Dave has worked the press line for twenty-two years, and he retires the second Friday in November. Everyone on the floor knows what that means, and nobody has done anything about it. Dave can walk past the number 2 press, hear a faint change in the stroke, and say "that ram's about to gall, back the tonnage off five percent and call me in an hour." He knows that the coil from the cheaper steel supplier needs the dies preheated longer or you get edge cracking on the first forty parts. He knows that when the cosmetic reject rate creeps up on second shift, it is almost never the operators, it is the upstream die wearing out, and he can tell you which die by the shape of the burr. None of this is written down. None of it is in the standard operating procedure. It lives in one man's head, and in roughly twenty weeks it walks out the door and drives home for the last time. The plant will then spend the next two years and a small fortune rediscovering, the hard way, things Dave already knew. This lesson is about the workflow that captures Dave before November, validates what he says so you do not enshrine a myth, and serves it to every shift so it keeps paying long after he is gone.

The Most Valuable Asset Walks Out the Door

Start with why this is the single highest-return move a thinning plant can make in 2026, because the math is not subtle. Roughly two million manufacturing workers need reskilling against about half a million unfilled roles, and 85 percent of manufacturers say staffing shortages are already hurting product quality. The retirees taking their knowledge with them are not being replaced one-for-one by people who know the same things. A green crew, meaning new hires who have never seen this machine throw this fault before, is inheriting machines whose quirks live only in the memory of people who are leaving. The gap is not a technology gap. It is a knowledge gap wearing a technology costume.

This reframes what AI is for on the floor. AI here is not a replacement for Dave and it is not a smarter machine. It is a knowledge multiplier: a way to take what Dave knows and make it available to a thinner, greener crew on every shift, in the moment they need it, grounded in the plant's own reality. The model is only ever as good as the institutional knowledge you feed it. Feed it Dave's twenty years, validated, and a first-year tech on third shift can ask "the number 2 press stroke sounds different and tonnage is drifting, what do I check" and get Dave's actual answer instead of a confident guess from the average of the internet.

AI on the floor is only as good as the institutional knowledge you feed it, and that knowledge is currently retiring in November.

The end-to-end pipeline has three stages, and the discipline is in treating it as a pipeline rather than a one-time recording. Capture gets the knowledge out of the expert's head and the plant's old records into a structured form. Validate confirms each captured item against data and a second source so you do not teach the next shift a habit that was wrong, or a fix that stopped applying when the process changed. Serve puts the validated knowledge where the crew can actually reach it, grounded so the AI answers from Dave's verified knowledge and not from its imagination. Skip capture and you lose the knowledge. Skip validation and you enshrine a myth at scale. Skip serving and you have an archive nobody opens, which is just a more expensive version of the binder on the shelf.

Capture: Getting It Out of Dave's Head

Capture has two channels, and a serious program runs both. The first is the structured interview with the living expert. The second is mining the historical records the expert created over the years without ever organizing them: paper travelers, the document that follows a job through the plant recording what was done; years of computerized maintenance management system history, the CMMS being the software that holds every work order and repair; old quality records; and the margin notes scrawled on procedures. Both channels feed the same knowledge base. The interview captures what Dave knows he knows. The records mining captures what Dave did so often he forgot it was knowledge.

The structured interview is where AI assistance changes the economics. A retiring expert is busy and not naturally a documenter, and a blank "tell us everything you know" session produces rambling and gaps. A better approach uses AI to run a structured, topic-by-topic interview and to transcribe it in real time, so the human interviewer can stay present and probe rather than scribble. The structure matters: walk the assets one at a time, and for each one ask the questions that surface tacit knowledge. What goes wrong on this machine that the manual does not mention? What are the early warning signs you notice before anyone else? What is the trick you use that a new person would never figure out? When this defect shows up, what does it actually mean upstream? What do you do differently with this material, this supplier, this season?

Here is the worked value. A plant scheduled six two-hour interview sessions with Dave over three weeks, AI-transcribed and auto-summarized into draft knowledge entries each evening. Twelve hours of Dave's time produced 140 candidate knowledge items. Compare that to the realistic alternative, which is what most plants actually do: nothing, until Dave is gone, followed by an estimated 18 months of the green crew rediscovering his knowledge through scrap, downtime, and a few customer escapes. If even a third of those 140 items each prevent a single avoidable incident over the next two years, and a single press defect escape that reaches a customer can run well past 20,000 dollars in containment and sorting, the twelve hours of interview time is the cheapest insurance the plant will ever buy. The constraint is not cost. The constraint is that Dave retires in November, and the interview has to happen now.

