What is a company brain?
The idea is older than the name: a knowledge base, an intranet or a wiki. The difference is how you use it. You ask a question in plain language and get an answer from your own documents, with a link to the passage where it is written.
A general chatbot answers from what the language model learned from the internet and knows nothing about your quotes, procedures or client agreements. A company brain searches your sources first and only then writes an answer.
How it works, without jargon
Behind every question are four steps. You can check what happens at each step, and that is the difference with a chatbot that only gives an answer.
- 1. Splitting
- Documents are divided into passages of a few paragraphs, each with its title, date and the place it came from.
- 2. Searching
- A question searches two ways at once: on the words you use and on what you mean. That way you also find the passage that says “notice period” when you asked “how far ahead do I have to cancel”.
- 3. Ranking
- The best passages come first, and people only see passages from sources they already have access to.
- 4. Answering
- The language model writes an answer from those passages only and adds the source. If nothing fits, it is built to say so instead of guessing. The test set measures how often that works.
What I learned from my own company brain
My own company knowledge sits in a company brain: meeting notes, decisions, project notes and client files, in September 2026 over 2,500 documents and over 15,000 passages. I query it daily, and my own AI agents do too before they claim anything about my business. A question takes about a second.
Three lessons from it I apply with every client. One large log file with hundreds of passages pushed other sources out of the results, until I set a maximum per document. Notes without a date give old decisions as much weight as new ones, so every note gets the date it was last checked. And the index always runs slightly behind the source, so whoever just edited a document has to know the answer only changes a little later.
What a company brain is not
Not a system that knows everything. Connect every folder, mailbox and chat channel at once and you get answers from old versions and documents that contradict each other. A useful company brain starts with the sources that are right, for the questions that come up most.
Not a replacement for judgement either. It shows what is written and where; whether advice from last year’s document still applies is for a person to decide.
Five places where it goes wrong
The technology is rarely the problem. These five are.
- Old versions
- Three versions of the same procedure sit side by side, and the system picks the wrong one. Fix: remove outdated sources and put a date on every document.
- Contradiction
- Two documents say different things. A good system shows both with their source, instead of picking one.
- Permissions
- Leave out the permissions from your existing systems and a salary overview shows up for anyone who asks. Permissions belong in the search step, not in the answer.
- One dominant source
- A large document with hundreds of passages crowds out smaller sources that hold the right answer. A maximum per document fixes that.
- Nobody notices decline
- Without a test set you do not see answers getting worse after a new model or a new source. You only hear it from a user who already relied on it.
How to measure whether a company brain is right
With a test set of real questions, each with the right answer and the source where it is written. I draw up that list with the people who ask the questions, before I build anything.
Every question gets one of four outcomes: right with the correct source, right but without a source, a correct “that is not in there”, or confidently wrong. The last counts as a serious error, because someone relies on it. We agree the threshold up front, and the test set runs again with every new source and every new model version.
Where your data stays
Before a single document goes in, I put five things in writing.
- Processing agreement
- Signed before any data is processed. You are the controller of the data, I process it on your behalf.
- Location in the EU
- The region where the data sits is in the agreement. Not assumed, written down.
- Retention
- Per type of source: how long it is kept and what deletion means, technically too.
- Own environment
- Every client gets a separate environment of its own, no database shared with other clients.
- Access
- Who may see what follows the permissions you already have. What someone may not open in the source, they do not see in an answer either.
When a company brain is not the answer
If the knowledge is not written down anywhere, there is nothing to search. It has to be written down first, and a company brain can help with that but does not replace it. If the documents themselves are unreliable, you get faster access to wrong answers. And if three people know everything and sit next to each other, asking across the desk is quicker.
In my assessment, a company brain for internal knowledge does not fall into the high-risk categories of the EU AI Act. If it starts to help decide about people or critical infrastructure, I have that checked legally first.
Frequently asked
What is a company brain?
A company brain is a place where a company’s knowledge from documents, notes and decisions comes together, so people can query it with AI. Every answer points to its source. When the answer is not there, the aim is that the system says so instead of guessing; a test set measures how often it does.
What is the difference between a company brain and ChatGPT?
ChatGPT answers from what the model learned and does not know your internal documents. A company brain searches a chosen set of your sources first and states where each answer comes from. That makes the answer checkable.
How do you build a company brain?
Start with the questions, not the documents. Collect what people ask each other most, choose the sources that answer those questions and test the system on them. Only add more sources once the answers are right. That way you keep sight of quality with every expansion.
Which companies benefit from a company brain?
When the same questions keep landing with the same colleagues, when knowledge sits in mailboxes and folders instead of one place, or when new hires need months to find their way. Which clients I build such a system for is on my clients page.
What does a company brain cost?
I build a first version in a Pilot-to-Production sprint of four weeks at most, at a fixed price I name after seeing the sources and questions. If you first want to know whether it fits, it starts with a Scan from €1,995 ex VAT.
Who maintains a company brain after delivery?
Someone in your own organisation, by name. That person tracks which sources are in it, removes outdated documents and watches the test set. If the score drops, that person sees it first. A company brain without an owner is, six months on, an archive with a search bar.
Where to start
For one week, write down the questions colleagues ask each other. If the same questions keep coming back and the answers sit somewhere in documents, a company brain is worth it. If they sit nowhere, the knowledge has to be written down first.
Knowledge stuck in heads and folders?
Book a first conversation. I look at your sources and the questions that keep coming back, and tell you whether a company brain fits or something else has to happen first.
30 minutes · free, no obligation