Company Brain

What Is a Company Brain?

An AI with three things: your company's context, connections to the tools where the work happens, and skills for the jobs you repeat. So it does more than answer questions about your business. It does the work, and leaves it in your drafts.

Try the DemoGet a free setup plan
owed replies it caught
626
meetings it processed
460
emails it sent on its own
0

Nexrizen's brain, June 4 to September 30, 2026

Written by Gabe Pérez. I run my company on one. Updated October 2026.

The definition

Three Parts and the Rules

Each part is close to useless alone. A skill with no context writes a generic proposal. Context with no skills is a search box. Connectors are how it reaches the work in the first place.

context.dat
_ □ ×

Context

What your company knows: clients, calls, email, prices and the decisions you've made. It arrives on a schedule, so nobody has to remember to paste anything in.

Ours holds 727 call transcripts, 1,295 emails and one log of rulings.

connectors.sys
_ □ ×

Connectors

Access to the tools where the work happens: Gmail or Outlook, Drive, Zoom, Notion, QuickBooks. It reads by default, and anything meant for a client stays a draft.

Ours syncs email three times a day and call transcripts every hour.

skills.exe
_ □ ×

Skills

Written instructions for the jobs you repeat, like turning a call into a proposal or answering a lead with your real prices.

Ours has one command that reads each founder's day, files it and drafts the replies.

rules.ini
_ □ ×

And the rules that hold it together

  • It drafts. It never sends.
  • Each person sees only the clients they should.
  • Every answer cites a file, or says it isn't there.
  • Decisions live in one file everyone's AI reads.
Why the rules get their own file ›
Do I need Claude Code?

Runs on the Claude You Use

The brain is plain files and written instructions in accounts you own, so it isn't tied to one app. Your team reaches it from whichever Claude fits how they work.

claude.exe
_ □ ×

Claude, the regular app

Desktop, web or phone. A project holds the instructions, connectors reach your email and Drive, and skills hold the jobs you repeat. Nobody opens a terminal.

Where most teams should start.

Connecting email the safe way ›
claude_code.exe
_ □ ×

Claude Code

The same brain as a folder of files, with rules a script enforces and every change reviewed like code. It can run on a schedule with nobody's laptop open.

What we run Nexrizen on.

Claude Code vs. the Claude app ›
claude_duet.exe
_ □ ×

Claude Duet

Anthropic's new app that works in the background: it checks your connected accounts on its own and sends you a morning brief. For now it's a waitlist for personal Pro and Max plans in the US.

Good for one person's inbox today. Team plans are listed as coming later.

Duet vs. our day-end run ›

Whichever one your team uses, company accounts get connected from a company plan, and the rule doesn't change: it drafts, and a person sends. Our connector policy.

Isn't that just ChatGPT?

Answers vs. Work

Most AI at work stops on the second rung. The third rung is where the hours come back.

ladder.exe
_ □ ×

A lead from this morning's call wants a quote.

  1. Search Finds the file.

    14 documents mention pricing.

    Drive search, enterprise search

  2. A chatbot Answers from it.

    Your rate card says $150 an hour for website builds.

    ChatGPT, Claude or Copilot, out of the box

  3. A company brain Does the job.

    The proposal is in your drafts, scoped from the call and priced from your rate card, with the source under every line.

    The same Claude, with your context, connectors and skills

Watch It Do the Work

Same job, two AIs. It plays through on its own. Pick one to take over, or open a source to see where each line came from.

chatgpt.exe
_ □ ×

Draft a proposal for Bluebird Bakery from this morning's call.

Happy to help. Paste in your notes from the call and your pricing, and tell me what they asked for.

company_brain.exe
_ □ ×

Draft a proposal for Bluebird Bakery from this morning's call.

It's in your drafts: a new site with online pre-orders, live before their spring menu launch. Priced at your $150 rate, paid in three installments. Nothing goes out until you send it.

Try it on a whole agency's files
calls.exe — Bluebird Bakery intro
Today, 9:30 AMBluebird Bakery + Rosa

Bluebird: People keep calling to pre-order cakes. We want that on the site.

Bluebird: The spring menu launches in March, so ideally before then.

Rosa: Makes sense. I'll send you a proposal this week.

drive.exe — 2026 rate card
Drive / SalesLast edited Jan 6

Website builds: $150 an hour.

