An AI agent doesn’t hesitate the way a new hire does. It doesn’t wonder whether a number is safe to use, or whether the table it just queried was quietly retired last year. It answers, confidently, with whatever the schema says, right or wrong: exactly the failure Atlan built its context layer to catch. Michael Polanyi named the underlying gap in 1966: people routinely act correctly on things they can’t fully put into words, the exception a veteran analyst applies without thinking, the judgment call nobody wrote down.[1] Tacit knowledge capture is the discipline of closing that gap before an agent runs into it: surfacing what’s in someone’s head and turning it into context a person can certify and a machine can use. This guide covers that discipline: why AI agents make it urgent, and the methods, old and new, for doing it well.
| Field | Value |
|---|---|
| What it is | Turning tacit knowledge, what people know but never wrote down, into context both people and AI agents can use |
| Why it matters now | AI agents can’t ask a clarifying question, so undocumented context that used to erode slowly now fails loudly, in production |
| Classic method | Exit interviews, apprenticeship and shadowing, communities of practice, decision logs |
| AI-era method | Context agents inferring definitions from query and usage patterns, then routing the judgment call to a human to certify |
| Real example | Meta: 50+ AI agents scanned 4,100+ files, taking AI-context coverage from 5% to 100% |
| Common mistake | Treating capture as a one-time documentation sprint instead of a continuous certification loop |
What is tacit knowledge capture?
Permalink to “What is tacit knowledge capture?”Tacit knowledge capture is the deliberate practice of surfacing knowledge a person holds but has never written down, and putting it into a form other people, and now AI agents, can use. Polanyi’s original example is recognizing a face in a crowd: you know instantly, and you’d struggle to list the exact combination of features that told you.[1] The same pattern shows up in data work: an engineer who spots a broken pipeline from the shape of an error log before reading a line of it, or an analyst who knows a dashboard number “looks wrong” before she can say precisely why.
The practice sits underneath two ideas that get used almost interchangeably, and shouldn’t be.
Tacit knowledge vs. tribal knowledge vs. institutional knowledge
Permalink to “Tacit knowledge vs. tribal knowledge vs. institutional knowledge”Tribal knowledge is the organizational subset of tacit knowledge: the specific undocumented exceptions, historical decisions, and workflow shortcuts inside one company. Atlan’s guide to tribal knowledge covers the five types and the production failures each causes; this page stays focused on the mechanics of getting knowledge out of someone’s head in the first place, not on cataloging what it looks like once it’s already lost. Institutional knowledge loss is the cost side of the same coin: what happens, in dollars and onboarding time, when tacit knowledge walks out the door with a departing employee. Atlan’s institutional knowledge loss piece owns that number, so it isn’t repeated here. What follows instead is the practice: how capture actually happens, method by method.
What is the difference between tacit and explicit knowledge?
Permalink to “What is the difference between tacit and explicit knowledge?”Tacit knowledge is know-how someone can perform but struggle to fully explain. Explicit knowledge is the opposite: already written down, in a doc, a wiki, a semantic layer’s metric definition, in a form someone else can read without asking the person who wrote it.
The bridge between the two, turning tacit into explicit, is what Ikujiro Nonaka and Hirotaka Takeuchi called Externalization, one of four conversion modes in their SECI model.[2][3]
The four SECI modes, briefly
Permalink to “The four SECI modes, briefly”- Socialization: tacit to tacit, through shared experience and observation
- Externalization: tacit to explicit, through concepts, metaphors, and documents
- Combination: explicit to explicit, through combining and reorganizing existing documents
- Internalization: explicit to tacit, through practicing what’s been written down until it becomes instinct
SECI is worth knowing, not worth over-indexing on. The model assumes knowledge flows person to person, and a 2006 critique by Stephen Gourlay, cited on the same Wikipedia page, argues it doesn’t explain how genuinely new ideas form and treats a messy process as a tidy four-stage sequence.[2] Useful as one lens, not the whole frame, especially now that the “person” receiving Externalization is sometimes an agent. Once externalized, where knowledge lives structurally is its own question; see knowledge architecture for AI agents and the types of metadata AI agents need.
