How do you capture tribal knowledge before someone leaves?

You capture tribal knowledge by asking everyone, privately, while they are still there, and by keeping what they say in one place that stays current. Not by asking the person who is leaving to write a handover document in their last two weeks. Tribal knowledge is the way work actually gets done, as opposed to the way the manual says it does: which supplier's invoices always need a manual step, who really approves a large quote, why the month-end pack only runs on one person's laptop. It lives with the people doing the work and rarely reaches a system, so the only way to capture it is to hear the people, and the only way to keep it is to keep hearing them.

What tribal knowledge is, and why it matters more now

Wikipedia defines tribal knowledge as knowledge that is known within an in-group of people but unknown outside of it, and notes the term's roots in Six Sigma and manufacturing, where it names the unwritten information a plant needs to produce quality work (Wikipedia). In an owner-led company of 50 to 500 people, that is most of the company. The org chart shows reporting lines; the tribal knowledge shows the phone call that unblocks a shipment, the spreadsheet that is the real inventory system, and the person everyone asks.

It has always mattered when someone retires or resigns. It matters more now because AI agents are being deployed into companies, and an agent can only follow the process it has been given. If the given process is the manual and the real process is tribal, the agent is built on the wrong company. That is the shallow AI problem: automation bolted onto work nobody described. Capturing tribal knowledge is no longer an HR nicety; it is the precondition for AI that pays.

Why the usual methods fail

The handover document. Written by the leaver in their last fortnight, from memory, for a reader who does not yet exist. It records what they think matters, not what breaks when they are gone, and nobody opens it until something does.

The documentation sprint. A team is asked to write down its processes. Experts cannot write down intuition, the documents describe the process as it was on the day of writing, and they go stale. One of the largest Hacker News discussions on the topic, 547 points and 213 comments, keeps returning to the same complaint: pages and pages of outdated documentation describing things as they were years ago (Hacker News, 2020).

The workshop. A room, a whiteboard and the manager's version of the process. The workaround the buyer uses for urgent invoices is not on the board, because the buyer would have had to describe it in front of finance.

The survey. Fixed questions, one pass, no follow-up. It measures how people feel about the process, not how the process runs.

Each of these asks a few people, once, in public. Tribal knowledge is held by everyone, changes over time, and comes out only in private.

What works: ask everyone, privately, and keep asking

The method that works has four properties.

Everyone, not a sample. The knowledge is distributed. The dispatcher knows why the pick list is late; the fitter knows which machine faults recur; the person in finance knows which supplier never matches. Ask ten managers and you get ten official versions.

Privately. People describe the workaround when their manager cannot read the answer. That is not a courtesy; it is the mechanism. If the tool reports upward, people recite the manual.

With follow-up. Tribal knowledge comes out in the second and third question: "and what happens if she is off that week?" A conversation does this; a form does not.

Continuously. A snapshot of how the company works is out of date by the next quarter. A diary is not. If people keep talking about their work a few minutes a day, the record moves with the work.

Panopto's survey of 1,001 employees found people spend about five hours a week waiting to reach the one colleague who has the knowledge they need, and around six hours a week duplicating work someone had already done (Panopto, 2019). Those hours are tribal knowledge being fetched by hand. Capturing it once, and keeping it current, is the saving before any AI is built.

Finding the single points of failure

The most valuable part of tribal knowledge is knowing who holds it. Key person risk is the vulnerability that arises when critical knowledge, relationships or capabilities are concentrated in one individual; the diagnostic question is simple: if this person left tomorrow, what would break, and how fast (Prialto)? In our experience owners can name two or three such people. The company usually has ten, and several of them are not managers: the person who runs the month-end macro, the dispatcher who knows every driver, the engineer whose phone number is on the shop-floor wall. You find them by asking everyone who they call, because the answers converge.

Where it becomes a world model

Captured tribal knowledge is the first layer of a company world model: a picture of how work actually gets done, built from the people who do it. A snapshot of that picture is an ontology, what exists and what is attached to what. Kept in motion, with what usually happens, what stalls and what breaks when someone is away, it becomes a world model, and that is the thing your decisions and your AI agents can work from. We built Transcript to do this at company scale: it has a private conversation with every person in the company, in their own language, a few minutes a day, asks again where accounts disagree, and lets each person approve what was learned from their words. No manager can read a conversation. Everyone sees the world model. Read more in What is a company world model?

Our view

Most companies try to capture tribal knowledge at the moment it is about to leave, which is the one moment it cannot be done well. The leaver is busy, the reader is absent, and the knowledge that matters is the part the leaver never thought was knowledge. Capture it while people are there, from everyone, privately, and keep it moving. Then the resignation letter is a staffing problem, not a knowledge problem, and the agent you build next year runs on the company you actually have.

Questions people ask next

What is the difference between tribal knowledge and institutional knowledge?

Institutional knowledge is everything an organisation knows, documented or not. Tribal knowledge is the undocumented part, held by people and passed on informally. Every company has both; the risk sits in the second.

Can AI capture tribal knowledge?

Yes, if it talks with people rather than reading documents. Documents hold the official process. The workaround, the phone call and the person everyone asks come out only in conversation, and only when the conversation is private.

How long does it take to capture the tribal knowledge of a 300-person company?

Weeks, not months, if everyone talks a few minutes a day. The first picture is useful within weeks and keeps improving, because the conversations continue.

What do you do with it once it is captured?

Three things: find the single points of failure and give them a backup; fix the processes that most people said hurt; and use the picture as the ground for any AI agent, so it runs on the real process rather than the manual.

Is asking employees about their work a form of monitoring?

Not if no manager can read what an individual said. In Transcript that is enforced in the database, not in a policy. People describe the workaround precisely because the answer stays theirs.

References

1. Wikipedia, Tribal knowledge, https://en.wikipedia.org/wiki/Tribal_knowledge
2. Hacker News, Ask HN: Good ways to capture institutional knowledge? (2020), https://news.ycombinator.com/item?id=22454333
3. Panopto, How much time is lost to knowledge sharing inefficiencies at work? (2019), https://www.panopto.com/blog/how-much-time-is-lost-to-knowledge-sharing-inefficiencies/
4. Prialto, The Key Person Risk: What Happens When Everything Runs Through You, https://www.prialto.com/blog/key-person-risk