Every team runs on rules nobody wrote down. You probably recognise at least one of these.
- The private listthat someone keeps next to the official system
- Last year's folderwith the example texts everyone copies from
- The colleague you callbecause that's quicker than the procedure
None of it is in a procedure, it lives in people's heads, and it walks out the door when they leave.
And it isn't one department. Finance has them, marketing has them, HR has them, planning has them. Most organisations have dozens of processes like this, and I've spent the past few months working out how to get those rules into AI without losing them along the way.
The problemOne bot per process doesn't scale
The obvious route is one AI tool per process. A bot for invoices, one for social posts, one for the weekly report. I've built plenty of those myself, and each one works on its own.
The trouble starts somewhere around number five. Every bot has its own copy of the tone of voice, the audiences, the sources and the rules. Change the house style and you update five places, or you forget one. After a year you have a drawer full of tools and five versions of the truth.
It's the colleague test, really: if you can't hand it over to a colleague, you can't hand it over to AI. And you don't give a new colleague a separate brain for every task.
The ideaOne brain, filled one process at a time
So I build it the other way round. One company brain, and every process you hand over adds to it.
The brain has a shared layer that every process uses.
- Who we areBrand, tone of voice, audiences.
- Rules that hold everywhereWhat's never allowed, what stays confidential, who decides what.
- SourcesThe information that's fixed, and the systems it fetches fresh every time.
- PeopleWho owns which process, and who knows what.
On top of that, each process gets its own part, always in the same form: the steps, the rules for that process, and what good looks like. Plus a test set of real cases with the right answer.
The second process starts with everything the first one wrote down. The brand is already there, the sources are there, the rules that hold everywhere are there. That's where it scales: every process you add makes the next one faster.
Swipe sideways to see the whole diagram.
How a process gets in
Getting what people know onto paper is most of the work. Building turns out to be the small part, and only after the knowledge is down do limits and controls make the system safe to use.
For each process, that knowledge comes from three places.
- The documents. The current file, three good and three bad cases with their corrections, and the latest version of every source.
- The old data. If there's history, AI goes through it before the conversation and lists what keeps coming back and what deviates.
- The conversation. With the people who do the work, recorded, with the real file open and one tricky case on the table.
The conversation is the one you can't skip, and I learned that the hard way. On one project I asked the team to talk me through their process and got a written brief instead. The rules that really mattered, which sources count and what isn't a reason to drop something, came back as feedback on the output. Three rounds, over two months.
How much work a process takes depends mostly on what's already written down. If the documents are good and up to date, the job is largely checking them against how the work really goes. If they're thin, old or scattered across mailboxes, writing them is the bulk of the work. And even a good document describes the official route. The workarounds are exactly where the work drifts from it, so documents get the same treatment as the data: they propose, the people who do the work confirm.
From request to a system in use, it runs in eight phases: first out of heads and loose documents, then onto paper as a process, and only then the system.
Swipe sideways to see the whole diagram.
Out of heads. A request gets a case file before anyone spends time on it. The exploration asks two things: may we, and does it pay? That includes a baseline in hours and times per month, because without one you can't say later what it delivered. Not going ahead is an outcome too, with the reason written down. Then the process conversation.
Onto paper. A one-page proposal for a small first version, with an account in the organisation's name and an owner with a name. Then the process goes into the brain: its steps, its rules and what good looks like, plus a fixed test set of real cases with the right answer. Whatever turns out to be true beyond this one process goes into the shared layer. That's worth having even if you never build anything, because the work can now be handed over.
Into a system. AI builds the first version on top of the brain, in the tool the team already has. Every test round runs the same test set and makes one change, and that change goes back into the brain as well as the tool. Handover means the owner runs it once alone, without help. After two and six weeks you look back against the baseline.
The guardrailThe data proposes, the expert signs
Some tools learn these rules from data alone. Talky Labs, a Madrid startup that raised €1.5 million in 2026, reads a company's accounting history and the way its accountants resolved past queries, and applies that judgement to new documents. A person steps in when something is off or missing. Talky says it handles 86 to 93% of documents that way, by its own count (EU-Startups). It's a smart approach, and I'd still do it differently, at least from what's been published about how Talky works.
Patterns in old data are useful. AI is good at spotting them, and they often give you the baseline for free: how often something happens, and how long it takes each time. But a pattern is a question for the conversation, not a rule.
