Step one for AI in your business is not an AI tool

I have a day job in fintech that I love, two companies I am building on the side, and a personal agent assistant to help me run all of it. AI was supposed to mean less work. It has meant more, and somehow I am more energized than ever. This morning the information for those four things lived in a scattered Obsidian vault, a handful of code repos, Linear, my email, and my head.

All across X and LinkedIn right now, everyone is pushing products like a company brain, which clearly I need. But first you have to understand where all the data lives and get that context in order. That part is the problem, where the real work lives to make AI successful, and it is the part I want to dig into.

The foundation is the data

Adding AI agents in a business is a lot like building a house. It has to start with the foundation. If the foundation is not right, the rest will eventually fail.

The foundation is the data layer of the business. Not only is that layer the actual fields in a database but it is also all the context an agent needs to make a decision. That context is in emails, CRMs, paper files, and the CEO's head. AI agents are not effective if the data is not there.

Two examples from my own setup.

I run Claude Code, Codex, and a few custom agents I built every day. What they are doing and why should be data I can look at. Right now it goes off into the ether because I never set up the proper tracking when I started.

Product ideas. I start working on one, get distracted, and do not finish it. It lives in my head and scattered as markdown files in various repos and Obsidian.

What scattered looked like at bigger companies

One thing I learned that even in larger, regulated entities the data is even more strung out. There is context in SharePoint, ShareFile, actual servers, email, and Teams. Literally everywhere and not usually connected very well if at all.

It made it impossible to easily report on things or have a pulse of where everything was or where it was headed. Not because the systems were unreachable. Because nobody knew where everything was, so nobody could point anything at it.

I never got to run an agent there outside of CoPilot. My guess is it would have read the same mess. This is not a new idea. Bad data in, bad answers out has been true for decades. Agents just make the bill come due faster, because an agent acts on what it finds instead of a person who knows to double check.

The company brain is step two

A company brain can index what you point it at. It cannot tell you about the ShareFile nobody mentioned or the pricing rule that only the CEO knows. The pointing is step one. That is the part you have to do yourself, it's not fun, and it's not sexy but it is what will make or break your AI experience.

I decided to dogfood this and go through the pain of organizing my scattered context so hopefully I can share what I learn along the way to help others out and keep them from spending money on the latest AI trending topic that will cost an arm and a leg and still fail.

What step one looked like this week

Step one is an inventory, not a cleanup. Mine took a day.

The vault had too many top-level folders and I could not find anything in it. It now has five: one folder per company, one wiki for knowledge that crosses companies, one for the agent setup, one personal, one archive.

There is one registry file that answers "which project is this." Claude Code and Codex both read it, so they resolve a project the same way I do.

There is one shared memory store that every agent reads and writes. A lint check catches a bad entry before it is committed.

There is a trash folder of one-time items I had been keeping. It is still sitting there waiting for me to empty it.

What is still scattered: my head (it's scary in there), Linear, email, and the repos. Step one is not done. But a day of work got the vault from a mess to something I can find things in, the agents can find things in, and a foundation to keep building off of.

Same job at any size. The folder structure will not transfer to a company with eighty people and an ERP. The inventory will.

Where does yours live?

So, where does your data live today? This is my attempt to try and build in public and I would love to get any tips or tricks from anyone out there. Also, if there is anything I can share to help you out then let me know. The only way to really learn is to get dirty and do the work and if we do it together we can all get better faster!