What is tribal knowledge?
Engineering culture and software practice
Tribal knowledge is information an organisation relies on but has never adequately recorded, so it lives in people's memories, habits and informal conversations. The term has roots in manufacturing and quality practice and overlaps with the older idea of tacit knowledge. It becomes risky when essential work depends on knowing unwritten exceptions, history or reasoning. Knowledge silos are the same pattern at group scale, especially when other teams cannot discover what one group knows.
What this means
Every organisation has knowledge that is easier to demonstrate than to document. An experienced employee knows which customer complaint needs immediate escalation, how an old machine sounds before it fails or which data anomaly is harmless. That is normal tacit knowledge: skill and judgement acquired through experience cannot always be converted into a paragraph.
Tribal knowledge becomes a practical problem when the organisation depends on information that could have been made explicit but was not. Nobody wrote down why the invoice run happens on Tuesday, which spreadsheet is authoritative or why a particular integration must never be restarted before another process finishes. New starters learn by sitting near the right person and asking the right question.
The problem can exist inside one person's head or inside a whole team. A knowledge silo may contain well-documented information that simply is not accessible elsewhere, so "silo" and "tribal knowledge" are not exact synonyms. Both create the same awkward moment, however: the business discovers that something important is known locally rather than organisationally.
Why it matters
Unrecorded knowledge makes normal change expensive. Onboarding takes longer because newcomers must discover the organisation socially. Incidents take longer because responders do not know the historical workaround. Holiday cover becomes risky. A resignation triggers a frantic request to "write down everything you know", usually several years too late.
It also disguises the bus factor. A system may have several developers, but only one understands why an apparently unnecessary validation rule exists. A finance process may have written instructions, but only one administrator knows which exceptions require manual intervention. Headcount gives false reassurance when knowledge is concentrated.
AI adds a new reason to care. Retrieval-based AI can search, rank and present recorded material, but it cannot retrieve a reason that nobody captured. Connecting an assistant to a knowledge base does not convert undocumented experience into documents by magic. An empty shelf with excellent search is still an empty shelf.
The remedy is not "document everything". That creates warehouses of stale prose nobody can maintain. The useful aim is to capture the map: critical decisions, dependencies, exceptions, ownership and routes to deeper expertise, while accepting that some skill remains embodied in practice and conversation.
How it works
Where the term came from
"Tribal knowledge" has long appeared in manufacturing, process improvement and quality-management language, where it describes knowledge understood within a group but not formally captured or broadly available. The American Society for Quality includes the term in its quality glossary. Software organisations later adopted the phrase naturally because undocumented systems, oral onboarding and specialist knowledge create the same problem.
The underlying knowledge-management distinction is older. Michael Polanyi's 1966 work on tacit knowledge emphasised that people know things they cannot always fully articulate. Ikujiro Nonaka and Hirotaka Takeuchi's 1995 The Knowledge-Creating Company developed the relationship between tacit and explicit knowledge in organisational learning.
That history matters because it prevents a common mistake. Not all knowledge in someone's head is documentation waiting to be typed. Some expertise depends on pattern recognition, physical practice, accumulated context or judgement. The challenge is to make essential organisational knowledge transferable without pretending the whole of human expertise can be reduced to a wiki.
Why tribal knowledge forms
The usual cause is speed rather than secrecy. Somebody fixes a problem under pressure, explains the workaround to a colleague and moves on. The workaround becomes normal. Months later everybody close to the process knows it, so writing it down feels redundant. Nobody notices that "everybody" means four people who sit together.
Trust can reinforce the pattern. Teams often prefer asking a familiar expert to searching imperfect documentation. The expert gives a faster, contextual answer, so the document receives less use and less maintenance. Eventually the person's memory is better than the official record, which makes asking them even more rational. The organisation has created a self-reinforcing knowledge bottleneck.
Ownership gaps matter too. Documentation decays when everybody can edit it but nobody is expected to keep a particular area trustworthy. Recording knowledge is then treated as work to do after the real work, which means it usually loses.
Find the knowledge before the exit interview
The best clues are repeated questions and workarounds. Ask new starters what they could not discover without asking a person. Review incidents for phrases such as "only Sam knew", "we normally do this manually" or "the documentation does not mention that case". Watch where holiday cover requires a detailed private briefing.
Shadowing can reveal knowledge that experts no longer notice they possess. Ask an experienced employee to perform a routine but failure-prone task while another person records decisions and questions. The useful material is often not the sequence of clicks. It is the conditional reasoning: "I check this first because if the figure is older than yesterday, the next screen is misleading."
Decision history is another rich source. Teams often know what the system looks like but not why it looks that way. Without the rationale, a future employee may "simplify" something and rediscover an old failure. Short decision records can preserve the reason without requiring a novel about every project.
Exit interviews can still recover knowledge, but relying on them is poor design. A departing employee has limited time, incomplete recall of what others do not know and little opportunity to test whether the material makes sense to a successor.
Capture enough to make knowledge transferable
Different knowledge needs different forms. A runbook suits recurring operational procedures. A decision record preserves why an important choice was made. A short system map explains ownership and dependencies. A recorded walkthrough can capture demonstration and commentary quickly, then be turned into searchable notes after review. An SOP works for repeatable business processes with clear checkpoints.
The strongest test is use. Ask somebody who did not write the material to perform the task or answer the question from it. Their confusion reveals assumptions the expert did not know they were making. Documentation improves when it is part of onboarding, incident response and day-to-day work rather than an archive maintained for appearances.
