated productivity gains from artificial intelligence tools, represents a significant strategic miscalculation by organisations in both the private and public sectors. A miscalculation whose consequences will be felt most acutely between 2028 and forward, when the costs of reversing it will be substantially higher than the costs of avoiding it today.
Four converging factors underpin this assessment.
AI tools are transforming the developer role from code writer to code architect, and doing so faster than organisations can adapt their workforce structures to accommodate the change. The measured productivity gains from AI in real production environments are, at present, significantly lower than executive perception suggests. The most rigorous independent study to date (Joel Becker, 2025) found that experienced developers using frontier AI tools in 2025 took 19% longer to complete tasks than without them, against a developer-perceived improvement of 20%. Organisational hiring decisions are being made at scale on the basis of anticipated gains that have not yet materialised at the level assumed.
The second factor is the rapid contraction of the developer education and training pipeline, as students and institutions respond rationally to the signal that entry-level employment has collapsed. US entry-level technology job postings fell 67% between 2023 and 2024, and computer science program enrollment deposits have declined by over 25% in the most recent academic cycle.
The third factor is the accelerating demographic ageing of the existing professional developer workforce.The proportion of professional developers aged 35 and above has grown from 31% to 39% in just three years, while the under-25 cohort has shrunk from 33% to 23% over the same period. The aging of the workforce and the subsequent large drop in new engineers is a clear sign that the system is becoming top heavy. Coupling this with the higher levels of age discrimination, and ageism beginning earlier, in software development, it creates a scenario where the industry loses its most skilled practitioners at precisely the point it can afford to the least.
The fourth consideration is a significant and well-documented deterioration in the quality of the developer experience, a measure of developer’s work environment effectiveness and job satisfaction. Research shows this is accelerating attrition at the senior level precisely when retention of senior capability is most critical.
These four factors are not independent. They form a compounding system in which each accelerates the others, and in which the window for low-cost intervention is closing. The developers currently displaced by junior hiring freezes will not remain available indefinitely. Research consistently shows that when entry-level professionals cannot gain a foothold in their trained field within a reasonable period, a meaningful proportion redirect their careers and employment permanently into adjacent sectors. Once lost to the pipeline, they do not return.
There is a historical parallel that is instructive. The COBOL developer shortage, now costing financial and government institutions extraordinary sums in retention and recruitment, was the direct consequence of a previous generation’s failure to invest in succession. That shortage affected a single ageing language. The emerging shortage affects the foundational layer of all software capability, at a moment when AI implementation itself demands the most experienced engineering judgement available.
The recommendation of this paper is straightforward, though counterintuitive in the current business climate. Organisations that continue to eliminate junior development roles in favour of a leaner, AI-augmented senior workforce are optimising for a present that is already changing, at the cost of a future in which they will find themselves competing in an open market for a talent cohort that was never adequately formed. The more resilient path is to recommit to building developer capability in-house, treating junior developers not as a cost to be eliminated but as the senior engineers and technical architects of the next decade. The organisations that maintain this discipline now, both private sector firms and public bodies responsible for critical digital infrastructure, will be the ones best positioned to capture the genuine long-term productivity gains that AI will eventually deliver, because they will have the experienced human judgement required to direct it.
A compounding set of forces
Generative AI and in particular the use of Agentic AI is changing the landscape of Software Development at an unprecedented pace. We have found in our research that it is the area with the highest development rate as well as the area of AI with the most adoption. While the jury is still out on the effectiveness overall of the AI tools and overall functionality, we already observe a direct effect of this adoption in the form of lower hiring and even a shift towards layoffs and force reduction driven by the anticipated gains of AI. While it feels an obvious choice for a business to reduce dependence on human workforce in favor of laborers that never sleep, the research points to a situation that businesses will be hard pressed to recover from.
Companies never practice the sort of foresight that researchers can, and in doing so they tend to embrace the here and now because the economics of their financial cycles are tied to immediate performance, not long-term sustainability. The AI arms race is changing that scenario however, and companies would do well to look beyond today and tomorrow at the long-term scenario they are creating. In short, a reduction of junior-level staff today will lead to a massive uptick in hiring costs later, if there is even anyone left to hire. The stats back up the race to the endgame condition we are in, and even during the data science boom of the 2010’s and COBOL developer competitiveness years into the early 2020’s, we did not see this much of an imbalance on the horizon. The data is stark, showing a relative workforce age increase of 5 years on average, coming in just the last 3 years.
The factors influencing this condition fall into 4 basic issues:
- The real impact of AI tools on developers and the mismatch of perceived and realized gains.
- A rapid drop-off in training and education of developers because of the fear of being unable to obtain employment post education.
- The “aging out” of the majority of senior developers – in an industry where real age discrimination begins at 35, the majority of developers are now pushing 40.
- The significant and precipitous decline in the quality of the developer experience, both the tools and platforms they use, and their working environment as a whole, for developers of all levels, and in all organizations,
The solution is obvious, but is counterintuitive when it comes to the current business climate of reducing headcount on the back of AI. This behavior is an anti-pattern when it comes to survival in the long-term, however. Our recommendation is returning to the somewhat antiquated notion of building your employee’s careers “in house”, and fostering your entry level people of today into the seniors you will need tomorrow. By doing this through a combination of a quality work environment and a reasonable career path, you can hedge against the increasing pressures of AI adoption, and the inevitable arms race that is brewing for talent that can handle the increasingly complex and critical implementation of AI tools.
