Tacit Knowledge

The know-how born of experience and intuition — the kind Michael Polanyi famously distilled as:

"We can know more than we can tell."

Named as a first-order enterprise-AI problem in Teaching AI How People Work Is Fraught with Problems (Economist) — the missing layer of context that neither rules (explicit) nor process-mining (data-derived) reaches.

Why it matters for enterprise AI

The vault's Context Engineering and Data Democratisation threads treat context as something already existing that needs plumbing. Tacit knowledge is the exception: it exists only in workers' heads and hands, and getting it out is the harder problem.

The Polanyi frame reframes many stalled enterprise AI projects: not "the model isn't good enough," not "we don't have data governance" — but "the context that would make the model useful hasn't been extracted from the humans yet, and might resist extraction."

The three-tier taxonomy (Economist framing)

Explicit knowledge — rules, guidelines, SOPs. Easy to codify upfront.
Data-derived knowledge — process mining (Celonis reads ERP logs to see how invoices actually unfold vs. how the SOP says they should); corpus mining (brand voice from marketing copy).
Tacit knowledge — Polanyi's residual. "We can know more than we can tell."

Most enterprise AI plays live in tier 2. Tier 3 is where the durable advantage — and the political friction — sits.

The four routes to capture tacit knowledge (Economist)

Route 1 — Absorb from the corpus

Enrique Ide (IESE): AI's defining property is identifying patterns humans can't describe. Existence proofs: facial recognition, chess.
Trap: works when the corpus embodies the tacit knowledge (a brick wall encodes bricklayer knowledge). Fails when it doesn't (staring at brick walls doesn't teach you how to lay one).

Route 2 — Video-based embodied capture

Monumental (Dutch startup, bricklaying robots) — canonical case.
Interviewed masons: got "That's the way I've always done it." — pure Polanyi.
Hours of video revealed the tacit micro-move: masons vibrate their hands slightly when pushing brick into mortar, forcing mortar into the brick's pores for a stronger bond. Monumental's robots now replicate this.
Generalises to embodied work. For cognitive work, see route 3.

Route 3 — Fine-grained keystroke/screen tracking

Existing precedent: call centres (recorded calls, screen monitoring).
The frontier: Meta's Model Capability Initiative — track keystrokes and mouse clicks of Meta employees to train AI. Noisy employee pushback.
The Danielle Li (MIT) survey finding: workers believe they hold large amounts of uncodified knowledge and can withhold it. When shown a video of how their data will train AI, they became less willing to sell survey data — the political-economy problem of tacit-capture.
The information-asymmetry lock: workers know they're the substrate; the substrate is strategically withheld.

Route 4 — Expert evaluation of AI outputs

Some tasks self-verify — "a piece of code either runs or it does not" (see Code Is Free).
Fuzzier tasks (aesthetic judgment, research-brief quality) require human experts to rate AI output, which becomes the training signal.
Operationalised at scale: the EI Survey in Anthropic Economic Index Cadences Report (June 2026) — post-hoc user ratings feed back into training.

The three unresolved questions (durable)

The article's closing punchlist — worth pinning:

Who owns uncodified know-how? Worker vs. employer vs. the AI vendor whose model absorbed it.
How much surveillance is acceptable? Meta's MCI is the current bright line.
How does machine capture of tacit knowledge affect how humans acquire, practise, and pass on expertise? This is the Deskilling Trap (Juniors) question in a different key.

2026-08-10 — Cohen's management-side companion, and the intentionally-uncodified reframe

Seth Cohen (CIO, Procter & Gamble) is the vault's first management-register articulation of the codification thesis: strategic knowledge that used to live in a leader's head, dispensed meeting by meeting, will be written down as language and instructions that both humans and agents can execute against. This is the Polanyi problem posed as a management skill (see the Manager as Translator concept page).

The load-bearing addition to this concept comes from Raj Kumar's counter-comment on that same post, filed as Managerial Ambiguity as Coordination: a subset of what looks like "uncodified tacit knowledge" is not residual but intentionally uncodified because ambiguity is doing coordination work (letting managers vary the resolution of contradictory objectives across weeks/clients/quarters without formally deciding). This complicates the four-routes-to-capture framing above:

Routes 1–4 assume the extract-target is residual know-how the worker hasn't articulated.
Kumar names a fifth failure mode: some uncodified knowledge is load-bearing precisely because it's not codified. Extracting it (via any of the four routes) imports the fixed-rule failure mode into every downstream execution.

Practical implication for codification projects: identify which uncodified knowledge is serving a variable-tradeoff function before extracting it, and either (a) leave those items alone, (b) codify them as parametric-with-explicit-escalation rules, or (c) accept that codification will surface previously-metabolised contradictions that then have to be resolved once, at the top, in writing.

Cross-references

Context Engineering — the fourth lens: substrate below Keen's runtime pillars, Debois's lifecycle, and Prukalpa's enterprise-substrate view
What an Enterprise Context Layer Is (Prukalpa) — the "context layer" this concept says the substrate for is politically fragile
Data Democratisation — democratising meaning runs into the tacit-knowledge floor
Deskilling Trap (Juniors) — pipeline consequence of AI absorbing the tacit layer that used to be apprenticed
Cognitive Offloading — individual-level twin: the atrophy risk
Code Is Free — Route 4's self-verifying-tasks-first heuristic
Michael Polanyi — the philosopher whose formulation anchors the concept (probably doesn't merit an entity page — mentioned once, historical figure)
Anthropic Economic Index Cadences Report (June 2026) — Route 4 operationalised at scale via the EI Survey
Manager as Translator — management-register companion: the codification thesis stated as a management skill (Cohen, 2026-08-10)
Managerial Ambiguity as Coordination — the counter that reframes a subset of tacit knowledge from residual to intentionally load-bearing

Sources

Teaching AI How People Work Is Fraught with Problems (Economist) — canonical
Orchestrate Business Outcomes with AI and Human Talent (Seth Cohen) — management-side companion and the Kumar counter