A reliability lead at a chemical plant retired after 32 years. (Details anonymized.) Her expertise was never documented. Within months, an unplanned compressor shutdown cascaded into production delays that cost millions. Her replacement ran every diagnostic in the manual. The retiring engineer would have checked a specific upstream valve first - because she'd seen that exact vibration pattern three times before, and it was never the compressor.

She didn't take a manual with her. She took the ability to hear a motor drifting off-spec from across a unit. She took the instinct to check a specific valve when three unrelated symptoms appeared together. She took 32 years of pattern-matching that nobody had thought to write down - because nobody had designed a role where sharing that knowledge was valued, compensated, and worth her time.

This kind of story is playing out across industrial facilities worldwide.

APQC surveyed 1,000 professionals in mid-2025 and found that 92% of organizations don't consistently capture knowledge from departing employees. Forty-one percent rarely or never attempt it at all. An oft-cited IDC estimate puts Fortune 500 knowledge-sharing losses at $31.5 billion per year - a figure first published in the early 2000s that, if anything, understates today's scale.

The expertise that makes AI systems work in industrial settings doesn't live in a database. It lives in the judgment of people who've spent decades learning what matters and what doesn't. Many of them are about to leave. And the window to build a partnership - where their knowledge becomes a shared legacy instead of a quiet loss - is shorter than most organizations realize.

The Numbers Nobody Can Ignore

APQC found that 51% of the industrial workforce is expected to retire or leave within five years. Deloitte projects 3.8 million manufacturing jobs to fill by 2033, with the majority driven by retirements. Across frontline industries, 72% of managers lack confidence they can retain critical knowledge when experienced workers retire.

In oil and gas, the problem has a name: The Great Crew Change. The GETI 2026 report shows that 48% of the O&G workforce is aged 45 or older. Only 19% is 25-34.

But the demographics alone don't tell the full story. The 1980s oil downturn drove an entire generation away from the industry. In many energy organizations, Gen X is significantly underrepresented. The age distribution isn't a gradual slope - it's bimodal. Boomers on one end, millennials on the other, and a missing middle that should have been the bridge between them.

There is no experienced middle layer to ease the knowledge transfer. The gap isn't gradual. It's a cliff.

"Write It Down Before You Leave" Doesn't Work

Knowledge management research consistently estimates that only 15-20% of organizational knowledge is documented. The other 80% exists as tacit knowledge: unwritten rules, contextual understanding, judgment calls refined over decades of hands-on experience. Surveys cited by PBS and workforce researchers found that 57% of retiring boomers have shared less than half the knowledge their replacement needs. Twenty-one percent have shared none.

The current approaches aren't working. According to APQC, 74% of organizations rely on people-to-people transfer - the slowest, least scalable method. Forty-eight percent conduct exit interviews. But exit interviews capture facts, not intuition. They capture what someone did, not why they checked that specific gauge first, or how they knew that vibration pattern meant a bearing was about to fail.

The band-aid that proves the problem: knowledge management surveys consistently find that a large share of companies "solve" knowledge loss by rehiring retirees as consultants - often at double the pay. Harvard's Dorothy Leonard has argued this is "bad business" - you're paying experts to hoard knowledge rather than transfer it. That's not a knowledge management strategy. It's an admission that you didn't capture what you needed while you had the chance.

Without the Expert, the AI Is Useless

This is where the full series comes together.

The AI models I described in earlier articles didn't reach 93% accuracy because of better algorithms. They got there because domain experts - people with decades of P&ID experience - reviewed the AI's mistakes, corrected them, and fed those corrections back into the system. Without that expert knowledge in the loop, the model sat at 39% and was useless.

This isn't unique to engineering drawings. Consider something as common as classifying maintenance work orders in Maximo. A technician writes: "Chk vlv on P-101, weird noise, same as last spring." An off-the-shelf AI classifier sees words and guesses "valve inspection." The technician who's been on that unit for 15 years knows it actually means: recurring mechanical failure on the main feed pump, third time this year, root cause never addressed - escalate. The vagueness isn't a data quality problem. It's efficient communication between people who share decades of context. No vendor model trained on generic maintenance text will get that right. The only person who can teach the AI what that work order actually means is the person who wrote it - or someone with equivalent experience. And that expertise is worth more than any external annotation vendor, because external annotators will label it wrong confidently, and the model will learn to be wrong at scale.