Issue 1: The AI Productivity Question: Reality, Perception, and Consequence
Artificial intelligence tools have arrived in software development with genuine capability. That much is not in dispute. GitHub Copilot users in controlled studies completed isolated coding tasks up to 55% faster than peers working without assistance. Surveys consistently show that between 67% and 90% of developers report perceiving productivity improvements in their daily work, and by 2025 approximately 84% of professional developers were using AI tools regularly, with 51% doing so on a daily basis. The technology has moved from experimental to embedded in the development workflow faster than almost any preceding tool category, and organisations are right to take it seriously.
The problem is not that AI tools lack impact. The problem is the size of the gap between perceived impact and measured impact, and the consequential decisions being made in that gap.
The most methodologically rigorous study conducted to date on this question is the METR randomised controlled trial published in July 2025, examining how early-2025 AI tools affected experienced open-source developers working on real production codebases. The developers in the study are all experienced practitioners with an average of five years of active contribution to their specific repositories. Before the study participants estimated that AI tools would reduce task completion time by approximately 24%. After completing the study, participants estimated the tools had reduced their time by around 20%. The measured reality was that using AI tools increased their completion time by 19%. The tools, in this real-world context, made experienced developers slower, not faster.
This finding sits in deliberate tension with some of the more optimistic studies in the literature, and that tension is itself informative. Studies showing the largest productivity gains from AI, including figures of 56% and 65% improvement, were conducted using synthetic or simplified tasks specifically designed for automated testing. As the METR study authors note, such tasks are unrepresentative of most production software development work, and are likely to closely resemble training data used to develop the AI systems being evaluated, which creates a measurement advantage that does not exist in real codebases. A separate analysis of 800 developers by Uplevel found no significant gains in objective measurements such as cycle time or pull request throughput, and found that developers using AI coding assistants introduced a 41% increase in bugs, representing a measurable negative impact on code quality.
The code quality question extends beyond individual organisations into the open source infrastructure on which the entire profession depends. Open source projects underpin an enormous proportion of production software in both the private and public sectors, and they are maintained largely by volunteers working under significant time constraints. The arrival of AI-assisted development has introduced a new and serious strain on this infrastructure that has no clear precedent.
The most documented example is the curl project. Curl is a foundational internet data transfer tool embedded in billions of devices and as with many open source projects, it is maintained by a small team of volunteers. From early 2024 onwards, the project’s lead maintainer Daniel Stenberg documented a sustained increase in AI-generated vulnerability reports submitted via its bug bounty programme. By mid-2025, genuine vulnerability reports had declined from 15% of all submissions to approximately 5%, with the remaining 95% requiring triage time despite producing no actionable findings. In January 2026, the project shut down its bug bounty programme entirely. Stenberg described the volume of AI-generated submissions as effectively a denial-of-service attack on his maintainers’ time: a small number of trusted volunteers absorbing an unlimited volume of automated noise, each report superficially plausible enough to require human investigation before being dismissed.
Curl is not an isolated case. Seth Larson, a security triage worker for multiple open source projects, documented an equivalent pattern across the projects he managed, noting that AI-generated reports “appear at first glance to be potentially legitimate and thus require time to refute.” His assessment of the systemic consequence was direct: maintainers experiencing this burden become burnt out and increasingly averse to legitimate security work, with low-quality reports functioning in practice as if they were malicious regardless of intent.
Django updated its security documentation to reject AI-generated reports outright.
The Node.js project imposed minimum quality score thresholds to filter automated submissions.
The libxml2 maintainer ended support for embargoed vulnerability reports entirely in June 2025, citing unsustainable triage burden as an unpaid volunteer.
Jeff Geerling, who manages over 300 open source projects, reported that the problem had become serious enough that GitHub introduced a feature to disable pull requests entirely. This is a capability that, as he noted, is a striking response given that pull requests are the fundamental mechanism that made GitHub central to collaborative software development in the first place.
The open source signal-to-noise problem is directly related to the broader code quality argument. AI tools lower the cost of generating plausible-looking contributions without lowering the cost of evaluating them. That asymmetry is manageable when the humans producing AI-assisted contributions have sufficient understanding of the codebase to distinguish useful signal from noise before submission. It becomes structurally damaging when that understanding is absent. This is precisely the kind of contextual, codebase-specific understanding that junior developers acquire through structured development over time that is most difficult for AI tools to replicate and most consequential when it is missing. Currently, the data shows that entry-level developers are being displaced from formal employment. This is happening as AI tools make it easier to generate superficially convincing, but fundamentally uninformed contributions. Thus anyone hoping to contribute code can create code, whether it is objectively “good code” or not. This in turn causes a deterioration in the quality of the shared infrastructure that all software development ultimately depends upon.
The organisational picture is consistent with this ambiguity. McKinsey’s survey of C-suite executives conducted in late 2024 found that only 19% reported revenue increases of more than 5% from AI investments, with 36% reporting no change at all. Only 1% of organisational leaders described their companies as mature in AI deployment. A comprehensive Danish study linking AI chatbot adoption across eleven occupations to administrative labour market records found essentially zero measurable effects on earnings or recorded working hours through the end of 2024, despite widespread worker-reported adoption. The gap between what AI tools appear to do in controlled settings and what they deliver at organisational scale is not a rounding error — it is the central empirical question of the current moment in AI adoption.
None of this is an argument against AI in software development. It is an argument for proportionality. The Faros AI Productivity Paradox Report of 2025, drawing on real engineering metrics across organisations, found that even where team-level AI adoption drove measurable individual productivity gains, those gains were consistently absorbed by downstream bottlenecks in review, testing, and deployment pipelines. The organisation did not become more productive simply because individual developers touched more tasks. The gains require the surrounding system to be redesigned to capture them — a process that requires experienced engineering judgement to lead, and that takes time measured in years, not quarters.
