"If it lives in your head, it should live in the 'system.'"

An estimator opens a set of drawings from an architecture firm he has never worked with, and it takes him four hours to find things. He opens a set from a firm he has estimated a dozen projects for, and it takes him forty minutes, and the difference is not the drawings. The drawings are equally complete both times.

What he has, on the second set, is a model of how that particular firm thinks. Where they put the door schedule. Whether the glazing types live on the plan or in a separate sheet. Which dimensions they consider obvious enough to omit. Whether their revision clouds mean what revision clouds usually mean. None of that is written anywhere. He did not learn it from a document, because no document exists. He learned it by doing twelve projects and getting surprised eleven times.

Ask him to write it down and you get a page and a half of generalities that would not save anybody four minutes, let alone three hours. He is not withholding. He does not have access to it in that form. The knowledge is real, it is worth a great deal of money, it demonstrably exists because his estimates come in faster and cleaner, and it cannot be produced on request as text.

Now he takes another job.

Everything in the paragraph above leaves with him, and nothing in the company registers the loss, because there was never an artifact to notice was missing. His replacement starts from zero on twelve architecture firms and the company experiences this as the new person being slow.

That is the situation my own ism exists to prevent. If it lives in your head, it should live in the system. Agreed. The question the sentence does not touch, and which this entire chapter turns out to be about, is whether the thing in his head is the kind of thing that can live anywhere else.

Stated as a rule:

The variables are a piece of knowledge, the person holding it, and a store that is not that person. The store is the system: the record, the database, the document, the process definition, the place a second person can go without asking the first.

The quantity minimized is the number of facts whose only copy is in somebody. That is countable in principle, and the way you count it is unpleasant: pick a person, imagine they are gone tomorrow, and list what nobody can do. Every organization I have worked in can produce that list in about ten minutes and has never written it down.

The failure signal is the question that has exactly one possible recipient. When a thing can only be asked of one person, that thing lives in a head. The signal is available for free, all day, in any organization, and nobody collects it.

The domain of validity is the class of knowledge that can be externalized at all, which the original sentence assumes is everything and which is the chapter's whole problem.

Two words are doing more than they look. Lives is the right verb, because knowledge in a head is not stored the way a file is stored; it is maintained by use, it decays, it updates without announcement. And system in quotation marks was my hedge from the beginning. I meant the record of the business, the place where a fact is available to whoever needs it. Quotation marks because I have never been able to say precisely what qualifies, and a document nobody can find is not a system in any sense I care about.

Peter Drucker gave the reason this became urgent in the first place, and he gave it decades before most organizations noticed.

His observation was that the productive asset in modern work is knowledge, that it is owned by the worker rather than by the firm, and that it therefore walks out of the building every evening and may not come back. That was a genuinely new problem. A factory owns its machines. A firm whose value is in what its people know owns nothing in the relevant sense, and the arrangement holds only as long as the people stay.

Which sets up the question the rest of this chapter is about: if the asset is knowledge, can it be transferred to something that does not resign?

Michael Polanyi answered no, or rather answered that a large part of it cannot, and his formulation is the one everybody quotes because it does not need explaining. We can know more than we can tell.

His example is a face. You can pick a person you know out of a crowd of thousands, instantly, reliably, and you cannot describe the arrangement of their features well enough for anyone else to do it. The recognition is not vague. It is precise, fast, and correct. What it is not is available to you as a description, because you never learned it as one. You learned it by seeing the face, which is a different mechanism with a different kind of output.

Polanyi's claim is not that some knowledge is hard to write down. It is that tacit knowing is a dimension of all knowing rather than a separate compartment, and that the part which cannot be told is doing load-bearing work in the part that can. When you follow a written instruction, something not in the instruction is telling you how to read it.

David Autor named this Polanyi's paradox in 2014 and pointed it at automation. His argument is that the tasks resisting machine substitution are precisely the ones governed by knowledge nobody can articulate: common sense, flexibility, adaptability, judgment. And his conclusion is a correction to a mood rather than a technical result. The extent of machine substitution for human labor, he holds, has been overestimated by journalists and expert commentators, and the reason is that the hard part was never the computation. It was that the rule to be computed cannot be written down by the person who runs it.