Every correction became training data. Every edge case became a test case. Nine rounds of expert review turned a mediocre model into a production system. The data moat I described - accumulated expert corrections, evolved annotation guidelines, institutional knowledge embedded in every verified label - all of it required the domain expert to exist and participate.

Now ask the uncomfortable question: what happens when the person who knows why that specific valve configuration matters on that specific unit retires - and their corrections, their edge-case knowledge, their decades of pattern recognition never made it into the training data?

The answer: your AI stalls. The feedback loop that compounds improvement breaks. The model that processes the next facility is no better than the one that processed the last. You've lost the capability - not because the technology degraded, but because the human expertise that powered it walked out the door.

The Federal Reserve Bank of Dallas published research in February 2026 that crystallizes this: AI substitutes for codified, textbook knowledge - but augments workers with tacit, experiential knowledge. Wages are actually rising in AI-exposed occupations that value hands-on experience. The experienced worker becomes more valuable in an AI-enabled world, not less.

But only if their knowledge gets captured before they leave - and "captured" doesn't mean extracted. It means recognized, compensated, and structured as a legacy. The expert who teaches the AI should be credited as the author of that capability, not treated as a data source to be mined on the way out. Organizations that get this right design the role as a capstone - a senior technical position where the most experienced people spend their final years shaping the systems that will carry their judgment forward. Organizations that get it wrong send a retirement card and a SharePoint link.

The Judgment Gap: AI Might Actually Make This Worse

Here's the contrarian angle that should change how you design your AI rollout.

AI tools are accelerating the very knowledge loss they're supposed to prevent. This isn't an argument against AI copilots - it's an argument for redesigning how knowledge moves when AI changes the workflow. Junior engineers fix bugs by prompting an AI assistant instead of asking a senior colleague. Technicians look up procedures in a chatbot instead of shadowing an experienced operator. Every AI-assisted shortcut is a missed knowledge transfer moment - a conversation that would have happened between humans, where tacit knowledge moves through context, stories, and the back-and-forth of shared problem-solving.

IE University put it sharply in 2025: what happens when the next generation of knowledge workers learns through AI from day one? They'll produce equally polished work but lack the foundational judgment to distinguish good insights from compelling-sounding nonsense.

If juniors never develop tacit knowledge through practice and mentorship, who will know when the AI is wrong in ten years?

TSMC - the world's most advanced semiconductor manufacturer - recognized this. Despite heavy investment in automation, they expanded their apprentice program. The rationale: today's junior employees are tomorrow's experts who will know when the AI is wrong. You can't build that judgment from documentation alone.

Then there's the other path. We've already seen what it looks like when organizations treat knowledge capture as extraction - logging employee activity to train replacement systems, then cutting headcount the same quarter they post record profits. Thousands of workers discover their keystrokes and screen activity were feeding the models meant to make them redundant. No consent. No compensation. No legacy - just a dataset.

That approach might work for replicating routine digital tasks. It will never work for industrial expertise. The reliability engineer who can hear a bearing failing from across a unit floor didn't learn that from a screen recording. That judgment was built through decades of consequences - shutdowns avoided, incidents prevented, patterns recognized across hundreds of situations no training dataset contains. It has to be willingly shared by people who trust the process and see their contribution as a legacy, not a transaction.

The answer isn't "replace experts with AI." It's "use AI to capture expert knowledge so the next generation can build on it - while simultaneously building the judgment to know when it's wrong."

You need both the capture infrastructure and the human development. One without the other fails.

This Is a Partnership, Not an Extraction

Let's name the elephant in the room. Every industry is anxious about AI replacing jobs. And now this article is saying: we need to capture what experienced workers know and feed it into AI systems. That can sound like the worst version of the story - mine the workers for their knowledge, then automate them away.

That's not what this is. And the data doesn't support the fear narrative.

The World Economic Forum's 2025 Future of Jobs Report projects 170 million new roles created against 92 million displaced - a net gain of 78 million jobs globally by 2030. The Dallas Fed's research shows that wages are rising in AI-exposed occupations that value tacit knowledge - experienced workers with hands-on judgment are becoming more valuable, not less. PwC found that workers with AI skills now earn 56% more than peers in comparable roles.