This matters enormously in the context of the hiring decisions now being made. The Octopus Deploy 2026 AI Pulse Report documents that 73% of organisations have already reduced their number of junior developers over the past two years, with an explicit “seniors with AI” strategy emerging as the dominant model. This strategy rests on the assumption that experienced developers augmented by AI tools can sustain output whilst the organisation reduces headcount at the junior level. Given the evidence on actual versus perceived productivity gains, this assumption is being made at a significant scale on the basis of optimism rather than measured performance. The Brynjolfsson et al. analysis of high-frequency US payroll data through September 2025 found that early-career workers aged 22 to 25 in the most AI-exposed occupations show relative employment declines on the order of 15 to 16%, while senior employment remains largely stable — confirming that the adjustment is not distributed evenly but is concentrated precisely in the entry-level cohort.
The irony is sharp. Organisations are making irreversible structural decisions about their talent pipelines on the basis of productivity gains that, at the organisational level, have not yet materialised at the scale anticipated. The hype cycle around AI is real, well documented in the research literature, and directly traceable in the hiring data. The Danish study’s description of a “productivity J-curve”, in which organisational reorganisation and task restructuring delay measurable economic effects, suggests that genuine productivity gains from AI may indeed arrive, but will do so on a timescale that the current pace of junior hiring reductions is not designed to accommodate. By the time the gains are real and stable, the pipeline that would have produced the next generation of senior developers will have been disrupted for long enough that recovery will require competing in an open market for a cohort that was never properly formed.
It is worth also highlighting a hidden issue that will creep into organizations in the years ahead, as AI tools become the standard. Expertise in software engineering does not result simply from the accumulation of theoretical knowledge. In reality, it stems from the formation of “developer reflexes” born out of the repeated resolution of low-complexity difficulties. Historically, the tasks assigned to junior developers (writing unit tests, simple refactoring, first-level debugging or enhancements) served as an indispensable learning ground. It is by manipulating these fundamental building blocks that the human brain develops robust mental models, innovative strategies and, above all, technical intuition. Conversely, by systematically delegating these tasks to AI under the pretext of immediate time savings, organisations are inadvertently eliminating stages crucial to developing this technical instinct. We are witnessing a rupture in the learning chain: without the laborious practice of simple code, the capacity to supervise complex code withers.
As Nicholas Carr notes in The Glass Cage, automation converts tacit knowledge, meaning, the intuitive ‘feel’ for code built through manual debugging, into explicit procedures. This cuts out a key part of the learning process, leaving juniors able to generate code but unable to build the internal mental models required to oversee complex systems.
Automation creates a cognitive dependency trap. This phenomenon is analogous to the systematic use of GPS: by blindly following assisted directions to navigate one’s own city, one eventually loses the habit of memorising landmarks or constructing mental maps of journeys. Just as a driver becomes incapable of finding their way without assistance, the developer who delegates their logic to AI risks becoming lost when facing a complex system as soon as the tool is no longer able to guide them.
If a future senior developer has never learned to identify the subtle nuances of a logical bug for themselves because a machine did it in their place during their early years, they will find themselves unable to exercise the critical judgment necessary to validate the outputs of more advanced AI. We risk seeing the emergence of a generation of “surface-level seniors.” These will be “assisted experts”, certainly capable of piloting intelligent agents, but unfortunately devoid of the depth of analysis required when the AI fails, hallucinates, or drifts outside the specified framework. Expertise cannot be reduced to a supervisory role. It must be understood that if the supervisor does not possess sufficient technical mastery relative to the tool, they will quickly be overwhelmed. Ultimately, experience cannot be downloaded; on the contrary, it is forged through effort and concrete learning.
In the long term, this lack of direct confrontation with the raw material of code threatens the architectural innovation capacity of organisations. The understanding of large systems depends on an in-depth knowledge of their individual components. If the foundation of the teams expected to hold the skills and knowledge (including the ability to write code manually) is replaced by automated systems, however “intelligent” they may be, the overall coherence of the architecture becomes uncertain. Architectural decisions made by engineers who have never experienced the constraints of baseline implementation risks being disconnected from ground-level realities. It appears, therefore, indispensable to value the work of junior developers. The challenge is to preserve the integrity of the organisation’s global expertise and to ensure that the technical decision-makers of tomorrow possess the intellectual legitimacy (grounded in their lived experience as developers) necessary to direct increasingly opaque systems.
Issue 2: The Collapse of the Developer Formation Pipeline
Labour markets operate on signals. Prospective students, career changers, and educational institutions all make formation decisions based on their reading of where employment will be available, at what level, and at what compensation. When those signals turn negative, the pipeline does not slow gradually, it contracts sharply, and the contraction persists well beyond the point at which the underlying market conditions that caused it have begun to recover. The developer formation pipeline is currently receiving the most unambiguously negative set of signals it has experienced since the dot-com collapse of the early 2000s, and it is responding accordingly.
The proximate cause is the collapse in entry-level hiring documented in Section 1. US entry-level technology job postings fell 67% between 2023 and 2024, according to Stanford Digital Economy Lab analysis of ADP payroll data. The UK experienced a 46% decline over the same period. Employment among the youngest professional software developers sat approximately 20% below its late-2022 peak by the end of 2024. The consequences of these figures for the formation pipeline are not theoretical. They are already visible in enrolment data, institutional behaviour, and the accelerating closure of alternative development pathways.