The AI Context Stack
A breakdown of the layers between raw metadata and context an AI agent can actually query, and where captured tacit knowledge fits in that stack.
Get the BriefWhy does tacit knowledge capture matter more now that AI agents are involved?
Permalink to “Why does tacit knowledge capture matter more now that AI agents are involved?”A new hire who hits an undocumented exception asks a colleague. An AI agent hits the same exception and infers, often wrong, and with total confidence. Meta’s engineering team put this plainly after mapping tribal knowledge across its own codebase: “Without this context, agents would guess, explore, guess again and often produce code that compiled but was subtly wrong.”[4]
Ewa Zborowska, IDC’s research director for AI in Europe, argues most enterprise AI scaling failures trace back to exactly this: ignored tacit knowledge, not model quality, invoking Polanyi directly.[5] According to IDC (2026), the numbers back her up: 89% of organizations report ongoing data-quality problems, 52% call data quality the single most critical factor for AI success, only 6% of CIOs have completed every planned data initiative, and 70% cite data silos as a major adoption barrier.[5]
Teresa Tung, Accenture’s global lead for data capability, frames the upside: “Your next competitive moat will come from designing new systems that capture tacit knowledge and make it explicit.”[6] According to Accenture, one cosmetics company did exactly that with regulatory judgment calls that used to live in a handful of experts’ heads, scaling evaluations from hundreds to more than 40,000 a month while cutting expert workload by roughly 80%.[6]
This is the gap a context layer is built to close. Undocumented context doesn’t just slow agents down; it’s a direct contributor to AI agent hallucination and to definitions that quietly go stale, covered in context freshness.
How do you capture tacit knowledge with classic knowledge-management methods?
Permalink to “How do you capture tacit knowledge with classic knowledge-management methods?”Before AI agents entered the picture, organizations already had a toolkit for externalizing tacit knowledge. It wasn’t built to feed a machine, but it’s worth knowing before layering AI-era methods on top of it.
Exit interviews and structured knowledge-transfer sessions capture a real snapshot: what a departing expert can articulate in a room, once, before they leave. It’s one-time, though, and it only captures what the person remembers to mention. Deloitte’s Eyal Cahana and Evan Siegel found that 92% of organizations still fail to consistently capture knowledge from retiring employees, even with this method widely available.[7]
Apprenticeship and shadowing work differently. This is Socialization in SECI terms: tacit knowledge passed tacit-to-tacit through watching and doing, not writing anything down. It transfers judgment well. It produces nothing machine-readable, by design.
Then there’s the paper trail: communities of practice, decision logs, internal wikis, Externalization in its plainest form. Someone writes down what they know, after the fact. The problem isn’t the writing. It’s the shelf life. A wiki page goes stale the moment the underlying process changes, and nothing forces a rewrite.
When classic methods still make sense
Permalink to “When classic methods still make sense”Structured interviews and decision logs still earn their keep in succession planning, in regulated environments that require a documented decision history, and where no query or usage trail exists for an AI system to infer from. They’re a baseline, not a replacement for what AI-era capture adds next. That knowledge still needs somewhere to live that an agent, or a knowledge assistant, can query later; see how to build a knowledge base for AI agents and why systems of record were never built to hold it.
Context Maturity Assessment
Score how much of your organization's tacit knowledge has already been captured and certified, versus how much still lives only in someone's head.
Assess Your Context MaturityHow do AI agents capture tacit knowledge automatically?
Permalink to “How do AI agents capture tacit knowledge automatically?”The fastest-growing capture method skips the interview: it reads what people did, their SQL, dashboards, and ticket history, and asks a human to certify the judgment call, instead of asking them to write from scratch.
Atlan calls this Context Curation: building, refining, and certifying a company’s business logic, metric definitions, and policies, as human-on-the-loop work, not a documentation sprint. When finance and sales define “revenue” differently at the code level, that’s a curation problem: a metrics conflict agent reads all the SQL, surfaces the discrepancy, and one human makes one call that updates context everywhere. Nobody wrote a wiki page; the agent inferred the conflict, and a person resolved it once.