Against that stands the most influential attempt to say tacit knowledge can be converted. Ikujiro Nonaka and Hirotaka Takeuchi proposed that knowledge moves through four modes: tacit to tacit by shared experience, tacit to explicit by externalization, explicit to explicit by combining, and explicit back to tacit by practice until it becomes second nature. The middle one is the one if it lives in your head, it should live in the system needs. Externalization is the claim that what is in a head can be brought out into a form that does not require the head.

That model has been the foundation of corporate knowledge management for thirty years.

Not well, and this is the weakest evidential position of any ism in the creed.

Stephen Gourlay published a critique in 2006 that goes at the foundation rather than the details. Three of the four conversion modes, he allows, appear plausible. His finding is that none of them are supported by evidence that cannot be explained more simply, which is a specific and damaging claim: the phenomena the model is built to explain do occur, and the model is not doing the explaining. Simpler accounts cover the same observations.

His second charge is worse for my purposes. He argues the framework omits inherently tacit knowledge, which is to say it handles the knowledge that could have been written down anyway and quietly leaves out the part Polanyi's tacit dimension was actually about. If that is right, then the theory of externalization the creed leans on succeeds by narrowing its subject until the hard case is outside it.

So the honest statement of where this chapter stands: the book's title ism rests on a literature whose central model is not empirically supported and may misread the thinker it draws its key concept from.

Two things keep it from being a rout.

The first is that the specific mechanisms are better supported than the grand model. That documentation reduces single-point dependency is not really in dispute; it is close to a definitional truth and it is confirmed by every organization that has lost a person and discovered what they could not do. The strong claim that fails is not writing things down helps. It is tacit knowledge can be systematically converted into explicit knowledge by a managed process, which is a much bigger claim and the one the field built its consultancies on.

The second is that Polanyi's argument, correct as it is, is about a boundary rather than about the whole. A great deal of what sits in people's heads in an ordinary company is not tacit in Polanyi's sense at all. It is explicit knowledge that simply never got recorded: which vendor to call, what the exception is for that customer, why that field exists. None of that is ineffable. It is just undocumented, and the reason is ordinary friction rather than the structure of human cognition.

Conflating those two is the error I have made for years. I have used Polanyi's paradox as a reason a capture effort failed, when the actual reason was that the form had eleven required fields.

The design problem where I had to take all of this seriously was building a place for that knowledge to go.

The premise was simple and the premise was the easy part: people in the company know things that exist nowhere else, the company would like those things to exist somewhere else, and no amount of asking people to document their work had ever produced anything. That last fact is the important one. It had been tried. It failed the way it always fails, which is not with refusal but with compliance that produces nothing useful, because writing a procedure is a separate skill from having one and nobody has time to acquire it.

Three design decisions came out of that, and each of them is a concession to something in this chapter.

The first was video rather than documents. If a person cannot write down how they read an architect's drawing set, they can be recorded reading one, narrating as they go. That is a different request. It does not ask them to convert anything. It asks them to do the thing they already do, out loud, while a screen recorder runs. The knowledge stays tacit in their head and becomes observable to somebody else, which is not the same as becoming explicit and is worth more than nothing.

The second was cutting the intake. The specification for the submission form had eleven questions on it, each individually defensible, each contributing to a well-organized library. It shipped with eight. Three were removed because every required field on an intake form is a reason not to submit, and a knowledge-capture system that captures nothing is perfectly organized and worthless. That is a direct collision with defining every variable to its deepest identifiable condition, and the collision was resolved against depth, at the intake, on purpose.

The third was the honest time estimate. The ask presented to employees is five to fifteen minutes. Not an afternoon, not a documentation project, not a deliverable with a due date. Five to fifteen minutes, because the alternative to a five-minute video is not a thorough document. It is nothing, forever, and then the person leaves.

None of those three decisions makes the tacit knowledge explicit. Every one of them is a way of getting the knowledge into a place a second person can reach it, which is what if it lives in your head, it should live in the system actually wants and not what its wording says. The system does not end up holding the estimator's model of that architecture firm. It ends up holding forty minutes of him using it, which the next estimator can watch, and from which the next estimator will build their own model faster than twelve projects.

That is socialization with a recording device attached, and Polanyi's argument has no objection to it, because nothing is being converted.