More importantly, AI is creating entirely new roles that didn't exist three years ago - roles specifically designed for people with deep domain expertise and no computer science degree:

  • Knowledge Engineers who curate, structure, and maintain the information AI systems rely on to generate accurate responses
  • AI Trainers who provide expert human feedback to improve model outputs - a role where 20 years of P&ID experience is worth more than a PhD in machine learning
  • Domain AI Specialists who evaluate model outputs, test for failures, and write documentation that regulators require
  • AI Auditors who verify that automated systems are making sound decisions in safety-critical environments

These aren't hypothetical. AI training platforms are already hiring domain experts in energy, manufacturing, medicine, and law - and paying premium rates for industry-specific knowledge that no generalist can provide. The demand exists because AI has hit what the industry calls a "data wall" - models have consumed most high-quality public data, and what they need next is domain-specific expertise that can't be scraped from the internet.

For a 55-year-old reliability engineer, this isn't a threat. It's a career transition with leverage. The knowledge that makes you irreplaceable in your current role is exactly what makes you valuable in these emerging roles - and the salary premium reflects it.

But this only works if organizations design knowledge capture as a partnership, not an extraction:

  • Compensate it. If the expert's corrections become training data that generates value for years after they leave, the compensation should reflect that. A knowledge capture role should pay at or above the expert's current rate, not be tacked onto their existing responsibilities for free.
  • Credit it. The systems that result from expert input should be traceable to the people who taught them. This isn't just dignity - it's accountability. When the AI makes a decision on a safety-critical system, you need to know whose judgment trained it.
  • Make it voluntary. Mandating knowledge transfer breeds resentment and poor-quality input. The best knowledge capture happens when experts see their participation as a legacy, not an obligation. Design the role so they want to say yes.
  • Create a path, not a dead end. Knowledge capture shouldn't be the last thing someone does before being shown the door. It should be a senior technical role - a capstone position where experienced people shape the systems that carry their judgment forward. Some organizations are finding that these roles extend careers rather than ending them, as experts transition into part-time AI training and consulting roles.

And here's the part that makes partnership not just ethical, but operationally necessary.

The honest trajectory of AI in industrial operations isn't "AI replaces humans." It's a gradual offloading of cognitive work - from a small percentage to a larger one, over years. Today, AI might handle 5-10% of the cognitive load: basic classification, draft generation, pattern flagging. In three years, that could be 30-40%: routine work order classification, standard anomaly detection, first-pass RCA. In five to ten years, maybe 60-70%: autonomous scheduling, predictive maintenance decisions, self-directed inspection routing.

But at every stage of that curve, someone has to teach the AI what "right" looks like for your facility, validate that it's still getting it right as conditions change, and set the risk thresholds for when it should stop and ask a human. That governance role requires exactly the judgment that experienced workers have built over decades. If you extract their knowledge at Stage 1 and discard them, you have nobody qualified to govern at Stage 3. The AI doesn't just need human expertise once. It needs human judgment continuously - at every stage of the offloading curve.

This is the strongest argument against the extractive model: it's not just wrong ethically. It's operationally self-defeating. You're destroying the governance capacity you'll need in three years.

The question isn't whether to capture expert knowledge. It's whether you do it in a way that treats people as partners in building something lasting - or as resources to be processed on the way out.

What Capture Actually Looks Like

APQC found that 79% of organizations want AI for knowledge capture. Only 21% are doing it. That 58-point gap is one of the clearest first-mover advantages in industrial AI - and it narrows every month.

Expert knowledge exists in three layers, and each requires a different approach:

Explicit (~20%) - Documents, SOPs, manuals. Should already be captured. If not, start here.

Implicit (~35%) - Can be articulated when you ask the right questions. Capture through structured interviews with AI processing; mine collaboration channels for guidance patterns.

Tacit (~45%) - Embodied, intuitive, contextual - "you know it when you feel it." Put the expert in front of the AI's output and capture their corrections. Record the telemetry, have them label the anomalies. This is expert-in-the-loop learning on your actual domain.

That last row is critical. You can't interview someone about the sound a failing bearing makes. But you can record the audio, have the expert mark which frequencies matter, and turn that into training data. You can't ask an operator to write down how they read a control room display under pressure. But you can build a digital twin and capture how they respond. Tacit knowledge becomes training data not through documentation, but through expert-in-the-loop systems - the same feedback loops described in Articles 3 and 4.