The records-mining channel runs in parallel and needs the same grounded-AI discipline. Pointing a retrieval-grounded assistant, one that reads only the plant's own documents, at years of CMMS history can surface repeating patterns a human would never spot by hand: the number 2 press logs a galling-related repair every twelve to sixteen weeks, clustering in summer; the cheaper-steel coils correlate with a spike in edge-crack rejects on first run. These are Dave's lessons, written down all along in the data but never read. The AI assembles the pattern and the evidence. It does not get to declare the cause. That is the next stage.

One practical note on capture sequencing that experienced programs learn the hard way: run the records mining first, then the interviews. When you walk into the interview already holding "the press logs a galling repair every twelve to sixteen weeks, here are the dates," you can ask Dave the sharp question instead of the open one. He will say "right, that is the summer heat, and the trick is to back the tonnage off before the second-shift load peaks," and now you have both the pattern from the data and the tacit fix from the expert, joined into one entry. Asking the data its questions before you ask the human his is how twelve hours of an expert's time turns into 140 items instead of forty. The data primes the interview, and the interview explains the data.

Validate: Do Not Enshrine a Myth

This is the stage that separates a knowledge base from a liability, and it is the stage most teams want to skip because capture feels like the finish line. It is not. Twenty years of experience contains real gold and also contains habit, superstition, and fixes that were correct in 2009 and quietly stopped being correct when the process changed in 2018. If you capture Dave's knowledge and serve it unvalidated, you have not preserved expertise, you have mass-produced a myth and handed it to every shift with the plant's authority behind it. A wrong answer that a new tech would have questioned coming from a person becomes unquestionable when it comes from the official knowledge base.

Validation answers a simple question for each captured item: is this actually true, still true, and true for the reasons given? Three tests do most of the work. Test it against the data. Dave says the number 2 press galls every summer because of heat. The CMMS history either supports that clustering or it does not. If the data shows the galling repairs are evenly spread across the year, Dave's mental model has a bias, common and human, toward the failures he remembers most vividly, and the captured item needs correcting before it ships. Test it against a second expert or the engineering source. A torque value, a preheat time, a tolerance: confirm it against the drawing, the standard, or a second experienced tech, because a remembered spec is not a verified spec and a confidently wrong one is exactly the failure mode this whole program warns about. Test whether it still applies. A fix tied to the old supplier, the old die set, or a removed process step may be obsolete. Tag it with the conditions it depends on, or it will mislead the crew the day those conditions change.

A worked example shows why this pays. Of the 140 captured items from Dave's interviews, validation against CMMS data and a second senior tech found that 116 held up cleanly, 18 needed a condition added ("only with the older die set"), and 6 were flatly contradicted by the data, including a long-held belief that a particular reject was caused by operator handling when the records clearly showed it tracked an upstream die-wear cycle. Catching those 6 before serving them is the entire point. Serving even one of them would have sent the green crew chasing operators for a defect the operators did not cause, eroding trust in both the crew and the knowledge base, while the real upstream cause kept shipping scrap. Validation is not bureaucracy. It is the step that makes the knowledge base safe to trust.

Crucially, validation keeps a human accountable, because the customer audits the plant, not the vendor, and "the knowledge base said so" is no more defensible to an auditor than "the model flagged it." Each validated item carries a name and a date: who confirmed it, against what source, and when. That provenance is what lets the knowledge base survive a quality audit and what lets the next engineer trust it without re-litigating every entry.

It helps to grade each captured item rather than treat validation as pass or fail. A simple three-tier confidence label does the work on the floor. Confirmed means the data and a second source both back it; serve it plainly. Conditional means it holds only under stated conditions, such as a particular die set, supplier, or season; serve it with the condition attached so it self-limits. Expert-opinion-unconfirmed means a seasoned person believes it but the data is silent or thin; you may still serve it, but labeled as judgment rather than verified fact, so the crew weighs it accordingly. This grading matters because the alternative, a flat "everything in the base is true," is exactly the overconfidence that makes a rotting base dangerous. A tech who knows an entry is expert opinion rather than confirmed fact applies the right amount of skepticism, and the base stays honest about what it actually knows.

Serve: Grounded Answers on Every Shift

Validated knowledge that nobody can reach is the binder on the shelf with better formatting. The serving stage puts the knowledge in front of the crew at the moment of need, and the technology that makes this work without inventing answers is retrieval-augmented generation, abbreviated RAG. In floor terms, RAG means the AI does not answer from its general training; it first retrieves the relevant validated entries from your knowledge base, then writes its answer using only those entries, and it cites which entry it used. The grounding is the whole point. It is the difference between an assistant that gives you Dave's verified answer and a chatbot that gives you a plausible, confident, invented one.

