Teaching AI How People Work Is Fraught with Problems

Section: Business (Bartleby column adjacent) Edition: 2026-06-27 Print headline: Watch and learn URL: https://www.economist.com/business/2026/06/25/teaching-ai-how-people-work-is-fraught-with-problems

The Tacit Knowledge framing of the enterprise-AI adoption stall — the missing dimension that pure context-engineering (data + rules + docs) can't reach. Directly extends Context Engineering and Data Democratisation with the epistemic side that neither has surfaced yet.

The three tiers of enterprise context (the framing)

Explicit — rules, guidelines, policies. Easy to codify.
Data-derived — process mining (Celonis reads ERP logs to see how invoices/procurement actually unfold, not how the SOP says they should). Corpus-derived (brand voice from marketing copy). This is where most enterprise AI plays land in 2026.
Tacit — the Michael Polanyi category: "We can know more than we can tell." Know-how born of experience and intuition. This is what neither rules nor process-mining reach.

The Polanyi frame (the durable concept)

"We can know more than we can tell." — Michael Polanyi

The article treats tacit knowledge as a first-order enterprise-AI problem, not a footnote:

"How can AI learn about work when even humans cannot articulate why they do things a certain way?"

The three routes (and their traps)

Route 1 — Absorb tacit knowledge from the data corpus

Enrique Ide (IESE Business School): one of AI's defining characteristics is its ability to identify patterns in data that humans cannot describe.
Existence proofs: facial recognition, chess — activities models were never explicitly taught.
The trap: "Sometimes the details of a process matter." A brick wall is a corpus of bricklayer knowledge — but staring at brick walls doesn't teach you how to lay one.

Route 2 — Video-based embodied capture (Monumental case study)

The article's most vivid empirical anecdote.

Monumental — Dutch startup, bricklaying robots.
Interviewed masons early on. Answers were "maddeningly vague — 'That's the way I've always done it.'" — canonical Polanyi.
Hours of video footage revealed the tacit micro-move: masons vibrating their hands slightly while pressing brick into mortar, forcing mortar into the brick's pores → stronger adhesive bond.
Monumental's robots now replicate the vibration.

Generalisation: the closer the work is to physical/embodied action, the more likely a video approach captures the missing tacit layer. For cognitive work, the equivalent is Route 3.

Route 3 — Fine-grained keystroke / screen tracking

Existing precedent: customer-service call centres — calls recorded, screens monitored.
The frontier: everyone else. Meta's Model Capability Initiative — programme to track keystrokes and mouse clicks of Meta employees to train AI. Employees pushed back noisily.
The Danielle Li (MIT) survey finding — workers believe they are sitting on large amounts of uncodified knowledge about their organisations. And they believe they can withhold it.
The withholding experiment: researchers offered participants money for their survey data. Those first shown a video on how the data would train AI were less willing to sell. The information-asymmetry becomes economic.

Named as sensitive terrain. This is the political constraint on the Prukalpa-style enterprise context layer: the substrate the layer is supposed to sit on top of is strategically withheld by the workers who own it.

Route 4 — Expert evaluation of AI outputs (the "fuzzy verify" route)

The alternative when route 3 is politically blocked:

Some tasks self-verify — "a piece of code either runs or it does not." (Same argument Ryan Lopopolo makes on why coding was AI-solved first — see Code Is Free.)
Fuzzier tasks — aesthetic judgment, research-brief quality — require human experts to rate AI output on subjective criteria, then progressively refine.
This is what Anthropic's Economic Index's EI Survey is doing at scale for Claude usage: post-hoc quality ratings become the training signal for the tacit layer.

The three unresolved questions (the closing punchlist)

The article's most valuable line — the durable question set:

Who owns uncodified know-how? (Worker vs employer vs the AI vendor whose model absorbed it.)
How much surveillance is acceptable? (Meta's MCI as the current bright line.)
How will machines learning it affect how people acquire, practise, and pass on expertise? — this is the Deskilling Trap (Juniors) question. Brovich's org-level lens (A Leaders Guide to Advanced Team Structures (AWS Events)) is one answer; this Economist piece raises it as the industrial-relations version of the same problem.

Why this is a durable page

Every prior Context Engineering source in the vault (Keen / Debois / Prukalpa) treats context as something that already exists and needs to be plumbed. This is the first source that treats it as something that has to be actively extracted from unwilling and often unwitting humans. The tacit-knowledge frame will recur every time an enterprise AI project stalls without clear reason.

Connections

Tacit Knowledge — new concept page seeded by this article (Polanyi + the three-route taxonomy).
Context Engineering — this article is the fourth-lens extension: tacit knowledge sits below runtime pillars (Keen), lifecycle (Debois), and substrate (Prukalpa).
What an Enterprise Context Layer Is (Prukalpa) — the enterprise-substrate view; this article is the reason the substrate can be politically fragile.
Data Democratisation — the "democratise meaning before access" line acquires an extra layer: the meaning to democratise partly doesn't exist in codified form yet.
Code Is Free — verifiable-tasks-first heuristic; this article is why non-code tasks lag.
Deskilling Trap (Juniors) — the pipeline consequence of AI absorbing the tacit layer that used to be apprenticed.
Cognitive Offloading — individual-level twin of the tacit-atrophy risk.
Anthropic Economic Index Cadences Report (June 2026) — Route 4 (human rating) is what the EI Survey operationalises at Anthropic scale.
Meta (if we ever pin the entity page) — Model Capability Initiative belongs there.