Fixed-price projects are paid in three installments.

outlook.exe — Hosting after launch?
From: Harbor DentalTo: Tom BeckerTuesday, 2:12 PM

Hi Tom,

What will hosting cost us each month once the site is live?

Thanks!

outlook.exe — Menu PDF
From: Kiln CoffeeTo: Rosa DelgadoMonday, 10:41 AM

Can you send the menu as a print-ready PDF? Our printer is asking.

outlook.exe — Re: Photo shoot
From: Tom BeckerTo: Bluebird BakeryOct 3, 4:05 PM

Confirming what we said on the call:

Two rounds of product photos are included in the project.

Any extra rounds are billed at our photography rate.

Under the hood

What's in the Folder

Underneath, it's a folder of plain files. Ours is a git repository that Claude Code works in. Yours can just as well be a shared Drive folder your team's Claude reads. Here's the shape for Fernhill Creative, the made-up agency from the demo. Click around.

explorer.exe — fernhill_brain
_ □ ×

fernhill_brain / CLAUDE.md

The first page it reads

The instructions every session starts from. Claude Code reads it as CLAUDE.md, and in the Claude app the same text goes in your project's instructions. It says what's in each folder, where new things go and which rules to follow. Ours started at 58 lines and is 223 now, and most of those lines exist because something went wrong without them.

# Fernhill brain
Clients live in clients/<name>/. Raw sources never get edited.
Drafts go in drafts/. Nothing is ever sent.
Before quoting, read rulings.md and sales/rate-card.md.
What goes in a CLAUDE.md ›
12 items Mostly plain markdown
Who feeds it

How It Stays Current

A brain that depends on someone remembering to paste things into it stays empty. Ours is fed by scheduled jobs and one command each founder runs at the end of the day.

Scheduled Tasks
_ □ ×
TaskScheduleLast result
Sync emailThree times a day 1,295 emails filed by client
Pull meeting transcriptsEvery hour 727 transcripts so far
Confidentiality gateBefore any transcript is saved A required check on every pull request
Day-end processEach founder, end of day 39 runs, 460 meetings, 1,420 email threads
Nightly janitorNightly 245 open findings, no pull request since Aug 6
About a dozen scheduled jobs Heavy jobs run remotely

Feeding it is automated. Keeping it tidy still takes a person: the janitor won't move a file something links to, and most of its findings are exactly that. The janitor's numbers.

I already have a second brain

Second Brain or Company Brain?

A second brain is one person's notes with an AI on top. It works well right up until a second person writes to it.

When we started logging company decisions in July, six of the first ones came out of one person's private AI memory. One was the rule that the AI never sends client email. It had been governing that person's sessions and nobody else's.

compare.txt
_ □ ×
Second brainCompany brain
Who writes to itYouFounders, teammates and scheduled jobs
Where the rules liveYour head, or your AI's private memoryOne file everyone's AI reads
The same meetingFiled onceFiled twice, unless both runs share keys
Who sees whatEverything you can seeEach person sees only their clients
When someone leavesNot a problemA checklist against the access registry

More on this: Two founders, one brain · Shared, personal and secret settings · Where tribal knowledge hides

What You'd Try First

what_youd_try_first.txt
_ □ ×
  • Claude Team or ChatGPT BusinessGood search, and this runs on top of it. Each person searches what they can already see, and nothing carries over between chats unless someone files it. If your Drive sharing is a mess, the answers will be too.
  • Copilot or GeminiStrong inside Microsoft 365 or Google Workspace. Calls on Zoom and files kept anywhere else are yours to bring in.
  • Second-brain kits and coursesBuilt for one person, and you do the setup. This is for a whole team, and I do it.
  • Glean and other enterprise searchBuilt for companies with an IT team and an enterprise budget.
  • Hiring an ops personCosts more each month than this does once, and still has to ask around.

All of these stop at answering questions. At Nexrizen, ours does the work too: a discovery call becomes a proposal with a prototype the same week, and every evening the day's action items turn into tasks with the replies already drafted.

Do you need a knowledge graph on top? We tested one on 46 real documents and turned it off: two identical runs agreed on 16% of the labels. The test.

