A Thin-Brained Idea

I recently gave a few talks on the difference between fat and thin company brains. This post is the written version.

Every company I talk to wants an AI company brain. And not just any brain - what I call a fat brain.

The dream is a company brain that mimics a human one. One mega brain that holds all the information. CRM, ERP, finance, product, support, marketing. Everything. (Not every company thinks this way, but it's often the vision of the CEO or the internal AI champion.)

Intuitively, a fat brain makes sense. Humans have one brain that holds everything we know. Why shouldn't a company brain work the same way?

When delving into AI company brains, my first reaction was the same. A fat brain sounds amazing. If an AI brain had all of a company's information cleanly filtered, indexed and summarizable, imagine what it could do. A whole wave of AI brain products is selling exactly that. Give us everything and the brain figures it out.

After helping a number of companies build their company brains, my opinion has changed. Fat brains do not work at scale; thin brains do.

The combined simple brain works when you're small

After helping a number of companies build their company brains, my opinion has changed. Fat brains, which pull everything into one place, run into problems at scale. Thin brains - which leave the source data where it already lives and just know where to look - seem to work much better.

Look at X and you'll see dozens of posts about people building their own "AI Brain" for themselves or their small company. And yes, Fatbrains can work for small companies... They are cute, friendly and fun.

I built a personal brain the way a lot of people have lately, inspired by Andrej Karpathy and projects like GBrain. Markdown files. Simple folders. Obsidian on top. Custom CRM, custom processes etc. For the most part, it works. My AI tools read it, write to it, and know far more about my world than they did before.

Tiny teams can do the same thing. Put a markdown wiki into GitHub or Google Drive and give everyone on the team access to it. Everyone commits, everyone reads, and every agent reads too. Under about 20 people and assuming everyone has Github access, it can work decently well.

Past 20 people, the Fat Brain falls apart

But the simple brain that works for a small team breaks as the company grows.

Once a company gets bigger, it runs into a lot of new complexities. Here are a few:

• Permissions. Permissions become a mess with larger companies. With 3 to 10 people, everyone can basically see everything. As the company grows, private data starts showing up that needs to be gated: salaries, customer contracts, employee data, finance. Per-file access gets messy fast.
• Auditing. Larger companies often have compliance requirements like SOC 2 that demand an audit trail. Google Docs and Sheets keep version history, but auditing still gets hard. People can't just overwrite each other in a shared doc. And Agents overwriting files complicates this further. You need to know who changed what, when and why, across thousands of files and a growing number of AI agents making edits.
• Existing tools break down. GitHub works for engineers, but sales and marketing won't touch a CLI, let alone have a GitHub Account. Microsoft Excel and Word (or Sheets and Docs) work until twenty people are editing the same file. You end up stretching tools past what they were built for.
• Search scale. An AI can search a few hundred markdown files with simple text search. But once the whole company is pouring notes, specs and meeting summaries into the same repository, it grows to thousands of files. Simple search on Google Drive or GitHub stops working, and the AI can no longer find the one pricing decision someone wrote down last spring.
• Conflicting truth. As a company gets bigger, you start seeing more and more conflicting data, and the AI gets confused about which answer is the right one. The CRM shows ARR. The investor deck shows ARR. They're different numbers. Which one is correct? Often both are. There's the number sales tracks, the number you report to investors, and the number you put in a press release. The same deal can look different in a tax view and a non-tax view. A pile of markdown has no idea which one you want.

Enterprise SaaS wants to sell you a fat brain

The intuitive answer to all of this is to build a fat brain. It will hold everything and index everything semantically! Hundreds of enterprise SaaS startups are lined up and foaming at the mouth to sell you one: a mega data project that holds all your data. All your customer data (no need for a CRM!). All your product data, support data and development data. Everything organized and cross-referenced. It will be beautiful (not just for you but also for their company as they will lock you in forever!).

The brain becomes the canonical source of truth. Workflows live inside it. It's centrally managed by a new team of brain architects and brain managers, either in-house or run by a vendor.

Fat Brain Problems in Production

The fat brain is a great dream. It runs into trouble the moment you try to build it. Here are the patterns I see stop it:

Tool loyalists

Some people have built entire careers on their expertise in one specific tool. Many salespeople are Salesforce experts, often for ten years or more. They are never happy switching to another CRM, let alone a vibe-coded custom one or to go without their well loved shortcuts and tools. They drag their feet and make any migration to a brain as hard as possible. RevOps, retention and marketing teams are just as tied to their tools.

The IKEA effect

A problem just as strong as tool loyalists. A lot of internal systems start bottom-up. Someone builds a spreadsheet. It moves to Airtable, grows, and gets customized. Like someone who spent hours putting together an IKEA table, the employees who built it love what they made and don't want to get rid of it. Whoever built it won't be eager to watch their baby get shut down. On top of this, every company routes its work differently, so moving to a centralized company brain means migrating each of these homegrown tools, and in many ways recreating them. AI can do a lot of the replication, but getting it to match what exists still takes iterations and a lot of testing. That's a large amount of time spent replacing something that worked just fine.

