AI Tribal Knowledge: What It Is and Where It Hides

Take a twelve-person insurance agency. The office manager, call her Dana, started using ChatGPT about a year ago for the renewal reminder emails. The first ones were terrible: too long, too chirpy, and once it invented a discount the agency doesn’t offer. So Dana fixed it, one correction at a time. “Shorter.” “Never mention pricing.” “Three paragraphs, and the second one is always the deadline.” After a couple of months the reminders were good. After six, she had the same routine running for claim follow-ups and the monthly summary for the owner.

The owner sees the results. Renewals go out on time, the summaries read well, Dana is faster. What the owner has never seen is how any of it works.

If Dana gave notice on Friday, the owner would discover on Monday that the renewal reminders don’t come out of ChatGPT. They come out of Dana’s ChatGPT, tuned over a year, and the tuning is not in any place the company can reach.

AI tribal knowledge is the working know-how behind good AI results (the instructions, exceptions, examples, and judgment calls) that lives in one person’s head and one person’s account rather than in documents the business owns.

Every small business already has tribal knowledge: the bookkeeper who knows which client wants the PDF and not the link, the technician who knows the back door sticks in August. AI has added a new layer of it, and this layer has a property the old kind never had. It can be moved into a document in an afternoon, and it can also vanish with one password reset.

Where Dana’s Knowledge Actually Lives

Ask Dana where the renewal reminder instructions are and she’ll say “in ChatGPT.” That answer covers three different places, and none of them belong to the agency.

1In the conversations

Most of the tuning happened as corrections typed into individual chats. “Cut the second paragraph.” “Don’t say ‘reach out.'” “Use last year’s format.” Each correction improved that one conversation and nothing else. A model starts every new chat blank, so the next day Dana opened a fresh window and, without thinking about it, retyped the corrections that mattered. Her chat history holds a year of these. Nobody else can open it.

2In the settings nobody opens

ChatGPT, Claude, and most business AI tools have a settings page, usually labeled Personalization, Memory, or Custom Instructions, where the tool keeps standing notes about the user. Some Dana wrote herself: “I work at an insurance agency. Keep emails under 120 words. Never mention pricing.” Some the tool wrote on its own after noticing her habits: “User prefers renewal reminders in three short paragraphs with the deadline in the second.” This page is why Dana’s ChatGPT behaves differently from a brand-new account. It is also tied to her login, editable only by her, and redesigned whenever the vendor decides to redesign it.

3In Dana’s head

Which clients get a phone call instead of an email. Which claim types the model gets wrong, so she always checks those by hand. What “good” looks like for the monthly summary, which she knows because the owner once said “this one was great” and she kept that one. None of this was ever written down, because nobody asked.

Here is what the same knowledge looks like in each place, using one rule from Dana’s routine:

Only the third column is company knowledge. A plain document in a shared drive, with a name and a date on it. Boring on purpose.

Why This Is a Business Continuity Problem, Not a Tech Problem

Key person risk is familiar to any owner who has watched a bookkeeper give two weeks’ notice. AI concentrates that risk, because the person who builds the AI routine is usually the one who runs it, and the routine leaves fewer traces than a spreadsheet would. A spreadsheet at least sits in a folder. Dana’s routine sits in Dana.

Picture the handover. Dana’s replacement inherits a login they can’t use and chats they can’t read. The instructions that made the reminders good were refined over forty or fifty corrections, and none were saved. The replacement types “write a renewal reminder,” gets the long, chirpy version Dana got on day one, and concludes the AI “doesn’t really work for us.” The agency is back to square one, and it doesn’t know it, because nobody at the agency ever saw what square one looked like.

There is a second cost that shows up while Dana is still there. Only Dana gets the benefit. The producer who could use the same email standard and the part-time assistant who could run the same claim follow-ups can’t, because the know-how has no address inside the company. One person’s AI fluency keeps rising while the team’s stays flat.

The agency is paying for the results and owns none of the machinery.

The Five Things the Agency Should Own

Moving AI tribal knowledge into company ownership takes five plain documents, most of which fit in one shared folder. Here is what each one looks like for Dana’s agency.

1Shared company instructions

One page. Dana’s settings, cleaned up and generalized so anyone can use them. In practice it reads something like: “We are a family insurance agency. Emails under 120 words, plain language, no exclamation points. Never mention pricing or discounts. Never enter client names, policy numbers, or claim details into any AI tool. Renewal reminders are three paragraphs; the deadline goes in the second.” Anyone at the agency pastes this into their own tool’s instructions page. If you already have an AI manual, this is its first section.

