The Great Knowledge Drain

In 2025, a federal effort set out to rewrite the Social Security Administration's core benefits system — an estimated 60 million lines of COBOL — in a matter of months. Veteran technologists publicly called the timeline unrealistic, not because the code was too complex to ever replace, but because the people who actually understand how it behaves, workarounds and all, are retiring, and a rewrite that size has failed before on a much longer clock [1][2]. In the same year, a Stanford analysis found employment for the youngest software developers had fallen nearly 20 percent since late 2022, even as employment for experienced developers in the same AI-exposed field kept growing [4].

Those two facts are usually reported separately — one as a government-IT story, the other as a labor-market story. This investigation asks what happens when they're read together: organizations across very different sectors are losing the people who hold their institutional knowledge at the same moment the entry-level roles that used to grow replacements are contracting. Not a prediction. A pattern already visible in the data, in more places than one.

Organizations have always run partly on knowledge that isn't in any manual. Researchers call it tacit knowledge — the workarounds, historical context, and unwritten judgment calls that experienced staff accumulate and rarely transcribe, because most of it only becomes visible when something goes wrong and someone who's seen it before knows what to check first. A substantial peer-reviewed literature on knowledge retention from retiring workers agrees on one point across decades of study: tacit knowledge is the hardest knowledge type to capture or transfer formally, and it typically survives only through direct mentoring or working alongside someone who already has it [9].

That has always made retirements a quiet risk. What's different now is scale and timing. Large cohorts of experienced specialists across multiple critical sectors — mainframe programming, nuclear operations, air traffic control — are aging out of the workforce within roughly the same window, a pattern researchers describe in the nuclear sector as an "age bathtub": a generation of senior experts, a generation of new hires, and comparatively few in between to bridge the two [6]. Organizations built their handoff processes assuming a steadier trickle of retirements. What they're facing instead, across several sectors at once, looks more like a wave.

Start with the systems everyone still depends on and almost nobody is trained to maintain. Roughly 220 billion lines of COBOL remain in active production across banking, insurance, airlines, and government, and nearly a third of the programmers who maintain it are expected to retire by 2030 — the average COBOL programmer is about 55, and more than 85 percent of US universities dropped the language from their curricula decades ago [2]. The Social Security Administration's own attempted rewrite ran into this directly: roughly 30 percent of its CIO-office staff were eligible for retirement at the time the project was proposed, and a prior modernization attempt in 2017 — planned around a five-year timeline — never reached completion [3].

The pattern repeats in physical infrastructure. Nearly 40 percent of the global nuclear workforce is approaching retirement age, the legacy of a hiring lull from the late 1980s through the early 2000s that left the industry short on the mid-career staff who would normally mentor new hires [6]. The US Department of Energy projects the sector needs to nearly triple its workforce by 2050 — about 375,000 additional workers — while 63 percent of nuclear-sector manufacturing employers already call hiring "very difficult" today [7]. Air traffic control shows the same shape from a different angle: a mandatory retirement age of 56, 1,282 controllers retired over five years with 819 more expected through 2028, and roughly 91 percent of US facilities currently staffed below FAA-recommended levels — serious enough that Congress introduced legislation to let experienced controllers work longer [8]. And organizations broadly are aware of the exposure: APQC's 2025 research finds employers expect 51 percent of their workforce to retire or leave within five years, yet 41 percent of organizations say they rarely or never attempt to capture departing employees' knowledge before they go [1].

Competing hypothesis Whether AI closes this gap or widens it is genuinely contested. One camp treats AI as the knowledge-drain's most promising countermeasure: generative tools can translate undocumented COBOL into modern languages, help capture retiring experts' explanations in searchable form, and pair junior staff with AI-assisted mentoring the way a 2025 peer-reviewed study of intergenerational knowledge transfer describes [9]. On this view, AI doesn't replace the mentor — it makes whatever time remains with the mentor more productive.

Competing hypothesis The other camp points to a harder problem: AI tools are simultaneously reducing the volume of entry-level work available to the people who would eventually become those mentors. Employment for developers aged 22–25 fell nearly 20 percent from late 2022 to mid-2025 even as employment grew for developers aged 35–49 in the same AI-exposed occupations, according to Stanford's Digital Economy Lab [4]. A 2024 survey found 70 percent of hiring managers now believe AI can do the work traditionally assigned to interns, and internship postings in tech fell roughly 30 percent since 2023 [5]. If that pattern holds, AI doesn't just fail to solve the knowledge-transfer problem — it removes the on-ramp that used to produce the next generation of people capable of receiving that knowledge at all. Which effect dominates likely varies by sector, and by whether an organization treats AI as a mentorship tool or a junior-staff substitute; the evidence assembled here doesn't resolve that for any single employer, only shows both dynamics operating in the data at once.

This investigation sits at the junction of three systems that don't normally get analyzed together: workforce demographics (retirement timing, set decades ago by hiring patterns no one can now undo), entry-level labor economics (the hiring contraction Part 2 of this Convergence, "Why Hiring Feels Broken," examines directly), and sector-specific infrastructure risk (nuclear, aviation, and government software each running on staff nearing retirement, as the Age of AI's other investigation into restarted reactors also documents in the nuclear workforce specifically [6][7]). The reason these three systems interact rather than merely coexist is structural: the same AI tools being weighed as a knowledge-capture solution [9] are, in the adjacent labor market, also the tools reducing the junior hiring that would otherwise replenish the mentors these sectors will need [4][5]. A policy or corporate decision made in one system — say, replacing interns with AI assistants — has a measurable effect on a seemingly unrelated system: how many people will be available to receive a retiring nuclear engineer's tacit knowledge a decade from now.

Two clocks are running in every sector this piece examines — one counting down to retirement, one counting whether a replacement is being trained in time. Nothing below predicts which sector lands where; it traces how that race could actually resolve, drawn from the sector data already on the table.

The knowledge drain is real and already measurable — visible in mainframe demographics, nuclear and aviation workforce plans, and an organizational survey in which fewer than half of employers even attempt to capture what departing staff know [1][2][6][8]. What the evidence doesn't yet support is a single verdict on outcome, since the mechanism is two-sided and playing out differently by sector. These are the trackable indicators worth watching over the coming months and years.

• Federal & legislative action — Whether the SSA modernization effort publishes a revised, realistic timeline or a public account of what the rushed approach actually encountered [3]. Whether the Control Tower Continuity Act or similar retirement-age legislation for critical specialists advances [8].
• Workforce research updates — Updated APQC or equivalent workforce-research findings on whether the 41% "rarely or never capture knowledge" figure improves [1]. Any newly disclosed data-center or federal-agency knowledge-capture program tied explicitly to retirement risk.
• Sector hiring data — Whether entry-level tech and engineering hiring data (Stanford Digital Economy Lab and equivalents) shows the 22–25 age-cohort employment decline stabilizing or continuing [4]. Nuclear-sector hiring-difficulty rates against the DOE's workforce-tripling target [7].
• Operational track record — Whether documented operational incidents (grid, aviation, government IT) show a measurable rise in root causes traced to lost institutional knowledge, versus organizations successfully closing the gap through mentorship or AI-assisted capture [6][9].

If organizations are losing the people who hold their institutional knowledge, and the entry-level roles that used to grow their replacements keep shrinking, how are those same organizations actually trying to hire around the gap right now — and why does that process feel so broken? That's Part 2. In the meantime: have you watched institutional knowledge disappear from an organization you've worked in, or watched someone try to catch it before it left? Tell us. We read every message.