What we learned running a company brain on ourselves

Before we asked anyone else to trust a company brain, we moved our own company into one. Every AI tool we use now reads the same rules before it writes in our name: how we phrase emails, what we offer and what we don’t, how we check facts before we send them. Coding agents, chat assistants and a voice agent all pull from one place.

We started with the area we use every day: how we communicate. Most of those rules already existed as scattered notes and corrections we’d given our AI tools over months. Getting them into one place taught us more than the planning did. This is what we’d tell anyone starting out.

1. Topics are areas of responsibility, not projects

Our first import sorted rules by what they were about: one customer project, one tool, one website. It looked tidy and was useless. A rule about how we quote fixed prices applies to every project, and nobody owns “the website project” in the sense of deciding what’s true.

Topics only work when they map to someone who decides: communication, sales and offers, quality checks. If you can’t name the person who would approve a change, it isn’t a Topic.

2. Rules and facts are different things

A lot of what we had written down was status: which version is deployed, which account ID belongs to which service, what a customer said last Tuesday. That’s useful, but it changes weekly and nobody approves it.

A rule says what applies and why. “Customer emails name a fixed price, never hours” is a rule. “The current price is X” is a fact that lives in a price list or a CRM. When we mixed the two, the brain went stale within days and the real rules got buried. Now facts stay in the systems they belong to, and the brain only holds what someone decided.

3. Keep the contradictions

Our first extraction step was too helpful. When two statements disagreed, it either threw one away as “opinion” or merged both into a single, smoother rule. Both moves hide exactly the thing an Owner needs to see.

A contradiction is information. Someone believes the return window is 30 days and someone else believes 14, and both have been telling the AI their version. We changed the extraction so contradictions survive as Conflicts and land in front of the people who have to decide.

4. Numbers are decisions too

“We don’t go below this price for a website” got filed as a fact and dropped, because it looked like data. It isn’t. A price floor, a discount limit or an approval threshold is a decision someone made and someone owns. If your extraction treats every number as a fact, it loses some of the most important rules you have.

5. Better instructions beat a bigger model

At one point a smaller model kept missing a rule about vouchers. It guessed which Topic to open, opened the wrong one, and answered confidently. The obvious fix was a more expensive model.

What fixed it was one sentence in the instructions every AI receives: search first, once for each aspect of the task. A question about a discount email touches pricing, tone and what you may promise, so that’s three searches, not one guess. ChatGPT showed the same pattern over the same connection. Try the instructions before you pay for a bigger model.

6. Copies drift, live connections don’t

Before, our rules lived in several places at once: a project file in one assistant, custom instructions in another, a pasted document in a third. Each was a copy, and each copy was a little different.

The only setup that stayed consistent was a live connection. Every tool asks the brain at the moment it needs a rule, today over MCP (Model Context Protocol). Project files, custom GPTs and uploaded documents are snapshots. They’re fine to start with. They’re not a source of truth.

7. A hard approval gate stalls everything

We said from the start that nothing counts until its Owner approves it. Then we imported months of rules that the AI was already following. Holding all of them back until someone reviewed each one would have meant the AI suddenly knew less than the day before.

So we added a status for it: already live. A rule that is in use but not yet approved shows that label everywhere: in the AI’s answer, in the list, in the graph. Owners review them in batches. In our own setup, most rules still carry that label, and we’d rather see the honest number than pretend everything is signed off.

8. Someone still has to read the output

Automatic extraction gets most things right and some things strangely wrong. One of our “rules” turned out to be the extraction’s own comment on a candidate it had rejected. Interviews have a quieter trap: if the AI asks a leading question and the person says “yes, roughly”, the rule came from the AI, not the expert. Now only what the interviewed person said counts as evidence.

Neither problem is a reason to skip automation. Both are reasons to keep a person between draft and published.

What we’d do differently

We’d start even smaller. One Topic, one Owner, twenty rules, and every AI tool connected to it on day one. The value showed up the first time we stopped correcting the same mistake for the third time, and that happened with very few rules in place.

If you want to try this on your own company, the quick start in our company brain guide works without any software. When you outgrow a single page, that’s what CtxCore Brain is for.