Meta’s engineering team built this at a scale worth naming in full: 50+ specialized AI agents read 4,100+ files across four repositories and three languages, producing 59 compact context files that took the company’s AI-context coverage from 5% to 100% and documented more than 50 non-obvious patterns.[4] In Meta’s preliminary testing: roughly 40% fewer agent tool calls per task, because agents stopped guessing and started reading certified context.[4]
| Method | How it works | What it captures well | What it misses for AI agents |
|---|---|---|---|
| Exit interviews / knowledge-transfer sessions | Structured conversation before departure | Deep expertise, decision history | One-time snapshot, doesn’t update |
| Apprenticeship / shadowing | Socialization (tacit-to-tacit) | Judgment, pattern recognition | Doesn’t produce anything machine-readable |
| Wikis / decision logs | Manual write-up after the fact | Explicit process steps | Goes stale the moment the process changes |
| Context agents reading usage patterns | AI reads SQL, dashboards, ticket history, surfaces conflicts | Definitions actually in use, at the moment they diverge | Needs a human to certify the judgment call |
| Glossary certification workflows | A steward reviews AI-drafted definitions before they ship | Governance and accountability | Only as good as the steward’s review cadence |
Glossary certification workflows
Permalink to “Glossary certification workflows”The role this leaves for a human steward is different from the one it replaces: a queue, not a document. Atlan’s Context Engineering Studio reads what exists, metric definitions, catalog metadata, query history, BI dashboards, and drafts a semantic layer domain experts refine rather than write from scratch. Stewards used to write documentation; now they certify it: sampling AI-drafted definitions, resolving the ones an agent flags as ambiguous, and storing it in versioned, domain-scoped context repos so every agent reads the same definition.
The mechanism has names worth knowing: context mining for extraction, context engineering techniques for making it agent-consumable, the context development lifecycle for how a definition moves from draft to certified. Certified definitions land in systems of semantics, governed like metadata management for AI governs everything else, in a versioned context repository, tested before use. Agent Skills show the pattern plainly: a SKILL.md file, one engineer’s procedure, written down and governed, surfaced through semantic search and retrieval orchestration.
What goes wrong when tacit knowledge capture is treated as a one-time project?
Permalink to “What goes wrong when tacit knowledge capture is treated as a one-time project?”Most tacit knowledge capture efforts fail the same way: they run once, as a sprint, instead of running continuously as new decisions and exceptions pile up. The knowledge-management industry has had decades to solve this with conventional tooling, and it hasn’t. Ninety-two percent of organizations still fail to consistently capture knowledge from retiring employees.[7] Deloitte projects $6.9 trillion to $9.6 trillion in lost economic output as the current retirement wave compounds the problem, though that wave is one driver among several, not the AI-specific one.[7]
There’s a steel-man worth taking seriously. Harang Ju, of Johns Hopkins Carey Business School and MIT’s Initiative on the Digital Economy, argues capturing knowledge without redesigning the workflow around it doesn’t help: “It is easy to buy the motors. It takes real work to redesign the factory.”[8] His four tests for what should stay human-owned, checkability, stakes, judgment, duration, guard against automating every judgment call because an agent can now infer one.
The AI-specific failure mode compounds the KM one: a definition gets captured once, then quietly drifts out of date, and an agent keeps citing the stale version with full confidence because nothing told it otherwise, the problem context freshness exists to catch. The fix isn’t a bigger documentation sprint. It’s treating certification the way long-term context management treats memory: ongoing, with context noise filtered out and portability built in, so a certified answer travels with the agent instead of living in one team’s tool, an argument harness engineering makes from the runtime side of the same problem.
AI Agent Context Readiness Checklist
A short checklist for whether your organization's tacit knowledge is captured, certified, and ready for an agent to query, or still living in someone's head.
