Company brains / 20 Jun 2026

How to clean and structure company data for AI

Last time I argued that the first AI transformation is making your company queryable. The fair reply is to ask where the queryable knowledge actually comes from. It does not exist yet. Right now it is scattered across a dozen tools and a handful of people's heads, and a lot of it contradicts itself. Before a company can be queried, someone has to drag everything it knows into the open and clean it. That job is boring. It is also the whole game.

Two reasons the knowledge isn't there

Two forces sit between you and a queryable company.

The first is the model. It is brilliant, and it knows nothing about you. Ask it anything specific to your business and it answers like a confident stranger, because everything that makes your business yours is invisible to it.

The second is your own software. Decades of business tooling have left your information in shards. The customers in the CRM. The policies in a Google Doc. The real returns window in a Slack thread from March. The product specs in a supplier email, and a different number on the website. The pricing logic in the founder's head. Each tool holds a piece. None of them agree, and none of them talk to each other.

That second force has a plainer name. Dirty data.

Fragmentation is the dirty data

We treat dirty data like a technical problem. It is really just the natural result of running a business across a decade of disconnected tools. It looks like this:

The same fact in three places with three different values.
Knowledge stored in formats nothing can read together. A PDF here, a spreadsheet there, a rule that was only ever spoken aloud.
Entries that were true two years ago and were never corrected.
Duplicates, near-duplicates and contradictions nobody has noticed, because no two of them were ever in the same room.

This is why tipping it all into a vector database changes nothing. Retrieval can only hand back what was clear to begin with. Dirty in, confidently dirty out.

Dump it all out first

Here is the part people want to skip and cannot. To begin, you have to dedicate a person to this. Not a tool. A person. Their only job, for a while, is to get every bit of the company out into the open.

Dumping looks like:

Exporting the systems. The helpdesk macros, the CRM fields, the spreadsheets, the policy docs.
Emptying the heads. Sitting with the founder and the longest-serving people and writing down what only they know.
Gathering the scraps. The email threads, the decisions buried in Slack, the PDF on someone's desktop.

Do not clean as you gather. Get it all out first, mess and all. The dump and the clean are two different jobs, and trying to do both at once means you do neither well.

Clean it with the machine

You do not clean a pile like that by hand. This is exactly the work a model is good at, and it is one of the most useful things a business could point one at right now. Whether you do it in Claude or in GPT does not matter.

The loop is simple. You hand the model the raw, contradictory pile and it does the grind:

Pulls it apart into small, self-contained notes, each answering one question.
Finds the duplicates and near-duplicates.
Surfaces the contradictions instead of smoothing them over, so a human can decide which version is true.
Flags what looks stale, and what is plainly missing.

You supply the judgement: which number is right, which rule still holds, what to throw away. The model supplies the labour: the restructuring, the reconciling, the reformatting nobody wants to do. That division is the whole trick. The machine cannot know the returns window changed in April. You cannot face restructuring four hundred notes by hand. Together it gets done.

What I am doing about it

This is not theory I am admiring from a distance. It is most of the unglamorous work in front of me right now.

In the business, it means pulling scattered knowledge into one place and working through it with a model, file by file, until the Company Brain says one true thing per question instead of three half-true ones.

In my own second brain it is the same move at personal scale. The same idea as building a digital brain, turned on myself: get enough context out of my head and into plain markdown that a model opening the folder has something real to grip.

The tool matters less than you think

A word on software, because it is where people stall. Markdown files in a folder. Obsidian. Whatever comes after it. The means by which we store and explore this context will keep expanding, and it will keep getting better. None of that is the point. The point is that the context exists, written down and cleaned, in a form plain enough to outlive whatever tool you happen to keep it in today. Pick the simplest thing that lets you start, and start.

Prompt you can use

I am cleaning a messy pile of company knowledge so AI tools can actually use it.

I will paste in raw, unstructured information from different places: exported docs, email threads, notes, things I have typed from memory. It will be contradictory, duplicated and out of date in places.

Do not tidy it silently. Work through it with me like this:
1. Break it into small, self-contained notes, each answering one question, each understandable with no prior context.
2. Group the notes into clear categories.
3. Where two sources disagree, do not pick one. Show me the contradiction and ask me which version is true.
4. Flag anything that looks stale, and list the obvious gaps where a question has no answer yet.

Give each note a short title, the answer in plain language, and the reason behind it where there is one. Start by telling me what you want me to paste first.

The rule

The first hire for AI is not a model. It is someone to clean the data.

Get everything out of the systems and the heads. Clean it with the machine. What you are left with is structure, and structure is the only thing you can build on.