Your knowledge sits in people's heads. What if one of them leaves?

The colleague who spent eight years knowing which supplier gets what wrong just handed in their notice. A company knowledge base is built for exactly that day: so the answer doesn't leave with them. It isn't a folder or a notebook. It writes down how you actually work, and not only for human readers. The difference isn't storage. It is that the pieces point at each other.

What is a company knowledge base?

A company knowledge base is the written version of how your firm operates: decisions, processes, customer knowledge and templates in one place, linked to each other. It isn't an archive, it is a working tool. What makes it a knowledge base rather than a pile of files is that every claim carries its source and its neighbours.

Start with what it isn't. It isn't the shared drive. It isn't the wiki someone installed once and nobody has opened since. It isn't the document where you wrote the processes down once, three years ago. Those are containers. A knowledge base is a layer above the containers: it describes how you work, and behind every statement it says where that came from.

In practice, two things separate it from a pile of files. First, the raw source (the quote, the email, the meeting notes, the supplier price list) lives apart from the tidy article written out of it. The raw material is never edited, because that is the evidence. The article keeps getting better. Second, the articles point at each other: from a customer page you can reach the product you shipped them, from there the decision behind it, and from there the reason you priced it the way you did. That difference decides whether the thing helps you later, or becomes one more place to search.

We run our own the same way: raw sources in one layer, linked articles in another, and a translation step between the two that turns one into the other. That is our daily practice, not a diagram in a deck. A client gets the same structure when we build a second brain for their company.

Why aren't folders and email enough?

Because folders organise files, not knowledge. Email holds the answer, but only for someone who remembers the person and the month. Neither one tells you which version is current, and neither connects a decision to the reason behind it, so every question ends up back at a human being.

It breaks in three places. First, the question of versions: if the price list exists in three places, then you don't have a price list, you have three, and whoever goes looking is just as likely to open the oldest one. Nothing on it says that it is old.

Second, the missing why. The contract is there in the folder, but nowhere does it say why you agreed to those payment terms, or whether that turned out well. The problem isn't that the paperwork is missing. The problem is that paperwork keeps the outcome and throws away the thinking. A year later the same question comes back, and memory decides it again.

Third, search. An inbox searches for words, while the good answer often sits in a sentence that never uses the word you typed. If you don't remember the subject line and the sender, your email isn't a source, it is an excavation site.

And there is a fourth one, which always shows up late: folders and inboxes stick to people. When someone leaves, the files stay, but the order, the context and the arguments you already settled do not. Our first article is about what a written decision is worth: a model could write a usable text in a single pass because every decision had its reasoning sitting next to it.

What is worth putting in, and what isn't?

Put in what repeats and what has reasoning behind it: decisions, processes, customer knowledge, pricing principles, templates. Leave out anything that updates itself somewhere else, or that nobody will ever look for. An invoice belongs in your accounting, a meeting in your calendar. A company knowledge base points at those systems instead of copying them.

There is a simple filter for it: have you explained this to someone twice already? If yes, write it down. If no, it can wait.

• Decisions with their reasoning. What you chose is not enough. You also need what you ruled out, and why. This is the one thing that cannot be reconstructed later from anywhere else.
• Repeating processes. Quoting, onboarding, handover, handling a complaint. Anything that runs the same way for every client is worth thinking through exactly once.
• Customer and supplier knowledge. Who asks for what, where things usually stall, what each one is sensitive about. This kind of knowledge lives almost entirely in people's heads, and it is the first thing to disappear.
• Templates and good examples. A strong quote, a well written reply, a precise technical description. New colleagues learn from these, and so does the machine.

What stays out: confidential material, until you have decided who may see it; personal data you have no legal basis to keep; and any figure that lives and updates somewhere else. Stock levels belong in the stock system, invoices in accounting. Copy them across and you own two truths, and the worse one tends to be the one at hand.

Volume is not a virtue. A small, maintained knowledge base gives you the answer faster than a large one where nobody can tell any more which part still holds. So let real questions drive the growth: whatever somebody asks you a second time earns its own article.

How long does it take to come together?

It isn't a calendar question. The first useful version doesn't cover the whole company, only one recurring question: the thing you explain to somebody several times a week. After that it stops being a project and becomes a habit, a paragraph after every closed job. Regularity sets the pace, not headcount.

The first step is not picking a tool. An average tool you actually write into beats the perfect one nobody ever touches.

The second step is picking the topic. Don't start with the company history, start with the question that interrupts your day most often: how we quote, what happens after a new order lands, what we say to someone who complains. One topic written through moves more than ten half finished ones.

What really takes time is the very first round. That is when it turns out that the thing everybody knows is known three different ways by three people. This is not a fault in the system, it is its first benefit: you have to talk the differences through before anything can be written down at all.

After that, most of the work does not start from a blank page. The raw material is already there: quotes, emails, contracts, minutes, notes. Out of that you write tidy, linked articles, and today that translation step is largely machine work with human approval. All the machine needs is something to read. What goes into an article and what stays out is decided by a person.

From there, rhythm keeps it alive. A paragraph after every closed job, every quote and every argument, written while it is fresh and you still remember why it went that way. Skip it and the knowledge base doesn't collapse, it quietly goes stale, which is the worse outcome: anyone reading outdated material is confidently wrong.

Why has this become infrastructure now?

Because a new kind of reader showed up, one that reads everything: the language model. As long as only people used written knowledge, a question asked in the corridor filled the gaps. An AI can't walk into the corridor. It knows exactly as much about your company as you wrote down.

Writing things down was a good idea twenty years ago too. It just was never urgent, because a missing document could be patched with a question: you called across to the next room and got your answer. So documentation always slid to the bottom of the list, and that was a rational choice.

What changed is the arithmetic. An AI assistant or an agent is worth exactly as much inside your company as the context it gets. The same model writes the generic paragraph and the precise one that fits you; the difference does not come from the model, it comes from what you hand it. That is why a second brain is not a note taking fashion but infrastructure: it is the layer every other piece of AI work sits on.

The order of operations follows from that. A chatbot or an automation may be the first idea, but with no written knowledge behind it the machine guesses, and it guesses wrong exactly where your trade begins. We have a separate article on choosing your first agent, and the precondition there is the same: give it something to read first.

For us this isn't theory. Everything anyone can ask about the company sits on that same layer: our quotes, our pricing principles and the reasoning behind our decisions. Not because it looks good, but because a person and a machine can both read it end to end.

One decision, one place.