Check Your ReadinessCapture is a certification loop, not a documentation sprint
Permalink to “Capture is a certification loop, not a documentation sprint”Polanyi wrote “we can know more than we can tell” in 1966 to describe a human-to-human problem, knowledge passed imperfectly from one head to another over years. AI agents changed who’s on the receiving end, and the economics of getting it right: what used to erode slowly enough for a hallway conversation to patch it now fails the first time an agent queries it. The organizations doing this well have stopped treating capture as an event and started treating it as a certification loop that runs continuously. Most, per the 92% retiree-knowledge stat above, still haven’t made that switch. See Atlan’s context layer in action, or explore context engineering and the semantic layer it depends on.
FAQs about tacit knowledge capture
Permalink to “FAQs about tacit knowledge capture”1. What is the difference between tacit knowledge and explicit knowledge?
Permalink to “1. What is the difference between tacit knowledge and explicit knowledge?”Tacit knowledge is know-how someone can perform but struggle to explain, like recognizing a face or debugging a familiar pipeline by instinct. Explicit knowledge is already written down, in a doc or a certified metric definition, readable without asking the person who wrote it.
2. What is the SECI model?
Permalink to “2. What is the SECI model?”The SECI model, from Ikujiro Nonaka and Hirotaka Takeuchi, describes four modes of converting knowledge: Socialization (tacit to tacit, through observation), Externalization (tacit to explicit, through documentation), Combination (explicit to explicit, through synthesis), and Internalization (explicit to tacit, through practice). Tacit knowledge capture is mostly Externalization.
3. Why is tacit knowledge important for AI?
Permalink to “3. Why is tacit knowledge important for AI?”AI agents can’t ask a colleague why a table looks canonical but is actually deprecated, the way a new hire can. Undocumented context that used to erode slowly now produces a confident, wrong answer the first time an agent queries it. That’s an AI-agent infrastructure problem now, not only an HR one.
4. How do you capture tacit knowledge in an organization?
Permalink to “4. How do you capture tacit knowledge in an organization?”Classic methods, exit interviews, apprenticeship, decision logs, communities of practice, still work but produce a one-time, human-readable snapshot. AI-era methods have agents infer definitions from usage patterns, then route ambiguous cases to a human to certify. Who does that certifying is a governance question worth deciding deliberately.
5. What is tribal knowledge, and how is it different from tacit knowledge?
Permalink to “5. What is tribal knowledge, and how is it different from tacit knowledge?”Tribal knowledge is the organizational subset of tacit knowledge: undocumented exceptions, historical decisions, and workflow shortcuts inside one company, rather than tacit knowledge in general. All tribal knowledge is tacit; not all tacit knowledge is tribal.
6. What happens when tacit knowledge is lost?
Permalink to “6. What happens when tacit knowledge is lost?”The knowledge doesn’t just disappear, it stops being usable the moment someone leaves or forgets it: onboarding slows, decisions get remade from scratch, and any AI agent grounded on the surrounding data starts producing outputs that are technically correct but organizationally wrong.
7. Can AI agents learn tacit knowledge on their own?
Permalink to “7. Can AI agents learn tacit knowledge on their own?”No. Agents can infer patterns from query history, dashboards, and usage logs, and do this increasingly well, but inference isn’t judgment. The ambiguous cases, like which of two conflicting metric definitions is correct, still need a human to certify. That’s the loop this page describes, not a one-time handoff.
Sources
Permalink to “Sources”- Polanyi’s paradox, Wikipedia
- SECI model of knowledge dimensions, Wikipedia
- Nonaka, I. & Takeuchi, H., The Knowledge-Creating Company, Oxford University Press (1995)
- How Meta Used AI to Map Tribal Knowledge in Large-Scale Data Pipelines, Meta Engineering (April 2026)
- The Knowledge Your AI May Never Have, Ewa Zborowska, IDC
- Tacit Knowledge Is Your Next Competitive Moat, Teresa Tung & Philippe Roussiere, California Management Review (March 2026)
- Capturing Institutional Knowledge, Eyal Cahana & Evan Siegel, Deloitte Insights
- The Hidden Cost of AI Agents for Companies Is Lost Expertise, Harang Ju interview, MIT Sloan Management Review Middle East