2Approved reference material

The three or four files the model should be reading when it works: the current product list, the renewal timeline, the FAQ page, the one policy about what the agency will and won’t say in writing. Small and current beats big and stale. A twelve-page brand book nobody has updated since 2023 does worse than a one-page sheet that is right.

3A reusable examples library

Dana’s ten best real outputs, saved as files with plain names: “renewal reminder, standard,” “claim follow-up, delayed adjuster,” “monthly summary, the one the owner liked.” Examples do more work than instructions. Anthropic’s engineers, writing for developers, make the same point: a handful of well-chosen examples beats a long list of rules. Dana already has these. They’re in her chat history. Pull them out.

4A named owner

One line at the top of the instructions page: “Owner: Dana. Next review: first Monday in December.” Ownership is what turns a folder into a maintained document. Without it, the shared version drifts within a quarter and everyone goes back to their private setups.

5A handover step in offboarding

One line added to the offboarding checklist, next to “return laptop” and “transfer client list”: “Export AI instructions, memory entries, and best examples to the shared folder; walk successor through them.” Ninety minutes. Most companies have a knowledge transfer plan for passwords and nothing at all for the AI routine that produces half the reports.

What the AI Labs Already Know About This

In September 2025, Anthropic’s applied AI team published a guide to context engineering, written for developers building AI agents. Strip the vocabulary and it is advice about the same problem Dana’s agency has: how to give an AI what it needs to do the job well, every time, without depending on one person’s memory. Three of its points translate directly.

• Keep instructions clear and short. The guide warns against instructions so detailed they break and instructions so vague the model guesses. Dana’s one-page standard should be specific about tone, length, and what’s off-limits, and silent about things a competent adult would work out.
• Show, don’t list. The guide’s phrase is that engineers should “find the smallest set of high-signal tokens.” For the agency, that’s the examples library. Ten real reminders beat a rulebook.
• Keep notes outside the conversation. Anthropic’s agents write what they’ve learned to a file that survives when the chat resets, then read it back. Dana should do the same on purpose: when a correction finally works, it goes into the shared page, not just into today’s chat.

None of this requires an engineer. It requires deciding that AI know-how is company property, then about a day of moving it out of one person’s account.

Whatever lives only in one person’s AI account is not a company asset yet. It becomes one the day it’s written down, named, and owned.

Dana did real work building that routine. The way to honor it is to make sure it outlasts her: a folder in the shared drive with five documents in it, her name at the top of each, and a date for the next review. And on the Monday after she leaves, the renewal reminders go out anyway.

Frequently Asked Questions

What is AI tribal knowledge?

AI tribal knowledge is the working know-how behind good AI results (the instructions, exceptions, examples, and judgment calls) that lives in one person’s head and one person’s account rather than in documents the business owns.

Where do I find what my employee’s AI tool has saved about them?

Look for a settings page labeled Personalization, Memory, or Custom Instructions in ChatGPT, Claude, or whichever tool they use. It holds standing notes the person wrote plus notes the tool added on its own. Only the account holder can open it, so ask them to read it to you.

Is ChatGPT memory the same as a company knowledge base?

No. Saved memory is a personal setting tied to one login. It makes that person’s tool behave consistently, but the company can’t see it, review it, or keep it when the person leaves. A company knowledge base is a set of maintained documents the business controls and can hand to anyone.

Does a team or enterprise AI plan solve the problem?

It helps with sharing, since projects and instructions can be visible to colleagues instead of locked in private accounts. It doesn’t settle ownership: the shared workspace still belongs to the vendor’s product and changes when the product does. Keep the instructions, reference material, and examples as plain documents in your own drive, and load them into whichever tool the team uses.

What happens to AI knowledge when an employee leaves?

Unless it has been moved into company documents, it leaves with them. Chat histories stay in their account, saved memory is tied to their login, and the refined instructions were never written down. The fix is a handover step in the offboarding checklist: export instructions, memory entries, and best examples into the shared folder and walk the successor through them.

What should a small business do first?

Ask the team’s best AI user to read their custom instructions and saved memory out loud. Everything that surprises the owner is tribal knowledge. Copy it into one shared document, assign an owner, and put a review date on the calendar. That single afternoon moves most of the risk.