Build a Company Brain

Most of the AI implementation work I do for SMBs is unglamorous. Sitting with people, watching them work, asking how they make decisions, writing it all down in markdown files, and dropping them into a Claude project organized by function. Markdown is just plain text with light formatting, the kind of thing you'd write in TextEdit on a Mac or Notepad on Windows. I wrote about one of these implementations last week and used the term "company brain" to describe what I'd built.

Yesterday, Y Combinator dropped its Summer 2026 RFS and "company brain" was on the list, with Tom Blomfield making the case for it as a category they want to fund. So this week, I thought I'd go deeper on what a company brain actually is, what it looks like in practice, and why Jack Dorsey thinks the entire shape of a company is about to change because of it.

Blomfield argues the biggest blocker to AI automation isn't the models. The models are really impressive right now, and they’re the worst they’ll ever be. The blocker is domain knowledge. Every company has critical know-how scattered everywhere. Some of it lives in people's heads. Some is buried in old email threads, Slack channels, support tickets, and databases. Companies work because humans vaguely remember where the knowledge is and how to apply it.

AI agents don’t operate like that. If you want a company to actually run on AI, you need a new primitive. A living map of how the company works. How refunds get handled, how pricing exceptions get decided, how engineers respond to incidents. Said differently, a structured representation of how the business actually operates that AI can read and act on.

Blomfield calls it a company brain. Jack Dorsey calls it a mini-AGI. Different names, same primitive. In an increasingly AI-native world, this will be one of the few remaining moats.

As I mentioned at the top, last week I described what I built as a "company brain." Reading Blomfield's RFS, I think that was a misnomer. What he's describing is more ambitious than what exists in any tool today. Not a search layer over connected systems or a knowledge base with documented processes. An executable map of how the company works, kept current as the business changes, structured so AI agents can use it to do the work safely and consistently.

The AI knowledge base I typically build has two kinds of projects working alongside each other inside Claude.

The first project type is enterprise search. On Team and Enterprise plans, Claude has a preconfigured project called "Ask [Your Company]" that an admin sets up by connecting work tools, like Google Drive, Slack, and Microsoft 365. When someone asks "What's our policy on contractor travel reimbursement?", Claude makes live calls to the connected systems and synthesizes an answer with citations from across them. Permissions are respected at the source, meaning users only see what they already have access to.

The second is a set of Claude projects organized by function. Operations, finance, sales, HR, client policies. Inside each, plain markdown files for the standard operating procedures, the policies, the contacts, the recurring decisions. This is where you put the knowledge that lives only in people's heads, the kind of thing enterprise search can't find because it was never written down anywhere.

The work isn't technical. It's simply documentation. Most of what I do in the first week of an implementation is sit with people and watch them work, then write down the processes that matter. Most of the time, the person who runs them has been doing it for ten years and it lives in their head. Once it's written down, the company has something it didn't have before. So does Claude.

The two project types are perfectly complementary. Enterprise search finds what already exists in connected tools. The function-organized projects capture what didn't exist anywhere except in someone's head.

Jack Dorsey and Roelof Botha published an essay at the end of March on how Block is rebuilding itself as a company. The thesis is that organizations have been hierarchically structured for two thousand years because humans were the only available coordination mechanism. Span of control, layers of management, information routing up and down. The Roman army figured it out, and the Prussians refined it. The American railroads commercialized it. Every modern company still runs on it.

Dorsey's argument is that AI is the first technology powerful enough to actually replace what hierarchy does. Block is building what they call a company world model, a continuously updated representation of how the business operates, and using it to coordinate work in ways that previously required humans relaying information through layers of management. This is the closest thing to Blomfield's company brain that any company is publicly working on.

If it works, the org chart collapses to three roles.

ICs build and operate the capabilities, the model, and the interfaces. DRIs own specific cross-cutting problems and customer outcomes for fixed periods, with authority to pull resources across teams. Player-coaches combine doing the work with developing the people around them. There is no permanent middle management layer because the world model carries the information that managers used to relay.

In February, Dorsey cut 40% of Block's workforce, one of the largest AI-driven layoffs yet in the S&P 500. The framework above is the thinking behind it. Whether or not Dorsey's full vision plays out, the underlying primitive is the same one your company needs. A real company brain is what makes that future (and associated org chart redesign) possible.

Pick five processes in your work that live only in your head. Write them down in plain markdown. Drop them in a Claude project.

That's the smallest possible version of a company brain. It's yours, scaled to one knowledge worker. You don't get the enterprise search layer at the individual level, but you get the part that matters written down — maybe for the first time.

The anti-Grammarly went viral. Sinceerly, spelled with a second “e”, is a Chrome extension that takes AI-generated emails and adds typos, kills em-dashes, and breaks the "not just X but Y" construction so your messages stop reading like AI. It was built in a month by Ben Horwitz, an HBS student and Dorm Room Fund partner. The tool has three modes (Subtle, Human, CEO) and Horwitz claims he tested it by cold-emailing five Fortune 500 CEOs. Four replied. Two of the replies had typos. Maybe they were also using Sinceerly. For more, check out his recent TBPN appearance.

Taylor Swift trademarked her voice. TAS Rights Management filed three trademark applications on Friday. Two are sound marks ("Hey, it's Taylor Swift" and "Hey, it's Taylor"). The third is the pink-guitar-and-bodysuit Eras Tour image. The strategy follows Matthew McConaughey, who filed similar trademarks in January as the first A-lister to use trademark law as an AI defense. Copyright protects existing recordings. It doesn't protect against AI generating new content that mimics a voice. Trademarks might. The legal theory hasn't been tested in court yet, but expect more of this.

Working on a Claude implementation for your business? Shannon Advisory is the AI consulting practice I run for SMBs. Schedule a call here or email [email protected].