The university pipeline is contracting at pace. MARKETview data, tracking the entire enrolment funnel across public, private, and selective national institutions in the US, recorded an 8% year-on-year decline in computer science programme deposit volumes in 2023-24, followed by a further decline of more than 25% in the most recent cycle. Application volumes fell sharply in 2024-25, and the yield, the proportion of admitted students who actually enrol, has declined steadily for two consecutive years. The Computing Research Association’s own pulse survey of 130 computing departments, conducted in autumn 2025, found that 64% of responding academic units reported declining bachelor’s degree enrolments in computing, with 40% reporting declines of up to 10%, a further 29% reporting declines of 11 to 20%, and 31% reporting declines of more than 20%. Five institutions reported declines exceeding 35%, with some losing as much as 75% of their undergraduate computing student body in a single year. The University of California system reported a 6% decline in computer science(CS) majors in 2025, down 9% over two years. This is the first sustained retreat in that system since the dot-com bust, and the first time in over twenty years that the number has fallen.
Critically, this decline is not occurring because students have lost interest in technology as a field. It is occurring because they have made a rational assessment that the traditional computer science degree no longer reliably produces the employment outcome it once promised. CS graduate unemployment reached 6.1% in 2025 against a general graduate unemployment rate of approximately 4%, placing computer science seventh highest among all college majors in unemployment — below philosophy, art history, and journalism. Federal Reserve Bank of New York data confirms the shift. Students are not abandoning technology; they are redirecting towards AI-specific programmes, data science, and cybersecurity, all of which carry more immediately legible employment signals. The US now offers 193 bachelor’s degree programmes in artificial intelligence and 310 AI master’s programmes, numbers that have grown rapidly as universities respond to student demand. This redirection will eventually produce a different kind of technology professional, but it will not produce the foundational software development capability that production systems require. Worth noting is that the results of this shift will not be seen for a decade at the earliest.
The alternative formation routes have fared considerably worse. The coding bootcamp sector, which at its peak in the mid-2010s provided a substantial and inclusive parallel pipeline into entry-level development roles, has undergone a structural collapse. Between December 2023 and mid-2024, more than a dozen prominent bootcamps closed their doors across North America, including Codeup, Kenzie Academy, Momentum Learning, Rithm School, Epicodus, Code Fellows, Women Who Code, and Toronto’s Juno College. Southern New Hampshire University shuttered its coding programme in 2023, citing AI adoption as a direct factor in declining placement rates. The most symbolically significant closure was 2U’s December 2024 decision to exit the bootcamp sector entirely, shutting down rather than selling a portfolio of partnerships with over 50 universities that had produced more than 96,000 graduates. This marks a decision that followed a 40% drop in bootcamp enrolment and a 23% revenue decline in its alternative credentials segment. The company’s $750 million acquisition of Trilogy Education in 2019 was effectively written to zero. These closures represent not merely business failures but the elimination of accessible, non-degree pathways that had served career changers and those without access to traditional university education, disproportionately affecting groups already underrepresented in the profession.
What makes this contraction structurally dangerous, as distinct from simply being a cyclical correction, is the timing mismatch between pipeline signals and market need. The students currently deciding not to pursue computer science degrees or bootcamp programmes are making decisions that will manifest as workforce absence in 2028 to 2032. The organisations currently eliminating junior roles will not begin to feel the downstream consequences of those decisions for a similar period. By the time the market signal reverses, as the BLS projection of 15% employment growth for software developers through 2034 suggests it will, the pipeline will have been suppressed for long enough that recovery will be limited. Degree programmes take three to four years to produce graduates. Experienced developers take five to ten years to mature from entry-level hires. Neither timeline accommodates the kind of short-cycle correction that organisations typically deploy in response to talent shortages. Salary competition and international recruitment, the standard tools of talent crisis management, will be available but at substantially elevated cost and without guaranteeing the depth of institutional and domain knowledge that internally developed talent provides.
There is a further multiplier that the headline enrolment figures do not fully capture. Research on prior talent pipeline disruptions consistently shows that a proportion of displaced early-career professionals who cannot gain a foothold in their trained field within two to three years do not return to it when conditions improve. They redirect permanently into adjacent fields where their skills transfer. Areas like financial technology, defence and security, healthcare informatics, and industrial automation among them, where compensation and stability are competitive. The 67% collapse in junior developer hiring between 2023 and 2024 therefore does not simply delay the entry of that cohort into professional software development by one or two years. It permanently removes a fraction of them from the profession’s future supply. The size of that fraction is not yet measurable, because the disruption is recent, but historical precedent from comparable labour market dislocations suggests it is not trivial.
Issue 3: The Ageing Out of the Senior Developer Cohort
Every profession has a demographic centre of gravity. This is the age band where the largest concentration of experienced practitioners sits, where institutional knowledge is deepest, and where the capacity to mentor, lead, and make consequential architectural decisions is most reliably found. For software development, that centre of gravity has been shifting steadily upwards for a decade, and is now moving at a pace that the broader workforce planning conversation has not yet caught up with.
The longitudinal data is unambiguous. Stack Overflow’s annual developer surveys, the most consistently sampled source in the field with responses from between 65,000 and 100,000 developers each year, show a profession that has aged measurably and continuously since at least 2015. In that year, the average developer age across respondents was approximately 29, and more than half of all developers were under 30. By 2016, analysis of the same survey noted that barely a quarter of practitioners had more than a decade of experience, leading commentators at the time to describe software development as a “field of newbies.” That description is no longer accurate. By 2018 and 2019, approximately three-quarters of professional developers were still under 35. By 2021, 48% of professional developers were concentrated in the 25 to 34 band. By 2022, 31% of all respondents were aged 35 and above. By 2023 that figure had grown to 35%, and by 2024 it had reached 39%. The 2025 survey confirms that 66% of professional developers now fall between 25 and 44, with the overall population measurably older than in any prior survey year.
