In the early 2000s, the United States needed to refurbish its nuclear warheads. Routine maintenance. One component, a classified material called Fogbank, had not been manufactured since the 1980s. The engineers who originally produced it were retired or gone.
The new team followed the documented procedures. The material failed. They tried again. Failed again. For a full year, some of the most capable engineers in the country tried to reverse-engineer a material their own organization had invented — and could not.
The cost reached $69 million before someone found the answer. Modern purification processes were removing impurities that were essential to the material’s function. The original makers had intuitively known not to over-purify — but that knowledge had never been written down. It was the kind of thing you learned by standing next to the person who made it and watching what they did.
The people who stood next to them were gone. The knowledge went with them.
The case shows what can happen when experienced people leave before others have learned from them. Right now, that story is playing out across every industry — on a scale that makes Fogbank look like a rounding error.
The Double Hit
Two forces are converging. Neither alone would be a crisis. Together, they are.
From the top: 4.1 million Americans are turning 65 every year through 2027. That is 11,200 per day. By 2030, all 73 million Baby Boomers will have crossed the traditional retirement threshold — concentrated in senior roles, the positions where tacit knowledge lives.
From the bottom: entry-level job postings in the U.S. plunged 35% between January 2023 and June 2025. In tech, the drop was 67%. A third of positions still labeled “entry-level” now require prior work experience. The category is becoming a fiction.
Research, analysis, drafting, and data crunching were often tedious work for a junior. They were also a way into the profession. Working beside an experienced colleague, a newcomer could see what happened when a plausible recommendation met reality and make mistakes while someone was there to help. Over time, that practice became judgment.
AI automates that grunt work. Microsoft’s own research found that the time savings come from automating the exact activities that historically functioned as on-the-job training. The efficiency gain is real. So is the learning loss.
Experienced people are leaving at the top of the pipeline while organizational choices are closing the entry points below. The middle, where knowledge transfer used to happen, is vanishing.
The Talent Doom Cycle
The consulting industry is running this experiment in real time.
BCG’s AI tool, Deckster, now handles roughly 80% of what junior analysts used to do. DeciBio, a life sciences consultancy, cut its entry class from 15 to 4 despite growing revenue. The logic is straightforward: why hire fifteen juniors when AI handles most of what they would have done?
The answer takes about five years to become obvious. Those juniors were going to become your future partners.
55% of employers now admit regret over AI-related layoff decisions. Not because the AI stopped working — it works fine. The regret is that they destroyed the pipeline that produces the people who know whether the AI’s work is right.
Replace juniors with AI. The internal pipeline starves. Future senior talent never develops. The organization becomes dependent on external hiring at inflated costs — but there is no external supply, because every organization is doing the same thing. CNBC called it the “talent doom cycle.” The risk is losing the internal route from junior work to senior expertise.
IBM is the contrarian bet. They are tripling entry-level hiring. Their CHRO: “Companies that doubled down will be most successful.” One company wagering that the apprenticeship pipeline has value that does not appear on quarterly reports.
Many are wagering it does not. We will find out around 2030.
Can AI Bridge the Gap?
One possible answer is that AI itself might solve the problem it created. If AI can compress four months of call center learning into two, perhaps it can compress judgment and expertise at scale.
The evidence cuts both ways.
In a controlled study by Shen and Tamkin, participants who used AI scored 17% fewer points on unassisted assessments afterward. The AI had become a crutch — remove it, and they performed worse than if they had never used it.
I saw the same thing in miniature two months ago: a 25-year engineer who couldn’t change a directory or list files in a terminal. I caught myself asking, “You never used DOS?” Of course he hadn’t. He’d always had the mighty mouse doing it for him — and two and a half decades later, the foundation simply wasn’t there.
The researchers also identified six interaction patterns with different learning outcomes. Three destroyed learning: delegation, instant answers, uncritical acceptance. Three preserved it: Socratic questioning, guided struggle, feedback loops. The learning outcomes differed according to how participants used the tool.
A PNAS study on math education confirmed this. Unrestricted GPT access made students 17% worse on subsequent tests. Guardrailed GPT — designed to question rather than answer — preserved learning completely.
The variable is not whether AI is present. It is whether anyone designed the interaction for learning rather than productivity. Market incentives overwhelmingly favor productivity — reduce costs this quarter, not build capabilities over five years. Deliberately designed AI apprenticeship works. It is also vanishingly rare.
What Cannot Be Written Down
Some expertise has to be learned through practice. Neither AI nor documentation can transfer it fully on its own.
NASA provides the companion case to Fogbank. When the Orion capsule’s uprighting system failed during testing, engineers needed Apollo-era documentation. Enterprise search at Johnson Space Center returned nothing. Months of consulting retired engineers. Nothing. A new semantic search tool eventually found 200 relevant documents in three hours — sitting in NASA’s systems all along, invisible to anyone who did not already know to look for them.