Some tacit knowledge should be transferred socially. Pairing, shadowing, rotation and communities of practice allow people to observe judgement in context. The goal is resilience, not the elimination of conversation.
AI can expose the gap, but not erase it
Retrieval-augmented generation, or RAG, combines a generative model with retrieved external material. That makes recorded organisational knowledge much more useful to an AI assistant, because relevant documents can be supplied when a question is asked. It does not help with information that was never recorded or made accessible to the retrieval system.
AI can assist the capture process by turning a reviewed walkthrough into a draft runbook, summarising a decision discussion or grouping repeated newcomer questions. Human review remains essential because the model cannot know which undocumented detail is business-critical merely from how confidently it can phrase a summary.
A useful AI knowledge project therefore often begins before the AI. Identify authoritative material, owners, gaps and access boundaries. The retrieval layer is downstream of knowledge management.
Examples
An illustrative retailer has one administrator who knows that refunds from a legacy sales channel must be reconciled against a separate report on Monday mornings. The written finance process never mentions it because everybody in the original team learned the rule verbally. When the administrator takes extended leave, discrepancies accumulate. The repair is a short runbook covering the exception, its reason, owner and evidence to check, plus another employee practising the process before the next absence.
A charity's new operations manager spends six months asking which spreadsheet, policy note or email thread contains the current answer. Rather than launch a giant documentation programme, the charity records the questions asked during onboarding, identifies the 20 most frequently depended-on pieces of information and assigns owners to make those sources explicit and searchable.
A software team connects an AI assistant to its internal documentation. It answers questions about documented APIs well but invents plausible explanations for an old billing rule. The failure is traced to missing decision history: nobody ever recorded why the rule exists. The team adds a reviewed decision record and teaches users that retrieval can surface recorded knowledge, not reconstruct forgotten organisational history reliably.
Common misunderstandings
Misunderstanding: Tribal knowledge means all tacit knowledge is bad. Correction: tacit expertise is unavoidable and often valuable. The risk arises when essential organisational information is concentrated or inaccessible and could reasonably have been shared.
Misunderstanding: A knowledge silo is exactly the same thing. Correction: not quite. Tribal knowledge is unrecorded or informally transmitted knowledge. A silo can contain explicit material that is well documented inside one team but unavailable to others.
Misunderstanding: Tribal knowledge is simply the bus factor. Correction: it is not. Bus factor describes the risk created when critical capability is concentrated in too few people. Tribal knowledge is the undocumented knowledge itself, which is one important cause of a low bus factor.
Misunderstanding: A knowledge base fixes the problem. Correction: a knowledge base organises recorded material. It cannot contain a decision, exception or technique that nobody has captured.
Misunderstanding: AI can interview employees and document everything automatically. Correction: AI can help transcribe and structure material, but experts and users still need to identify what matters, correct mistakes and test whether the record is usable.
Risks and boundaries
Documentation has a carrying cost. Capture too little and the organisation depends on memory. Capture too much and employees face thousands of stale pages whose authority is impossible to judge. The practical target is high-value knowledge: critical procedures, unusual exceptions, system boundaries, decision rationale, ownership and routes to deeper expertise.
The phrase "tribal knowledge" itself can sound dated or exclusionary in some settings. "Tacit knowledge", "undocumented knowledge" or "institutional knowledge" may be clearer when the metaphor distracts from the problem. Whatever term is used, avoid treating the people who hold knowledge as the cause. Organisations usually create the incentives and structures that keep it informal.
Some knowledge will always resist codification. Judgement built through years of experience may be better transferred through practice, mentoring and observation. A document can show where expertise lives and when to seek it without pretending to replace the expert.
What to do next
Start with dependency, not document volume. Ask which tasks, systems and customer commitments would become difficult if a particular person or team were unavailable for a month. Those areas deserve attention first.
Use new starters as sensors. Record the questions they have to ask because the answer cannot be discovered from an authoritative source. Repeated questions are a ready-made documentation backlog.
Capture rationale as well as procedure. A runbook saying what to do is useful; a short note explaining why a surprising step exists can stop a future colleague from deleting the very safeguard the organisation once learned to need.
Give important knowledge an owner and a use case. Material used for onboarding, holiday cover, incident response or AI retrieval is more likely to stay current than material written simply because somebody ordered a documentation drive.
For AI knowledge work, test the corpus before the assistant. Ask representative questions and see whether authoritative written material actually contains the answers. Retrieval quality cannot compensate for knowledge that never made it into the record.
FAQs
What is tribal knowledge in a business?
It is important information or know-how that employees rely on but that remains mainly in people's memories, habits and informal conversations.
Is tribal knowledge the same as tacit knowledge?
They overlap, but not completely. Tacit knowledge can include expertise that is inherently difficult to express, while tribal knowledge often includes information that simply was never documented.
What is a knowledge silo?
It is knowledge concentrated within one person or group and not readily accessible elsewhere. The knowledge may be unwritten or formally documented inside the silo.
Why is tribal knowledge risky?
It increases dependency on particular people, slows onboarding, makes incidents harder to resolve and leaves gaps when staff are absent or leave.
How do you find tribal knowledge?
Look at newcomer questions, incident reviews, holiday-cover briefings, repeated workarounds, shadowing sessions and tasks that always require the same expert.
What should be documented first?
Prioritise critical procedures, exceptions, ownership, system dependencies and the reasoning behind surprising or irreversible decisions.
Can RAG retrieve tribal knowledge?
Only after the relevant knowledge has been captured in material the retrieval system can access. RAG cannot retrieve information that exists only in someone's head.