SlashData’s global developer population data, drawn from surveys of over 11,000 developers across 126 countries, corroborates the Stack Overflow trajectory with a sharper lens on the most recent period. Between early 2022 and early 2025 alone, the proportion of developers aged 18 to 24 fell from 33% to 23% of the global professional population. That’s a ten percentage point drop in three years. Over the same period, the 35 to 44 cohort grew from 22% to 26%. The direction of travel is consistent across every major data source, and the rate of change has accelerated since 2022.
That demographic shift is unfolding within a profession that has historically skewed young, and the contrast with the broader workforce makes the trajectory more pronounced than the raw numbers alone suggest. The US occupational average for software developers currently sits at approximately 39 years, placing the profession at or near the median of the general workforce for the first time in its history. Globally the average remains younger, with the 25 to 34 bracket still constituting the largest cohort at around 32%, but the US figure is the relevant one for any organisation considering talent strategy in the world’s largest technology market. The Pave compensation management data cited by LeadDev is perhaps the most striking single data point in this regard: the average workforce age at large public tech firms rose from 34.3 years to 39.4 years between January 2023 and August 2025 alone, a shift of five years in under three real years. The junior hiring freeze has not merely slowed the pipeline, it has already produced a measurable and rapid demographic step-change in the workforce itself.
This ageing trajectory is directly at odds with a well-documented but underacknowledged dimension of the technology labour market: the effective age threshold at which developers experience discrimination in hiring is considerably lower than in almost any other professional field. The US Age Discrimination in Employment Act sets its legal threshold at 40, and the EEOC’s own reporting confirms that the proportion of tech workers over 40 declined from 56% to 52% between 2014 and 2022. This was during a period when that same cohort represented 53% of the overall US workforce, meaning tech had moved from over-representing the 40-plus group to under-representing it. AARP’s 2023 survey found that 64% of workers aged 40 and above reported seeing or personally experiencing age discrimination, the highest level since AARP began tracking the question.
The tech-specific threshold is, however, considerably earlier than the legal one. A 2021 study by the University of Gothenburg found that 35 is considered old by tech industry standards. The CWJobs survey of UK technology workers found that developers begin experiencing age-related discrimination at 29, a full twelve years earlier than the cross-industry average of 41. Further it found that by 38, colleagues consider a tech worker to have passed their peak. These are consistent findings across multiple geographies and multiple methodologies, and they carry a direct implication for the workforce ageing argument: the cohort most at risk of being structurally pushed out of the profession is precisely the 35 to 44 bracket that, as the SlashData survey shows, is now growing as a proportion of the overall developer population. The same demographic trend that is shifting the centre of gravity of the profession upward is simultaneously moving more developers into the age range where hiring discrimination becomes an active factor. The pipeline contraction described in Section 2 is not merely reducing the inflow of new talent, it is doing so at the moment when the outflow pressure on mid-career developers is also increasing.
This matters for two distinct but compounding reasons. The first is volume. The cohort that is now concentrated in the 35 to 44 band, the developers who entered the profession in the early to mid-2010s and who now constitute its most experienced working layer, will begin transitioning out of hands-on development roles in significant numbers between approximately 2030 and 2040. The US Bureau of Labor Statistics projects approximately 129,200 annual openings for software developers through to 2034, and explicitly notes that a significant portion of those openings will arise from the need to replace workers who transfer to other occupations or exit the labour force entirely. In 2025, a record 4.2 million Americans reached age 65, and approximately 10,000 Baby Boomers are estimated to be reaching retirement age every day through to 2027. The science and engineering workforce, of which software development forms a substantial part, shows labour force participation beginning to decline meaningfully between ages 55 and 60, with a marked reduction by 65. The retirement wave that has been forecast for years in the broader knowledge economy is not a future event. It is occurring now, and its heaviest impacts on the software development profession are approximately five to ten years away.
The second reason is less about volume and more about depth. Senior developers carry something that cannot be acquired quickly and cannot be transferred through documentation alone: the accumulated understanding of why systems are built the way they are, how architectural decisions made years ago interact with present requirements, and where the institutional knowledge of an organisation’s technical landscape actually lives. This is not a soft concern. It is a well-documented and quantifiable risk. Research consistently shows that fewer than 30% of organisations have a formal knowledge retention plan in place, and only 19% have formal succession plans at all. The consequence of that absence becomes visible at the moment experienced practitioners leave, and the knowledge gap is discovered rather than planned for. The NASA Apollo programme provides perhaps the most cited illustration of this dynamic: when engineers who had worked on the Saturn V rocket retired without structured knowledge transfer, the blueprints for the most powerful rocket ever built were effectively lost, and the capability had to be rebuilt from scratch at enormous cost decades later. The lesson was not learned at the organisational level. In 2025, the US Social Security Administration announced a three-year, one billion dollar AI-assisted programme to modernise its legacy COBOL codebase. It did this, not because the technology had failed, but because the human expertise required to maintain it was retiring faster than it could be replaced.
The COBOL situation warrants extended attention in this context because it is not an aberration. It is a preview. Approximately 220 billion lines of COBOL code remain in active operation globally, powering 43% of American banking systems, 95% of ATM transactions, and the majority of government benefits processing across multiple countries. The average age of a COBOL programmer today is between 55 and 60, and approximately 10% of the remaining COBOL-literate workforce retires each year. Around 70% of universities stopped teaching COBOL decades ago, assuming the language would become obsolete. The resulting shortage has driven COBOL specialist salaries to between $121,000 and $150,000 annually, with 60% of organisations that still rely on COBOL reporting that finding skilled developers is their single biggest operational challenge. During the COVID-19 pandemic, the state of New Jersey made a public appeal for COBOL programmers to help process an unprecedented surge in unemployment claims. It was not because the state’s systems had broken down, but because there were not enough people alive who understood them well enough to extend them under pressure. Researchers at the Dutch national institute for mathematics and computer science have characterised their country’s equivalent situation as a “knowledge crisis,” noting that the real problem is not the programming language itself but the loss of system knowledge. It is the loss of the understanding of decades of accumulated business logic, regulatory adaptation, and organisational context that retires with the developers who hold it.
