Let’s talk about the old man who could hear the machine about to fail.
He never wrote it down. He couldn’t. If you asked him, he’d say something useless like “it sounds wrong,” and then he’d go and adjust a thing you didn’t know was adjustable, and the noise would stop, and you’d stand there feeling slightly cheated.
That man is the most valuable person in the building and the worst documented. And this is the week I want to put a number on him.
Picture a workshop. Could be a farm shed outside Bethal, could be a fabrication shop in Germiston, could be an ICU, could be a lawyer’s office at eleven at night. Somewhere in it is a person who has done one thing so many times that the doing has moved out of their head and into their hands.
They can hear a bearing going. They can feel moisture in grain with a closed fist. They can tell from how a weld pulls whether the steel was too cold when it was tacked. They know which client is lying, not because of any tell, but because they’ve been lied to by that exact tone of voice before.
None of that is in a manual. It was never in a manual. And I want you to sit with why.
The philosopher Michael Polanyi put it in one line sixty years ago: we can know more than we can tell. That’s the whole problem in six words.
Knowledge comes in two kinds. Explicit knowledge is everything that can be written down — the procedure, the spec, the checklist, the torque value, the drug interaction, the code comment. Explicit knowledge is a document. You can copy it, email it, lose it, paste it into a model. It is finite and it is already mostly public.
Then there’s the other kind. Tacit knowledge. The judgement that lives in the body and in the pattern-matching that runs below language. It’s not hidden because someone is protecting it. It’s hidden because the person carrying it genuinely cannot get it out in words. The moment they try, they distort it.
Now here’s the thing nobody selling you the future wants to say out loud.
Every AI system on earth is trained on the explicit layer. It has to be. The explicit layer is the only thing that was ever written down, recorded, typed, timestamped, committed to a repository. Books. Papers. Tickets. Chats. Transcripts. Documentation. All of it.
That’s not a bug you fix by throwing more data at the model. The data doesn’t exist. It was never captured, because it was never capturable by writing.
So you have a machine that has read every manual ever written, and knows nothing about the moment the manual stopped being true.
That’s the gap. And here’s what we’re doing with the gap. We’re filling the explicit layer — fast, brilliantly, cheaply — and then rearranging the actual work so that the conditions that used to transfer the tacit layer are quietly being dismantled.
Think about how it actually moved, that second kind of knowledge. Not through instruction. Through proximity, over years, under someone with authority over you.
You stood next to a person while they worked. You watched the wrong way first, and then you watched the fix. You asked a question at the exact second it mattered, and got an answer you couldn’t have understood a week earlier. You were slow, and in the way, and somebody let you be, because somebody once let them be.
That’s the mechanism. It’s not a course. It’s not a video. It’s time, in the same room, with a person who already has the feel for it.
And now look at what we’ve done to the room.
The expert is off-site. The work happens on a screen. The junior’s first move — the reflex, the one nobody has to teach because it’s now automatic — is to ask the model instead of the person. And the model gives a fluent, confident, well-structured answer that is correct at the level of the document and slightly off at the level of the thing in front of you.
Fluency is the trap. Fluency reads exactly like mastery. And it will keep reading like mastery right up until the moment something fails in a way the document never covered.
Here’s what people are missing.
They think the loss will be visible. It won’t be. That’s the cruelty of it. When tacit knowledge leaves a business, nothing goes missing from the shelf. Every procedure is still there. Every manual. Every dashboard. The system still answers every question you put to it.
The loss only shows up years later, as a thing nobody can diagnose. A yield that drifts and nobody can say why. A machine that fails early and the report says “human error” because there’s no category for “the person who could hear it retired.” A company that keeps hitting numbers and loses the ability to build anything new, because nobody left in the room has the feel for the material.
The research on this is blunt. Roughly seventy percent of critical operational knowledge on a typical plant floor is undocumented — it lives in people, not systems. Millions of experienced workers in manufacturing alone are walking out the door over the next decade. One widely cited estimate puts the cost of that knowledge loss at tens of millions a year for a large organisation, through errors, rework, and duplicated problem-solving. And a handover done late — six months before retirement, documents only — routinely loses a fifth to nearly half of what the role used to know.
You can’t hire that back. You can’t buy it back. And you absolutely cannot prompt it back.
Now, I’m not writing an anti-machine piece. Read the thesis properly.
The correct move is to hand the machine the explicit layer, on purpose, and take the leverage. Do that. Every procedure, every spec, every ruling, every post-mortem, every repetitive question your team asks a human being ten times a week — give it to the model and get those hours back. That part is free and it’s stupid not to take it.
The error is a specific one. It’s assuming that because the machine can answer, the machine can teach. Those are different things. An answer is a product. Teaching is a transfer between two nervous systems that takes time and physical presence.
So here’s the operator’s version of this, and it’s the part that pays for itself.
First, map your tacit layer. Not your processes — those are documented. The other thing. Ask who in your business can tell a good one from a bad one by sound, smell, sight, or feel. Write those names down. That list is your real intellectual property and it is not on any balance sheet.
Second, protect co-presence. Time in the same room stopped being an accident and became a design decision. Schedule it. Pay for it. Put the learner next to the person who carries the unspoken thing, and let the senior person be slower than they’d like, on purpose, with that time costed into the job.
Third, make teaching part of the senior role — not a nice-to-have in the last three months before retirement, but a standing duty, with the same weight as delivery. The person who knows something nobody else can do has an obligation attached to holding it. So do you, to build the conditions.
Fourth, record the why, not just the what. Don’t capture the procedure — you already have that. Capture the decision. Ask: when did you break the rule, and why? What did you see that told you to stop? Where was the manual wrong? That’s the part worth keeping, and most of it surfaces only in a conversation, not in an interview form.
And underneath all of it, the principle.
Knowledge was never a download. It was an inheritance. It moved person to person, in a room, at cost, over time — and the people who received it received more than information. They received the standard. They received the caution. They received the reason to care about the difference between good enough and right.
That’s the thing worth defending. The right to learn beside someone. To be the junior standing in the way of a person who is good at the job, allowed to be slow, allowed to be useless for a season, while the feel of it moves from one pair of hands to another.
And the duty on the other side: if you carry something unspoken, hand it over. Out loud. While there is still somebody standing next to you to hear it. Because the one thing the machine cannot inherit is you.
You’ve got this — but only if you’re the one who passes it on.
The documents will still be there in ten years. That’s exactly how you’ll know what you lost.