The organizations getting results are building capture infrastructure, not running exit interviews.

Aker BP + Cognite built three specialized AI agents for root cause analysis on aging offshore assets - one generates visual cause maps, one retrieves and analyzes documents scattered across SAP and SharePoint, and one pulls sensor time-series data. The problem wasn't a lack of data; it was that engineers spent days manually gathering documents and tracing failure chains for a single equipment failure. Senior Reliability Engineer Tor Arne Amdal said it directly: "It's not enough to just have data. You need the right data, in the right context, at the right time." Results: a recurring mechanical seal failure that used to take days to diagnose was resolved in hours. Across their RCA process, they've confirmed a 70-97% reduction in time for individual analysis steps, with projected savings of thousands of engineering hours as the system scales across six brownfield assets.

A mid-sized utility company working with eGain saw a 46% reduction in onboarding time for technical roles, 38% improvement in problem resolution, and $4.2 million per year in productivity gains - largely by making captured expert knowledge searchable and contextual.

If you're starting from scratch, don't boil the ocean. Start with a 90-day pilot - designed with the experts, not imposed on them:

  1. Weeks 1-2: Identify who holds critical knowledge and have a conversation - not a mandate. Explain the role, the compensation, and what their contribution will become. Deliverable: a list of willing participants, knowledge domains, and timelines.
  2. Weeks 3-4: Pick your top three retiring experts. Score their knowledge areas by business impact multiplied by loss probability. Deliverable: a shortlist of 5-8 knowledge domains to capture first.
  3. Weeks 5-8: Shadow each expert with a junior engineer and an AI transcription/annotation tool. For implicit knowledge: structured interviews processed by AI. For tacit knowledge: have the expert correct the AI's predictions on real operational data. Deliverable: 15-20 structured sessions captured, indexed, and reviewed.
  4. Weeks 9-12: Validate captured knowledge with junior staff. Can they find answers faster? Can they resolve issues the expert used to handle? Deliverable: a searchable knowledge base with a feedback workflow and a named owner.

The goal at day 90 isn't a finished system. It's proof that the approach works - with measurable results you can take to leadership to fund the full rollout.

The Clock Is Running

Boeing is the cautionary tale that everyone in industry knows. Institutional knowledge destroyed through outsourcing and layoffs. The quality crisis that followed wasn't a technology failure. It was a knowledge failure - decades of process expertise lost for short-term cost savings, with consequences that are still unfolding.

The window isn't ten years. Industry estimates put it at 24 months before the steepest part of the retirement curve hits. Every month of delay means more expertise walking out the door that no model, no vendor, and no consulting engagement can replace.

If you've been in the field for 20 or more years - reading drawings, diagnosing equipment by sound, making judgment calls that kept facilities running safely - your plant-specific judgment is the one thing AI can't learn from the internet. Not because of title or tenure, but because you've built judgment under real constraints that no training dataset contains.

That doesn't make you replaceable. It makes you the most qualified person to shape how AI works in your domain - if the organization is willing to design that role with the respect it deserves. Not writing documentation in your last two weeks. Sitting in a senior technical role where you review the AI's output, correct its mistakes, and turn decades of experience into something the next generation can build on. That's a legacy worth leaving.

The most expensive AI failure isn't a bad model. It's a good expert who retired before anyone thought to ask - or worse, one who was asked but given no reason to say yes.

This is Part 5 of a 6-part series on AI in industrial operations:

  1. I've Trained AI on Engineering Drawings. The Model Was Never the Hard Part.
  2. You're Buying AI Like Software. That's Why It's Failing.
  3. Your AI Model Works. Your Organization Doesn't.
  4. Your Data Is the Model. Everything Else Is a Commodity.
  5. The $31.5 Billion Walk-Out: Why Your Best AI Training Data Is About to Retire
  6. Build, Buy, or Die: The AI Decision That Will Define Your Next Five Years

If you're managing a team where the average age is north of 50, this isn't a future problem - it's a current one.

Are your experienced people involved in teaching the AI, or are they just using it? I'd like to hear how you're approaching the transition - what's working, what's not, and where the gaps are. If you want the 90-day pilot framework, send me a DM. I'll share the template and the scoring rubric for prioritizing which knowledge to capture first.