The parallel to the current moment in general software development is direct and uncomfortable. COBOL’s crisis was the product of a single failure mode: a generation of practitioners was allowed to reach retirement without succession planning or structured knowledge transfer, in a field that educational institutions had stopped feeding. The emerging crisis in mainstream software development is the product of that same failure mode operating simultaneously across a much larger and more strategically critical domain. The difference is scale, not kind. The organisations now eliminating junior development roles are not simply reducing headcount. They are removing the mechanism by which institutional technical knowledge passes from one generation of practitioners to the next. When the senior cohort that currently holds that knowledge moves on, as demographic data confirms it will in increasing numbers from the late 2020s onwards, the knowledge will not transfer itself. It will simply be lost, at a cost that will be measured not in recruitment fees but in the ability to maintain, extend, and direct the AI-augmented systems that organisations will by then have become entirely dependent upon.
Issue 4: The Decline of the Developer Experience
Of the four factors driving the emerging talent crisis, the deterioration of the developer experience within organisations is simultaneously the least discussed in public policy terms and the most immediately actionable. It is also the factor with the most direct bearing on retention. It is the primary mechanism by which organisations either hold on to the experienced developers they currently have, or lose them into a market that will soon be competing intensely for exactly that cohort.
Developer experience, in the research literature, encompasses everything that shapes a developer’s daily working life: the quality of tooling, the degree of autonomy over their own work, the amount of time available for focused technical contribution versus administrative overhead, the clarity of direction from leadership, the health of the codebase they inherit, and the organisational culture within which all of this operates. It is not a soft metric. McKinsey’s Developer Velocity Index research found that organisations in the top quartile of developer experience outperformed those in the bottom quartile by up to five times in revenue growth. Gartner’s 2024 research found that teams with high-quality developer experiences are 20% more likely to retain their talent. The business case is not ambiguous, and yet the data on the current state of developer experience across organisations is deeply concerning.
The most comprehensive recent study is the State of Developer Experience Report, produced jointly by Atlassian and DX, drawing on surveys of 1,250 engineering leaders across the US, Germany, France, and Australia, and 900 developers globally. Its headline finding is stark: 69% of developers report losing eight or more hours per week to inefficiencies. That is 20% of total working capacity, or one full working day in every five, consumed not by productive development work but by the accumulated friction of poor tooling, technical debt, insufficient documentation, and unclear direction. For an organisation with 1,000 developers earning an average of $100,000 annually, the arithmetic produces a loss of approximately $18.5 million per year in wasted capacity. The principal causes identified by developers themselves are technical debt, followed by insufficient documentation, complex build processes, and lack of focused time. These are not novel problems, they are the predictable consequence of years of underinvestment in engineering foundations in favour of feature velocity.
The technical debt dimension deserves particular attention because it compounds directly with the ageing and attrition dynamics described in Section 3. Technical debt in US organisations alone was estimated at $1.52 trillion in 2022 by the Consortium for Information and Software Quality. Some 91% of CTOs identified technical debt as their biggest operational challenge heading into 2024, according to STX Next’s Global CTO Study. McKinsey research indicates that CIOs estimate technical debt amounts to between 20% and 40% of the total value of their technology estate, with approximately 30% reporting that more than a fifth of their technology budget nominally allocated to new development is actually diverted to resolving issues arising from it. Technical debt does not remain static. It accumulates. And it accumulates fastest when the experienced developers who understand the codebase well enough to manage it strategically are stretched across expanded responsibilities following headcount reductions, or are absent because they have left.
The layoff cycles of 2022 to 2025 have directly worsened this dynamic. LeadDev’s Engineering Leadership Report 2025, surveying 617 engineering leaders and developers, found that 65% reported expanded responsibilities following recent workforce reductions, with 40% managing more direct reports than previously. Only 3% reported a decrease in scope. The same survey found that 22% of respondents were experiencing critical levels of burnout, with a further 24% moderately burned out. This means nearly half of the engineering leadership population was operating under significant stress. The 2024 Stack Overflow survey captured a data point that had not appeared in prior years: for the first time, senior developers reported lower job satisfaction than junior developers. The profession’s most experienced practitioners, precisely the people whose retention is most critical, are the least satisfied with their working conditions.
The Haystack Analytics study found that 83% of developers report experiencing burnout. The LeadDev survey found that 40% of engineering leaders noted their teams were less motivated than a year previously. A 2025 survey cited in the JetBrains State of the Developer Ecosystem Report found that 66% of developers do not believe current organisational metrics accurately reflect their real contributions. This is a finding that speaks to a profound disconnect between how developers experience their own work and how leadership perceives and measures it. That disconnect is not merely a morale concern. It is a structural risk, because developers who feel their work is neither understood nor fairly measured by their organisations are the most likely to leave those organisations when the labour market improves which, as the BLS projections confirm, it will.
