Skip to content
Fritz Gerald ZephirinMenu

Social

Tacit Knowledge Is the Last Moat

4 min read

In 1966, the chemist-turned-philosopher Michael Polanyi compressed a career into one sentence: we know more than we can tell. A doctor recognizes a disease in a face before naming a single feature. A senior support agent reads three lines of a furious email and knows — knows — that this account is saveable and that one is gone. Ask either to write the rule down and you get something true, partial, and useless. The knowledge is real; it just doesn’t survive articulation.

For sixty years this was philosophy. AI made it a pricing question. Large language models are, mechanically, machines for absorbing everything that has been articulated — every documented procedure, every written answer, every policy that made it into text. The boundary Polanyi drew between what we can tell and what we merely know has become the boundary between what can be automated and what cannot. What can be articulated can be trained on. What can be trained on becomes a commodity. So the durable edge of a firm migrates toward exactly the knowledge that won’t go in the document.

You can watch the migration in any operating company, including mine. The documented escalation tree — who handles what, when to refund, which issues page legal — is in the knowledge base, and the AI executes it better than a tired human at 2 a.m. What the AI does not have is the agent who knows that this enterprise client’s procurement lead is the real decision-maker and is having a bad quarter; that this particular phrasing from this particular customer means churn risk, not confusion; that the policy should bend today and hold tomorrow. None of that is in a document. Most of it couldn’t be — not because nobody tried, but because the knowledge is indexed to context in ways its holder can’t enumerate. That residue is increasingly where the margin lives.

Two clarifications keep this from becoming mysticism — and the first one cuts against my own thesis. Most “tacit knowledge” in companies is fake tacit: ordinary documentable process that nobody bothered to write down. Undocumented is not inarticulable; it’s deferred maintenance, and the discipline is to write relentlessly until the genuine residue shows itself. This produces the paradox I find most useful in operations: you discover your real moat by trying to document everything and seeing what’s left. The writable part was table stakes all along; firms that refuse to write it down aren’t protecting tacit knowledge, they’re hiding ordinary knowledge in expensive heads.

Second: the boundary moves. Imitation learning extracts patterns from behavior without anyone articulating them — models learn from demonstrations what no manual captured, and distillation is forced articulation at industrial scale. The honest claim is comparative, not absolute: tacit knowledge erodes slower than explicit knowledge, not never. A moat that erodes slowly is still the best moat available — that has always been the actual standard — but anyone telling you taste is permanently safe is selling something.

The collision worth losing sleep over is with the entry-level problem. Tacit knowledge has exactly one transmission mechanism: apprenticeship — watching, doing, being corrected, accumulating the cases. It cannot be taught by document, by definition. And apprenticeship is precisely what the AI economy is dismantling, by automating the junior work that constituted it. Follow both lines and you reach an uncomfortable place: we are commoditizing explicit knowledge while shutting down the factory that produces tacit knowledge. The firms that notice — that keep humans accumulating cases on purpose, as a designed cost rather than a free by-product — are building the only inventory that will be scarce on both ends.

What this means in practice, compressed. Document everything; the documentation is the floor, and the floor is now executable. Locate the residue honestly — the judgment calls, the client-reading, the exception instincts — and concentrate your humans there, priced as judgment, not as typing. Protect the apprenticeship loop that replenishes it. And when someone shows you an org chart, ask Polanyi’s question of it: which boxes hold knowledge that would survive the person leaving? The ones that wouldn’t are your moat and your fragility, in the same ink.

Sources & further reading