So in the process conversation we go through the list, pattern by pattern. Is this a rule, a coincidence, a workaround, or does nobody know? The person who knows the process confirms it, with their name and a date. What nobody can explain stays a question.
Old data also holds the mistakes. A habit that's been wrong for three years still shows up as a very convincing pattern.
This matters more as the brain grows. With ten rules you can keep track in your head. With five hundred, you need to know which ones someone actually signed off. So everything in the brain carries a status: signed off, with a name and a date, or still a working note. A sign-off expires after six months. And when a rule is replaced, the old one is marked as replaced, so nobody builds on it by accident.
Connections to other systems, like an ad account, analytics or a CRM, follow the same line. They start read-only, on the organisation's own account. Writing starts with proposals: the system suggests, a person carries it out.
The rulesFive rules that run through every phase
- Conversation over briefA written description is preparation. What you don't ask about now comes back later as feedback on the output.
- Test set before buildingReal cases with the right answer, fixed before anyone builds anything.
- Together at one screenI see whether it works, the process expert sees whether it's right. In the same session.
- Account and owner firstAn account in the organisation's name and an owner with a name are already in the proposal.
- Honest statusWhat the system can't find, it reports as not found, with where it looked. Every change is logged.
What stays with people
Not everything fits on paper. What you can only show stays with people, and you catch it in examples sooner than in rules. Nonaka and Takeuchi described this back in 1995: knowledge moves from tacit, what people know but never wrote down, to explicit, what's on paper, and back again. This method follows that loop, including the way back.
That way back is the part I find most interesting. When the system does the work, the knowledge of the person who reviews it slowly wears away. So the owner still does the work themselves now and then, or reviews a fixed sample. Corrections from users go back into the brain once the process expert has approved them.
If a tool gets to 90%, that's the question I'd ask: who still knows how to do the other 10%, two years from now?
StatusWhere I am with this
I build AI systems, pretty much since ChatGPT came out. My own second brain runs on exactly this: one shared layer, a separate part for each area of my work, and a status on every file.
The method to set the same thing up in an organisation is written down, every phase has a template, and the first trial is under way now. Nothing has run end to end at a client yet, so for now this is a method, not a case.
Which rule in your team only exists in someone's head?
FAQFrequently asked questions
What is a company brain?
A company brain is one shared place that holds what an organisation knows, so every AI system can work from it. It has a shared layer with the brand, the rules that hold everywhere, the sources and the people, and a part for each process with its steps, its rules and what good looks like. The AI tools are built on top of it and can be replaced without losing that knowledge.
Why not build a separate AI tool for every process?
Every separate tool needs its own copy of the tone of voice, the audiences, the sources and the rules. Change the house style and you update every tool, or you forget one. With one brain, the second process starts with everything the first one wrote down.
Can AI learn the unwritten rules from old data on its own?
AI can spot the patterns, but a pattern is a question for the conversation, not a rule. Old data also holds the mistakes: a habit that has been wrong for years still looks like a convincing pattern. So the person who knows the process confirms each rule, with their name and a date.
What is the difference between tacit and explicit knowledge?
Tacit knowledge is what people know but never wrote down: the judgement calls, the workarounds, the things you can only show. Explicit knowledge is what's on paper: procedures, rules, documents. The terms come from the SECI model by Nonaka and Takeuchi (1995), which describes how knowledge moves between the two. Building a company brain is mostly turning tacit knowledge into explicit knowledge. What you can only show stays tacit, and you catch it in examples.
How do you get tacit knowledge out of people's heads?
From three sources per process: the documents, the old data, and a conversation with the people who do the work, with the real file open and one tricky case on the table. The conversation is the one you can't skip. A written brief leaves out the rules that matter most, and they come back later as feedback on the output.
How much work is it to build a company brain?
It depends mostly on the documents you already have. Good, up-to-date documents mostly need checking against how the work really goes. Thin or outdated ones need writing, and that's where most of the hours go. Each process after the first one is lighter, because the shared layer is already there.
Does the team lose its own knowledge once AI does the work?
It can. When the system does the work, the knowledge of the person who reviews it slowly wears away. That's why the owner still does the work themselves now and then, or reviews a fixed sample, and corrections go back into the brain once the process expert approves them.