The interaction between poor developer experience and the junior hiring freeze deserves explicit attention. One of the most consistently documented benefits of investing in junior developers is the forcing function it creates on developer experience itself. Onboarding a new developer requires that processes be documented, codebases be navigable, tooling be coherent, and mentoring structures exist. Organisations that maintain a flow of junior developers into their teams are, whether they recognise it or not, continuously incentivised to keep their technical foundations in reasonable order. Organisations that eliminate junior hiring remove that incentive entirely. Technical debt accumulates uninhibited. Documentation degrades. Onboarding processes atrophy. The codebase becomes progressively less navigable to anyone other than the small number of senior developers who already carry its institutional history in their heads — until those developers leave, at which point the knowledge gap becomes an operational emergency rather than a planning problem.
The 2025 State of Developer Experience report found that 63% of developers consider developer experience to be either important or very important when deciding whether to remain in their current role, and that nearly two in three would consider leaving a role if dissatisfied with their developer experience. Some 86% of engineering leaders agreed that attracting and retaining top talent will be difficult without meaningful improvement in developer experience. The gap between that stated recognition and the actual investment being made to address it is where the fourth factor becomes an accelerant of the other three. Poor developer experience drives senior attrition. Senior attrition concentrates institutional knowledge risk. Institutional knowledge risk becomes critical at exactly the moment when the formation pipeline, already suppressed by the forces described in Section 2, fails to produce the replacements needed to address it.
Conclusion: The Counterintuitive Imperative — Building the Future In-House
The four factors examined in this paper do not operate independently. They form an interlocking system, and the interactions between them lies the real danger. AI tools are driving a hiring freeze on junior developers. That hiring freeze is sending a market signal that is already suppressing the formation pipeline. The formation pipeline contraction will reduce the future supply of experienced developers at exactly the moment when the current senior cohort begins aging out in significant numbers. And across all of this, the deteriorating quality of the developer experience within organisations is accelerating the attrition of the experienced practitioners who are, right now, the only buffer between present stability and future crisis. Each factor makes every other factor worse. The system is moving in one direction, and it is not self-correcting.
The timeline is the critical point for any organisation seeking to act. The consequences of the decisions being made today will not arrive as an emergency in the next quarterly earnings cycle. They will arrive quietly, between 2028 and 2034, as a realization that the experienced developers an organisation needs to direct and maintain its increasingly complex AI-augmented systems are simply not available. To be clear, it is not because those developers do not exist somewhere, but because the pipeline that should have been forming them for the past five years was deliberately closed. At that point, the organisation’s options reduce to three: pay a significant salary premium to compete for a constrained supply of experienced developers in an open market; recruit internationally at comparable cost and with no guarantee of institutional fit or domain knowledge; or attempt to rebuild internal capability from scratch, which carries both a multi-year timeline and the overhead of doing so from a degraded technical foundation encumbered by years of accumulated debt.
None of those options is cost-free. None of them is fast. None of them delivers the depth of institutional knowledge, contextual understanding, and organisational alignment that comes from developing talent internally over time. The organisations that will be best positioned in 2030 and beyond are the ones that treated the current period not as an opportunity to reduce headcount but as an opportunity to build the bench they will need when the arms race begins in earnest.
The recommendation of this paper is therefore straightforward in principle, even if it requires significant resolve to execute against the prevailing business climate. Organisations, both private sector firms and public bodies responsible for critical digital infrastructure, should recommit to the structured development of junior and early-career developers as a core strategic investment, not a discretionary cost. This means maintaining a deliberate inflow of entry-level talent into development teams even when AI tools appear to reduce the immediate operational need for them. It means creating structured career pathways that develop early-career developers into the mid-level and senior practitioners the organisation will need in five to ten years. It means investing in the developer experience conditions — manageable technical debt, coherent tooling, documented codebases, and protected time for deep work — that make those career pathways viable and that retain the senior developers already in place. And it means treating knowledge transfer from senior to junior developers not as an optional cultural nicety but as a formal operational continuity practice, with the same organisational seriousness applied to any other critical infrastructure risk.
The COBOL parallel, examined in Section 3, is instructive about what this looks like when it goes wrong at scale. Organisations and governments that failed to maintain succession in that domain are now spending extraordinary sums to maintain systems they no longer fully understand, at the mercy of a talent market they no longer influence. The Social Security Administration’s one billion dollar AI-assisted modernisation programme is not the cost of upgrading old technology. It is the cost of a generation of succession planning failures, now coming due. The emerging equivalent across the broader software development profession carries the potential to be an order of magnitude larger in its impact, because the systems at stake are not legacy administrative platforms but the AI-augmented production systems on which every sector of the modern economy will depend.
There is a further dimension that applies with particular force to public sector organisations. Private sector firms operating in competitive markets will eventually respond to talent shortages through compensation pressure. The response will manifest expensively, but it will be responsive. Public sector bodies typically cannot move at that speed. Government agencies, public health systems, defense establishments, and public utilities that allow their internal development capability to atrophy in the current period will find themselves structurally disadvantaged in any future talent market recovery, because they will be competing on compensation terms they cannot match and on employer brand propositions that the private sector will always be able to outbid. The organisations within the public sector that will retain digital resilience in the decade ahead are those that recognise now that their talent pipeline is a public infrastructure asset, not an administrative overhead, and invest in it accordingly.
The argument for returning to in-house talent development is sometimes characterised as nostalgic. It is seen as a preference for an older and simpler model of workforce development in an era when AI has changed everything. That characterisation misreads both the evidence and the moment. The case for building talent internally is not nostalgic. It is adaptive. The organisations that will most effectively harness the genuine long-term productivity gains that AI will eventually deliver are not those that have reduced their human development capability to a skeleton crew of senior practitioners. They are the ones that have maintained the human formation pipeline through which the understanding, judgement, and institutional knowledge required to direct AI systems effectively is created and transmitted. AI does not eliminate the need for experienced human developers. It raises the stakes for having them.