The harder question was what to record, and the answer turned out to be sitting in plain sight.

Asking people what they know produces nothing usable. Everybody knows this and everybody does it anyway, because it is the obvious move and its failure is not obvious: the responses are polite, plentiful, and generic, and they look like data. The reason it fails is that knowing something and knowing that you know it are different states, and the second one is rare for exactly the knowledge worth having. The estimator does not experience his model of an architecture firm as knowledge. He experiences it as the drawings being easy.

The instrument that works is the one named earlier in this chapter and it costs nothing to run. Watch where the questions go. Not what the questions are about, which is a topic and therefore vague. Who they go to. Every organization generates hundreds of these a week, and the routing is already visible in email, in messages, in who gets walked over to. A question that could go to any of six people is not interesting. A question that only ever goes to one is a piece of knowledge with a single copy, and the volume of traffic to that one person is a direct measure of what the company loses when they leave.

Run that for a month and the capture list writes itself, ranked, without anybody having to introspect. It also reliably surprises whoever runs it, because the heaviest traffic is almost never to the person whose expertise the organization talks about. It goes to somebody two levels down who happens to know how one system behaves, and who has never been described as the expert in anything.

What that instrument cannot do is tell the three bands apart. It says where knowledge is concentrated. It says nothing about whether the concentration is explicit knowledge that never got written, a skill that could be recorded, or judgment that cannot leave. That still requires sitting with the person and finding out, and the finding-out is a conversation rather than a form.

The boundary is the class of knowledge that survives leaving a head, and it has three bands rather than two.

The first band is knowledge that is explicit already and merely undocumented. Vendor contacts, exceptions, the reason a rule exists, who to call. This is the overwhelming majority by volume, it is fully covered by the rule, and it fails to get recorded for reasons of friction and habit rather than anything deep. Most of the value available here is in this band, and most of the conversation about tacit knowledge is a distraction from it.

The second band is knowledge that is tacit but demonstrable. The estimator's model of an architecture firm. How somebody handles a difficult call. What a person looks at first when they open a file. This cannot be written by its holder and can be captured by recording the holder doing it. The artifact is not the knowledge; it is evidence from which a second person can rebuild the knowledge, faster than the first person built it. The rule applies here in an amended form: it should live in the system as a demonstration rather than as a statement.

The third band is knowledge that is tacit and not demonstrable, because the performance does not reveal the discrimination. Judgment about people. The sense that a project is going wrong before anything measurable has gone wrong. Taste. Here Polanyi's tacit dimension wins completely, and the rule is not merely difficult to apply, it is a category error. The correct response is the one my own boundary ism gives: leave it manual, keep the person, and make sure they are talking to somebody who will inherit it the slow way.

The single most expensive mistake available in this area is treating band three as band one and building a form for it.

The second most expensive is misjudging which band something is in, and the misjudgment runs in one direction. Almost everything looks like band one from the outside, because the person holding it can always produce some description, and a description is what band one looks like. The estimator's page and a half of generalities is a band-one-shaped artifact produced by band-two knowledge, and an organization that collects it will file it, feel finished, and have captured nothing.

The test I have settled on is whether a competent stranger could act correctly from the artifact alone, with no access to its author. Not whether they would understand it. Whether they would do the right thing. That is a harsher standard than it sounds and most documentation fails it, including documentation that reads well.

Three things this rule cannot see.

It cannot see the cost of the store. Getting a fact out of a head and into a system creates an artifact that must be found, trusted, and maintained, and an out-of-date document in a searchable place is worse than nothing, because it answers with authority. The original prices the capture and not the custody, which is the same blind spot the chapter on defining every variable found, arriving from a different direction.

It cannot tell a fact that was recorded from a fact that was recorded and is still true. A system full of captured knowledge decays at a rate nobody measures, and the decay is silent for exactly the reason the capture was valuable: nobody else knows the subject well enough to notice.

And it is silent about what the capture does to the person. A worker whose knowledge is the reason they are needed has a rational, non-malicious interest in not putting all of it in a system, and pretending otherwise makes every capture program a negotiation conducted in bad faith by one side. I do not have a clean answer to that. The best I have is that the estimator who has recorded twelve firms is more valuable rather than less, because the recordings are evidence of a capability, and that argument is true and does not feel true to the person being asked.

