Company brains / 10 Jun 2026

How to make your company's knowledge AI-queryable

Every company wants the chatbot. The chatbot is step two. The first AI transformation is not a chatbot. It is making the work legible: turning what your company knows into something a machine can look up. Get that wrong and the chatbot is just a confident stranger with a search box. Get it right and almost everything else becomes possible.

The real constraint

The models are not the bottleneck anymore. Context is.

Ask a frontier model a question about your business and it answers like a smart stranger. Not because it lacks ability, but because everything that makes your business yours is invisible to it:

• The product truths live in the founder's head.
• The pricing logic lives in an old email thread.
• The tone lives in whoever has been answering customers the longest.
• The decisions live nowhere at all. They were made in a meeting and never written down.

A company like that cannot be helped by AI in any serious way. There is nothing for the intelligence to grip. The instinct is to reach for a smarter model or more data. Neither fixes it. The fix is legibility.

Queryable is not the same as having documents

Almost every company has documents. Very few are queryable. The gap between the two is where most AI projects quietly fail.

• A document is written for a human who already has context. Queryable knowledge is written for a reader who has none.
• A forty-page PDF is storage. Queryable means the same knowledge broken into small, self-contained units, each answering one question, each findable without reading the other thirty-nine pages.
• Tipping every document you own into a vector database feels like progress and usually is not. Retrieval can only hand back what was written clearly in the first place. Vague in, confidently vague out.

So queryable means three things at once. The knowledge is written down, structured into retrievable units, and legible to a reader with no prior context. Plain text is enough. It does not need a platform.

What to actually write down

Inside Evolve Skateboards this layer is called Company Brain. Plain markdown files, grouped by the kind of question each one answers:

• Product truths. Specs, compatibility, what fits what, the claims that must never be wrong.
• Pricing and policy logic. What costs what, how warranty works, refunds, shipping, the awkward edge cases.
• Brand voice. How the company sounds, and the things it never says.
• Routing rules. Which questions a machine may answer, which get handed to a human, which get ignored entirely.
• Postures for difficult situations. Safety, legal, complaints. The cases where the wrong word is expensive.
• Decisions already made, with the reasons behind them. So they are not re-argued from scratch every quarter.

The reasons matter as much as the rules. A rule with no reason attached gets overridden the first time it is inconvenient.

The model is the last layer, not the first

This is the part people skip. A queryable company is not consumed by handing everything to a model and hoping. The systems I trust read in a strict order:

1. Deterministic rules first. Is this even a question a machine should answer at all? Pricing, warranty, legal and safety cases are routed by fixed rules, never improvised.
2. The knowledge layer second. If the answer is written down, the agent looks it up. It does not guess.
3. The model last, and only when the case is safe and not already covered. Its job is to compose, inside the boundaries the first two layers have already set.

This is why one team can run agents for blogs, email, social comments, ads reporting and store operations without the agents contradicting each other. They are all reading the same brain.

It compounds, and it stays consistent

The best property of a shared knowledge layer is that improving it improves everything at once. Write down one new product truth and the blog agent, the comment agent and the support agents all know it the same day. Nobody updates five systems. They update one source.

The layer can even feed itself. The customer service agents write help centre articles as they work, and those answers, which used to live in one person's head, flow straight back into the layer. The company gets more legible as a side effect of simply operating.

You do not need to buy software for this

There is an industry forming around this idea, and most of it is unnecessary.

The expensive part is not storage or retrieval. The expensive part is judgement: deciding what is worth writing down, and the discipline to keep it true as the business changes. A folder of markdown files maintained by someone who understands the business beats an enterprise knowledge platform maintained by nobody.

A knowledge layer is not a project you finish. It rots the moment nobody tends it. The value was never the files. It is the person who keeps them honest.

Start embarrassingly small. Write down ten things: the decisions you keep re-explaining, the product facts that must never be wrong, the answers your team repeats every week. Here is a prompt to pull the first ones out of your own head.

The rule

A company that can be queried can be automated. A company that lives in people's heads cannot.

Make the work legible first. Everything else gets easier.