The data assembled in this paper points to a window that remains open, but is closing. The formation pipeline has contracted sharply but has not yet collapsed irreversibly. The senior cohort is aging but has not yet departed in the volumes that will define the crisis. The developer experience is deteriorating but has not yet driven the wave of senior attrition that poor conditions ultimately produce. There is still time for organisations that choose to act to do so from a position of relative strength, building the capability they will need before the market forces them to compete for it. That window will not remain open indefinitely. The organisations that act in the next two to three years will be the ones setting the terms of the talent market in 2030. Those that wait for the crisis to arrive before responding will be competing on terms they did not choose and cannot control.
The choice, at its core, is between building and buying. Building is available now, at manageable cost, through the deliberate and sustained investment in developing junior developers into the senior practitioners of the next decade. Buying will be available later, at substantially higher cost, in a market shaped by the collective failure of organisations to build. The research is unambiguous about which of those two futures is preferable. The only question is whether organisations will act on that research before the preferable future is no longer available to them.
List of Cited Works
Section 1: The AI Productivity Question
METR — Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (July 2025) Joel Becker, David Rein, Nate Rush, Elizabeth Barnes. 2025. “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.” Research Paper. METR, July. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/.
Faros AI — The AI Productivity Paradox Report (June 2025)
Brynjolfsson, E., Chandar, B., and Chen, R. — Canaries in the Coal Mine: Early-Career Workers in AI-Exposed Occupations (2025). ADP high-frequency payroll microdata analysis through September 2025.
Humlum, A. and Vestergaard, E. — AI Chatbot Adoption and Labour Market Outcomes: Evidence from Danish Administrative Records (2025). Difference-in-differences study across 11 occupations through December 2024.
Doron Yeverechyahu, Raveesh Mayya, and Gal Oestreicher-Singer — GitHub Copilot Rollout Study: Commits, Pull Requests and Code Contributions (2024).
Industry Reports and Surveys
GitHub / Microsoft and MIT Sloan School of Management — AI Developer Tools: Productivity Impact Study (500+ developers, August 2024).
Stack Overflow — Developer Survey 2024 and 2025 (65,000+ and 49,000+ respondents respectively).
Open Source Infrastructure and Code Quality
Django Security Team — Updated Security Reporting Policy Rejecting AI-Generated Reports (2025).
Node.js Security Team — Minimum Quality Score Thresholds for Vulnerability Reports (2025).
Nick Wellnhofer (libxml2) — End of Embargoed Vulnerability Reports Statement (June 2025).
Jeff Geerling — AI Slop and the Collapse of Open Source Pull Request Quality (2025). Public writing.
Section 2: The Education and Training Pipeline
Entry-Level Hiring Collapse
Brynjolfsson, E., Chandar, B., and Chen, R. — Canaries in the Coal Mine (see Section 1). Early-career employment declines of 15–16% in AI-exposed occupations.
Layoffs.fyi — Tech Layoff Tracking Data: 260,000+ layoffs in 2023, 150,000+ in 2024.
The Glass Cage (Blog): https://fs.blog/the-glass-cage-nicholas-carr/
Enrolment and Degree Completion
National Center for Education Statistics (NCES) — CS Bachelor’s Degree Completions Data: 51,696 in 2013–14 rising to 112,720 in 2022–23.
Computing Research Association (CRA) — Pulse Survey on CS Graduate Enrolment (2024).
University of California System — Computer Science Undergraduate Enrolment Data (2024).
Bootcamp and Alternative Pipeline
Momentum Learning — Closure Announcement Citing AI Impact on Job Placement Rates (2024).
Launch Academy — Closure Announcement Citing AI Impact on Job Placement Rates (2024).
2U — Financial difficulties and programme closures (2024).
Graduate Employment
Federal Reserve / Burning Glass Institute — CS Graduate Unemployment Rate: 6.1% in 2025 versus 3.6% overall rate.
Section 3: The Ageing Workforce
Global Developer Demographics
Evans Data Corporation — Worldwide Developer Population and Demographic Study, 34th Edition (2023–2028).
US-Specific Workforce Ageing
Pave Compensation Data (via LeadDev) — Average age at large public technology firms rose from 34.3 to 39.4 between January 2023 and August 2025.
Lemon.io — US Developer Age Statistics: Average 39.2 years (male), 38.5 years (female).
Age Discrimination
US Equal Employment Opportunity Commission (EEOC) — Technology Sector Workforce Report (late 2024). Workers aged 40+ in tech declined from 56% (2014) to 52% (2022); overall US workforce 53% over 40.
University of Gothenburg — Age Discrimination in the Tech Industry: The 35-Year Threshold (2021).
CWJobs — Age Discrimination in UK Tech Survey. Discrimination reported beginning at age 29, considered ‘over the hill’ by 38.
COBOL Parallel
Various industry sources — COBOL: 220 billion lines in production; 43% of banking systems; 95% of ATM transactions. Average COBOL developer age: 55–60.
Section 4: Developer Experience Decline
STX Next — CTO Survey: 91% cite technical debt as their biggest challenge.
McKinsey & Company — Technical Debt: Tackling the Hidden IT Problem (2022). Technical debt represents 20–40% of technology estate value; 30% of CIOs report more than 20% of new product budget diverted to debt resolution.
Haystack Analytics — Developer Burnout Survey. 83% of developers report experiencing burnout.
Note on sources
Where a source exists in both a primary form (such as a researcher’s own published writing or an official organisational report) and in secondary coverage (such as trade press reporting), the primary source is listed above. Secondary sources consulted during research included reporting from The Register, BleepingComputer, LeadDev, and Computerworld, which provided corroborating coverage for several findings, particularly in the open source infrastructure and age discrimination sections.