The microwave problem

I think of a fat brain like a microwave. It's a wonderful general-purpose tool that heats everything. But not everything heats the same way. For some things, like popcorn, a microwave works incredibly well. For others, like a nice steak or a piece of salmon, it's the wrong way to cook.

Data works the same way. Different kinds of data work better with different architectures. A CRM is built around pipeline stages and forecasting. A finance system is built around double-entry ledgers and audit controls. Product analytics is built to handle millions of events. Each of those tools has had years of optimization for its one job. They're all tables underneath, but they aren't doing the same job. Move them into an all-purpose fat brain and you throw away the years of tuning that made them good at it.

Department fiefdoms

As a company grows, every department (and often every sub-department) develops its own strong opinions about which tools to use. You can try to mandate a company brain from the top down, but you'll run into major resistance from department heads. Each of those heads has strong opinions about which toolsets work and what matters most for their department, and a one-size-fits-all brain rarely matches them.

The champagne tower

Finally, a fat brain is top-down. Someone at the top has to make the organizing decisions: the schema, the taxonomy, what flows in, who owns what. You often end up needing brain administrators and managers just to make sure everything lands canonically in the right place, and then to keep maintaining it. Like a champagne tower, the value only reaches the glasses at the top after every row below has been filled.

But that isn't how AI is being adopted. It's happening bottom-up. Teams and individuals are already making their own AI decisions, picking tools, building workflows and wiring up agents. A fat brain asks all of them to stop and wait for the top of the org chart to catch up.

The thin brain leaves the data where it lives

These complications make a fat brain almost impossible to implement today at a larger company. What I see working much more easily is what I call a thin brain. Canonical data stays at the endpoints. Customer data stays in Salesforce. Product data stays in the product database. Support stays in Zendesk.

The brain is a layer on top. It holds the policies, the playbooks and the skills. It knows where to look. It reads from those systems and can also push changes back, so everything stays in sync.

If the fat brain is the Library of Congress, the thin brain is the librarian. A good librarian doesn't own the books or memorize every page. They know where everything is, have a working sense of what's in each section, and can give you the gist before sending you to the right shelf. Ask which number is right and they know which book has the answer, and which edition to trust. And when something is out of date, they get the catalog updated.

Go back to the ARR problem. A thin brain doesn't need to own the number. It needs corporate knowledge: ARR for sales comes from one system, defined one way. ARR for investors comes from another, defined another way. Tax has its own rules. Where it helps, the thin brain can also pull those into named views (sales ARR, investor ARR, tax ARR) so no one, human or agent, has to guess which number they're looking at.

If the thin brain sounds familiar, that's OK. It's exactly the direction Anthropic and OpenAI are going. Neither is building a big company brain for you to import all your data into. They keep their own storage light (summaries, tables, artifacts) and connect to everything else through MCP connectors.

The obvious pushback is that this is just a data lake with a new name. It isn't. A lake copies data out of Salesforce, NetSuite and Zendesk, and the permissions don't come with it. Suddenly one system holds everyone's salary and every customer contract. A thin brain can query each system as the person asking, so Salesforce still decides what a sales rep sees. And a lake only feeds dashboards. A thin brain writes back.

Data, analysis, action

If the brain is thin, the stack becomes three layers people can build on.

• Data. Where the data already lives. Salesforce, HubSpot, Excel, Postgres, Supabase, Notion, Google Drive, TableThis. It's messy and spread out, and that's fine.
• Analysis. Summaries, churn scores, prioritization, dashboards. Pulled from the sources on demand instead of stored in one giant copy.
• Action. Agents on top that do the work.

(Quick Shameless Plug) Based on our analysis of what we've seen missing in company brains, we built out TableThis, a shared backend for tables and markdown documents that people and AI use primarily through MCP connectors. Think of it as Supabase plus Git for your intranet. It's one player in that layer, a home for the data that doesn't have a good one yet. It isn't trying to be the brain.

The company brains that win won't store everything. They'll send each question to the right system, more like a midbrain passing signals than a giant memory. It's a messier world than one big brain. I think it's a better one, because the knowledge stays with the people on the front lines who use it.

Thin wins

Plenty of companies will try to build the fat brain over the next couple of years. Some small ones will make it work. My bet is that most larger companies end up with a thin brain, whether they planned one or not.

If you're figuring out a company brain for your own team, this is a big part of what I do through Second Coffee consulting. You can see how I approach it at consulting.secondcoffee.ai/ai-brains.

If you're building a company brain and see it differently, I'd like to hear why. DM me on X at @imcharliegraham.