Here is the case against, at full strength.

Polanyi's claim is not that tacit knowledge is hard to externalize. It is that the tacit dimension is a feature of all knowing, that explicit knowledge is a surface resting on it, and that the surface does not function without the part underneath. Read that seriously and it does not carve out an exception to if it lives in your head, it should live in the system. It denies the premise.

Take the estimator's forty-minute recording. What the next estimator gets is a performance, and to learn from a performance they have to know what to attend to, which is itself tacit, which is why two people watch the same video and one of them extracts a method and the other extracts nothing. The recording did not transfer the knowledge. It gave a person who already had the relevant tacit capacity a faster route to building more of it. Which means the system is not holding knowledge. It is holding material that is inert until a sufficiently skilled human being interprets it, and the sufficiency is the thing that was supposed to be captured.

Follow that all the way down and every artifact in every system has this property. A written procedure is a set of marks that produces correct action only in somebody who already knows how to read it in context, including knowing when it does not apply, which the procedure never says. The organization believes it has moved knowledge into the system. What it has moved is a prompt that works on people who already have the knowledge and fails on people who do not, and the failure looks like the new person being careless.

And the harshest version, which I think is the real objection: if it lives in your head, it should live in the system is a proposition about ownership dressed as a proposition about efficiency. It says the firm should hold what the worker holds. Drucker's observation was that the knowledge worker owns the means of production; the creed's answer to that is to move the asset, and the honest translation is that the asset should be moved somewhere it cannot resign. Every capture program is an expropriation with a productivity rationale, and the people being asked to fill in the form understand this perfectly well, which is why the forms come back empty.

I concede the third objection almost entirely, and it changes how the rule should be applied rather than whether.

The asymmetry is real. A capture program run as an extraction produces exactly what extraction produces, which is compliance without content, and I have been on the wrong side of this. What makes it defensible is not an argument about the firm's interests. It is that the person filling in the form has to get something, and the company will be more resilient is not something. Faster onboarding of the people they will otherwise spend their own time answering, credit that is visible, and relief from being the only possible recipient of a recurring question are all things. Whether they are enough varies by person and I would not claim otherwise.

On Polanyi's main argument I concede the strong reading and keep a narrower claim, and the narrowing is the one the design arrived at before the reading did.

The system cannot hold knowledge. It can hold artifacts, and artifacts are inert without an interpreter. That is correct and it is not fatal, because the alternative is not a system that holds knowledge. The alternative is nothing, and a new estimator rebuilding twelve architecture firms from scratch over three years. The artifact does not have to transfer the knowledge. It has to lower the cost of rebuilding it. Judged against transfer, every knowledge-management effort in history has failed. Judged against rebuild cost, they routinely succeed, and the field has spent thirty years claiming the first thing while delivering the second.

So the sentence means something different from what I have been saying. Not the system should contain what is in your head, which is false and which Polanyi has shown to be false. Rather: what is in your head should leave a trace outside it that lowers the cost for the next person to build their own version. A procedure is such a trace. A recording is a better one for anything demonstrable. A conversation with a successor is the best one available for the third band and is not a system at all, which my sentence has no room for and should.

The band structure is the practical instrument, and it is the part I would want carried out of this chapter. Explicit but unrecorded: write it down, and most of your available value is here. Tacit but demonstrable: record the doing, never ask for the telling. Tacit and not demonstrable: keep the person, and make sure somebody is standing next to them.

What survives:

The original assumed everything in a head is transferable to a store, which Polanyi's tacit dimension has spent sixty years showing is false, and which the empirical record of knowledge management confirms by its results. The revision keeps the instruction and gives up the premise. It also gives up the tidiness: one of the three bands does not resolve to a system at all, and any version of it that has no room for keep the person is a version that will quietly destroy something it cannot see.

Three things go forward. The capture-versus-custody problem is the same decay hole Chapters 3 and 7 both reached, now from a third direction, and Chapter 16 is inheriting a great deal on this point. The ownership asymmetry is conceded and not solved. And the claim that recordings lower rebuild cost is one I believe from the design work and cannot support with evidence, because the program that would measure it would have to run for years and nobody has run it.