Technology solved retrieval, leaving the team with the work of interpreting what it found. Finding the documents required someone who knew what to look for. Understanding them required someone who knew what they meant. Organizations that want to survive this transition need to capture and structure their institutional knowledge before the people who hold it walk out the door.
42% of institutional knowledge is job-specific and never shared with coworkers. Not because people hoard it — it is the kind of knowledge that surfaces only when something goes wrong. The “we tried that in 2004” warning. The contractual dependency nobody writes into the manual because everyone in the room already knows it.
In the nuclear sector, nearly 40% of the workforce will be eligible for retirement within the next decade. A hiring lull from the late 1980s to the early 2000s left a generation-sized gap — a missing cohort of mid-career workers who would normally be stepping into senior roles. The knowledge has nowhere to go.
The $69 million Fogbank case shows why practice matters alongside documentation: some knowledge exists only in how people do the work. The critical variable was impurities that looked like defects to anyone following the documentation. Only the original makers understood, intuitively, what to leave alone. You cannot codify what was never conscious.
The Window
Every generation worries the next will not be ready. Socrates worried writing would destroy memory. Fair enough. But previous automation waves eliminated manual tasks while preserving cognitive apprenticeship. AI eliminates cognitive apprenticeship itself — the learning-by-doing that was the vehicle for tacit knowledge transfer. The way project phases are collapsing around AI only accelerates this: when entire coordination layers vanish, so do the roles where juniors once learned by watching seniors navigate complexity.
The data converges on 2027 to 2030. Peak retirements. AI infrastructure scaling. A timing mismatch that compounds the problem: each megawatt of AI data center capacity requires approximately 1,800 electrician-hours — nearly five times more than traditional facilities. The industry destroying the apprenticeship pipeline needs a massive expansion of the apprenticeship pipeline to build its own infrastructure.
Estimates suggest full Boomer retirement could reduce U.S. GDP growth by as much as 7.3% by 2030. Eighty-five million jobs globally could go unfilled — not because the work cannot be done, but because the people who know whether it was done right will be gone.
Organizations need to give people time to pass on what they know before they leave. Phased retirement, working in pairs, and AI tools designed to teach rather than replace are practical places to start. They require acting while an experienced colleague is still available to answer questions, rather than waiting until the missing knowledge stops the work.
If the first article in this series argued that your experience is more valuable than you think, this one asks a harder question: what are you doing with it?
The knowledge that makes experience irreplaceable is the same knowledge that makes it perishable. It exists in people, not systems. When they leave, it leaves.
If you carry knowledge that has never been written down, passing it on is part of the work ahead. Let someone watch how you handle an exception, explain why the obvious answer is wrong, and make room for their questions while you are still there to answer them.
The $69 million Fogbank case shows how expensive lost expertise can be to recover. Identify where your own organization depends on knowledge held by people nearing retirement.
References
-
“Fogbank: How the United States Forgot How to Make Its Nuclear Weapons.” SciTales, 2023. https://scitales.com/fogbank-how-the-united-states-forgot-how-to-make-its-nuclear-weapons/
-
“Peak 65: Baby Boomer Retirement Statistics.” Retirepreneur, 2025. https://www.retirepreneur.com/baby-boomer-retirement-statistics
-
“Entry-Level Jobs and AI: 2026 Report.” Rezi, 2026. https://www.rezi.ai/posts/entry-level-jobs-and-ai-2026-report
-
“AI Is Creating a Future Workforce Disaster as It Erases Entry-Level Skillbuilding.” AllWork, 2026. https://allwork.space/2026/02/ai-is-creating-a-future-workforce-disaster-as-it-erases-entry-level-skillbuilding/
-
“How Gen AI Could Change the Value of Expertise.” Harvard Business Review, 2025. https://hbr.org/2025/03/how-gen-ai-could-change-the-value-of-expertise
-
“The AI Layoff Trap: Why Half Will Be Quietly Rehired.” HR Executive, 2026. https://hrexecutive.com/the-ai-layoff-trap-why-half-will-be-quietly-rehired/
-
Shen, Tamkin. “AI’s Impact on Skill Formation.” SSRN, 2024. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4895566
-
“Remembering Apollo: Why Knowledge Management Is Mission-Critical at NASA.” APQC, 2023. https://www.apqc.org/blog/remembering-apollo-why-km-mission-critical-nasa
-
“The Cost of Lost Knowledge: What Organizations Lose When Employees Leave.” Learn to Win, 2024. https://www.learntowin.com/blog/cost-of-lost-knowledge
-
“Nuclear Talent Crisis.” Nuclear Business Platform. https://www.nuclearbusiness-platform.com/media/insights/nuclear-talent-crisis
-
“GenAI’s Human Infrastructure Challenge: Can the United States Meet Skilled Trade Labor Demand?” CSIS, 2025. https://www.csis.org/analysis/genais-human-infrastructure-challenge-can-united-states-meet-skilled-trade-labor-demand