Picture the third-shift tech, eight months on the job, at 2 a.m. with no senior person in the building. The number 2 press tonnage is drifting and the stroke sounds off. He asks the floor assistant on the panel, in plain language, what he is hearing. The RAG system retrieves Dave's validated entry, the one tagged to this asset and this symptom, and answers: "This pattern on press 2 has historically preceded ram galling. Recommended first step from the validated knowledge base: reduce tonnage 5 percent and inspect the ram surface. Escalate to reliability if the surface shows scoring. Source: knowledge entry PRESS2-014, validated by J. Ruiz against CMMS history on 2026-09-30." The tech gets Dave's actual judgment, with its source, at 2 a.m., on his first hard night. That is the multiplier. A green crew runs a little more like the crew that had Dave on every shift.

Three design rules keep the serving stage honest. It cites or it refuses. If the knowledge base has no validated entry for the question, the assistant says so and points to a human, rather than guessing. A confident guess at 2 a.m. is worse than no answer, because the tech will act on it. It stays advisory on anything that moves or hurts. The assistant recommends; the human acts and owns the action. It does not reach into the control system, because keeping AI out of direct control of the line is a hard constraint on a brownfield floor where 78 percent of plants cannot even centrally monitor their operational-technology network, the OT network being the machines and controllers themselves as opposed to the office IT. It is fast enough to use under pressure. An answer that takes two minutes to surface is an answer the tech abandons for a phone call. The serving layer earns adoption only if it is faster than the alternatives the crew already has.

Keeping It Honest: The Base That Does Not Rot

A knowledge base is not a monument you build once and admire. It is a living asset that rots if you ignore it, and a rotting knowledge base is dangerous precisely because it still sounds authoritative. The process changes: a new die set goes in, a supplier switches, a machine gets rebuilt, a procedure gets revised. The moment that happens, some validated entries become wrong, and an entry that was correct in September and serves a confidently wrong answer the following March is worse than no entry at all, because the crew has learned to trust it.

The discipline that prevents rot is lightweight but non-negotiable. Every entry carries a version, a validation date, and the conditions it depends on, so when those conditions change you can find and review the affected entries instead of hoping someone remembers. When the crew uses an entry, they get a one-tap way to flag it: "this did not match what I found." Those flags are the early-warning system that an entry has gone stale, and they route to the engineer who owns that area for a quick re-validation. New saves and new failures feed back in: when a tech solves a problem the base did not cover, that solution becomes a candidate entry, validated and added, so the base grows from the crew's own experience and not just from the retirees. The base that keeps learning is the base that keeps paying.

The payback the graduate defends to a plant manager is concrete and durable. The upfront cost is small: a dozen hours of the expert's time, the validation effort of a senior tech and an engineer, and the serving software. The return compounds. The 18 months of rediscovery-by-scrap that would have followed Dave's departure is largely avoided. A green crew gets to competence faster, which directly addresses the talent cliff that 85 percent of manufacturers say is hurting quality. And every prevented escape, here a single press defect containment can exceed 20,000 dollars, is a traceable win against a small, one-time capture cost. The knowledge base is the rare investment that keeps paying after the person who seeded it has retired, and on a floor losing its Daves faster than it can hire, it is the highest-return AI move available.

Key Takeaways

• The talent cliff makes knowledge capture the highest-return AI move a thinning plant can make: roughly 2 million workers need reskilling against about 500,000 unfilled roles, and 85 percent of manufacturers say shortages are already hurting quality. AI here is a knowledge multiplier for a thinner, greener crew, not a replacement for the expert.
• The pipeline is capture, validate, serve, and it must be treated as a pipeline. Skip capture and you lose the knowledge; skip validation and you mass-produce a myth; skip serving and you have an archive nobody opens.
• Capture runs two channels: AI-assisted structured interviews that surface what the expert knows he knows, and grounded mining of travelers and CMMS history that surfaces what he did so often he forgot it was knowledge. Twelve hours of interviews produced 140 candidate items against an alternative of about 18 months of rediscovery-by-scrap.
• Validation is the stage that separates a knowledge base from a liability. Test each item against the data, against a second expert or the engineering source, and against whether it still applies. In the example, 6 of 140 items were flatly contradicted by the data, including a defect blamed on operators that actually tracked an upstream die-wear cycle.
• Keep a human accountable for every validated entry, with a name, a date, and a source, because the customer audits the plant and "the knowledge base said so" is no more defensible than "the model flagged it."
• Serve with retrieval-augmented generation so the assistant answers only from validated entries and cites the source. It cites or it refuses, it stays advisory on anything that moves or hurts (78 percent of plants cannot centrally monitor their OT network), and it is fast enough to beat a phone call at 2 a.m.
• A knowledge base rots if you ignore it, and a rotting base is dangerous because it still sounds authoritative. Version every entry, tag its conditions, give the crew a one-tap stale flag, and feed new saves back in so the base learns from the floor and not just from the retirees.
• The payback compounds and outlives the expert: a small one-time capture cost against avoided rediscovery, faster green-crew competence, and prevented escapes that can each exceed 20,000 dollars, all traceable, which is the credential the graduate puts in front of a plant manager.