The part vendor pages skip

What Breaks

All of these happened to ours, and each one has a write-up with the numbers.

knowledge_graph.exe
×

Two identical runs agreed on 16% of node labels. A graph that renames itself every night can't be trusted.

Read what happened
tasks.exe
×

Created 1,042 tasks in four months. Fixing it took a script, not a better prompt.

Read what happened
transcripts.dll
×

46 confidential transcripts reached main. The gate worked. It just wasn't in the way.

Read what happened
learning.log
×

Three feedback loops built to make it learn. Two months later, two logs were empty and the third never fired.

Read what happened
README.md
×

Called a job "Planned" that was running every weekday. The AI believed it, and a new report nearly went blind to overdue projects.

Read what happened
notion.sys
×

Couldn't see 19 of 36 databases for months, and nobody noticed.

Read what happened

None of it made the brain less useful day to day. It does make the honest pitch smaller than most: it answers with sources and drafts the work. It doesn't improve itself on its own, whatever the brochure says.

Am I too small for this?

Do You Need One?

If you're one person with good notes, probably not. Tick what's true for your team.

brain_check.exe
_ □ ×
If you're one person with good notes, probably not. Tick what's true for your team.
0

of 6 are true for your team

You probably don't need one yet. A good notes app and a CLAUDE.md will carry you a long way.

Start with one folder
Is this a fad?

Where the Idea Came From

The name is new. The pieces have been arriving for years, and each one fixed part of the last one.

setup.exe — History
_ □ ×
  1. 2019

    Enterprise search

    Glean is founded to search across all of a company's tools at once. Finding things gets easier. Keeping them right is still your job. Source ↗

  2. 2022

    Building a Second Brain

    Tiago Forte's book makes personal knowledge systems mainstream. One person, one set of notes. Source ↗

  3. December 2025

    Context graphs

    Foundation Capital argues the valuable record is why decisions were made, not only what happened. Source ↗

  4. April 2026

    The LLM wiki

    Andrej Karpathy posts a pattern where an AI keeps a markdown wiki current instead of searching raw files every time. Source ↗

  5. 2026

    GBrain

    Garry Tan open-sources the memory layer he built to run his own AI agents. Source ↗

  6. Now

    A brain that does the work

    Context and connectors, plus skills that turn them into proposals, replies and follow-ups, sized for a team of 20.

Setup is complete.

Questions People Ask

Help Topics
_ □ ×
Is Glean a company brain?
It's the closest enterprise version: search, an assistant and agents across a company's tools, built for companies with an IT team and an enterprise budget. A company brain for a small team does the same job on the Claude and the tools you already pay for, sized for 20 people instead of 2,000.
How is it different from a second brain?
A second brain is one person's notes. A company brain has several writers, so it needs rules everyone's AI reads, keys that stop two people filing the same meeting twice, and walls so each person sees only their clients. How we handle two writers.
Do small teams need one?
Not a one-person shop with good notes. It starts paying off at the second person answering clients, or when one person is who everyone asks. The checker above is honest about it.
Do I need a knowledge graph?
We tested one properly and turned it off. We kept plain search with grep underneath, plus a simple link graph that gives the same answer every time. The test, with the kill criteria.
How does it stay up to date?
Scheduled jobs bring in email, calls and docs, so nobody has to remember. What still needs a person is tidying, and deciding which document is the real one when there are two.
Does it send emails or do things on its own?
No. It writes the proposal, follow-up or reply and leaves it in your drafts. You read it and hit send. Why it's built that way.
Will it make things up?
It answers from your documents and shows the source. When the answer isn't there, it says so. Ask the demo something it can't know.
What does it cost?
I build one for a team of up to 20 from $3,997, live in 14 days or all your money back. Ongoing care is optional at $497 a month, plus your own AI account. Every price on one page.
Do I need Claude Code?
No. Your team can use it from the regular Claude app on desktop, web or phone. Claude Code is for whoever wants the rules enforced by scripts and every change reviewed, which is how we run ours. Claude Duet can work from it too, once it reaches team plans. How the options compare.
Can I build one myself?
Yes. Start with one folder and a CLAUDE.md and the free Claude Code guide. If you want help, coaching is four 90-minute sessions for $1,997 and you do the typing.
Field notes

How Ours Was Built