How we run a company on a company brain. With the real numbers and everything that broke.

I priced 56 Fiverr gigs that promise an AI trained on your business, from $30 to $20,000. What each tier buys, and six questions to ask before you pay.

A real company brain example: the folders, the jobs that feed it and the commands that read it, after 3,474 commits. And what still breaks.

We built three ways for our AI to learn from its own work. Two months later two logs were empty and a hook never fired. Loops that fill ride on daily work.

Start with a folder of markdown and a CLAUDE.md. Once a team wrote to ours, we added folder rules, a hook and a nightly janitor. It still has 245 findings.

Our decision log is one append-only file of rulings our AI loads before drafting proposals. Six of the first came from one person's private AI memory.

Tribal knowledge is what your team knows and never wrote down. Ours hid in five places, including a README that said 73 rules when the file had 151.

Our company brain broke more often from bad inputs than from the model. Eight times a tool returned a wrong answer with no error, and the check we added.

Claude can send Gmail for you as of August 2026, with approval on by default. Ours found 626 replies we owed, drafts them, and sends none.

Our AI turned meetings and email into 1,042 tasks in four months, mostly noise. Telling it to bundle failed. A script that does the counting worked.

My co-founder and I each run an AI day-end process over the same clients. Keeping it from filing everything twice took a shared table and cross-inbox keys.

Our Claude Code GitHub Action reviewer and a nightly routine looped on one PR for 21 hours: 147 review runs, about $500. The fixes we made on both sides.

We piloted an LLM knowledge graph on 46 documents with kill criteria set first. 89% of edges were true, but a rerun matched 16% of labels. Search won.