Company brain: build it in-house or with an agency

You've realized a company brain is something you genuinely need. Your company's knowledge is scattered across chats, emails, policy PDFs, revenue spreadsheets, and above all, inside the heads of two or three people who'd leave a gaping hole if they quit tomorrow. The real question now is different: do you build it in-house, or bring in an agency?

This is a founder's decision, not a tech hobbyist's. A second brain isn't a well-organized shared folder. It's a system designed to be read and navigated by an AI, where every structural choice ripples through the quality of every answer for years. Get it wrong at the start and you're building on crooked foundations, and you won't notice until redoing everything gets expensive.

In this article I'll walk you through what it actually takes (structure, atomic notes, ontology, quality control, RAG, compliance), where DIY runs out of road, and why there's one thing you can never delegate, not even to the most powerful AI: understanding your own business.

Why "build or agency" is the right question

Most people treat this as a choice between two tools. It isn't. What's really at stake is the method. A company brain lives or dies on its architecture, and the architecture doesn't show itself while the system is still small. With 50 documents, almost any layout looks like it works. The problem explodes at 2,000, when the AI starts pulling the wrong note, mixing up two clients with the same name, or making things up because it can't find the right source.

That's why this isn't a matter of installing software, it's a matter of design. And design requires three skills that rarely live in the same person: understanding how a machine reasons over text, knowing how to structure knowledge so it stays reusable, and understanding your specific business. The third one, as we'll see, is always on you.

What it actually takes: the five layers nobody should improvise

Before deciding who builds it, you need to know what you're building. A serious company brain rests on five layers. Each one has a right way and a hundred wrong ways of being done.

1. Atomic notes: knowledge broken into reusable units

Knowledge doesn't pour into a system like water into a bucket. It's broken down into many small notes, one idea per note, all interlinked. It's the same principle sociologist Niklas Luhmann used with his Zettelkasten method to manage 90,000 paper index cards and write dozens of books from them: every card stood on its own and linked to the others. A well-made atomic note is reusable across different contexts and, crucially, navigable by an AI. A bad note (too long, three concepts crammed together) confuses the system and degrades every answer built on it. If you want to go deeper on the logic, we cover it in detail in the article on atomic notes as the building blocks of company knowledge.

2. Ontology: the difference between archiving and reasoning

This is where DIY falls apart. There's a real difference between taxonomy (how you file things, which folder they go in) and ontology (how concepts connect to each other). You already know how to do taxonomy: it's just organizing. Ontology is the web of connections that lets the AI move between notes and reason, going from a client to their industry, to the contract, to the policy that governs it. Without a designed ontology, you have a tidy but mute archive. With a well-designed ontology, you have a brain that connects the dots. It's the most technical part, and the easiest to underestimate.

3. Canon: a single source of truth

The canon, or single source of truth, is the one company truth the AI draws from. It's what stops the system from making things up: the AI doesn't imagine, it only reports facts present in the company's knowledge base, drastically cutting down hallucinations. You need a clear rule for what enters the canon, who updates it, and how contradictions get resolved (two documents stating different prices for the same thing). Without governance, the canon gets polluted and the AI starts answering with stale data.

4. Living memory: a system that updates itself

A company brain isn't a snapshot taken once. It's a living memory that updates itself through daily conversations and work sessions. It's what lets you, six months later, ask "what did we decide with that client back in March?" and actually get an answer instead of digging through a dead chat thread. But for that to work you need a reliable update mechanism and version control that keeps a history and guarantees a single source of truth even when your whole team is working in it.

5. RAG: semantic search to scale to thousands of documents

While notes are few, a content map and an index are enough. Once they multiply, you need RAG (retrieval-augmented generation): semantic search that, across thousands of documents, surfaces only the ones truly relevant to the question, efficiently. As a rough order of magnitude: under 500 notes, a content map and an index will do; between 2,500 and 20,000 notes, you need to introduce embeddings and RAG; past 20,000, you need a full RAG pipeline. Setting up RAG too early is wasted effort; setting it up too late leaves you with a system that stops answering. Knowing exactly when it's needed is precisely the kind of call that separates people who do this for a living from everyone else.

What going it alone actually costs you (in time)

Scattered knowledge already has a cost you pay every day, even without a company brain. McKinsey estimates that around 19% of the workweek, nearly one day out of five, is spent just looking for information. It's an order of magnitude, not gospel, but it makes the point: that lost day is the price of informational chaos.

Then there's the cost of onboarding. A new hire takes on average 8-12 months to become truly productive, with a curve that varies a lot: a top performer hits full speed in 3-6 months, an average profile in 8-12, a slower one in 14-18. The interesting part is that in the first stretch of that curve, the employee is the one gaining (absorbing knowledge); in the second, the company is the one gaining (getting output back). Shortening the ramp-up shifts that balance in your favor, and as I explain in the article on cutting onboarding time with AI, it also unlocks job rotation and lowers employee turnover.

The trouble with a DIY build is that you risk turning months of internal work into an archive the AI can't actually use well. You've spent the time of your most expensive people (the ones who know the company inside out) building foundations that then need to be redone. It's the classic case where the initial saving turns into the bigger cost.

The uncomfortable truth: understanding can't be outsourced

Here's the insight that reframes the whole build-versus-agency choice. With AI today, you can outsource two things: expertise (the AI writes code for you) and thinking (the AI proposes architectures, options, structures). But there's a third thing you can't hand off to anyone, not to the AI, not to an agency: understanding your business.

No outside partner knows why that client needs to be handled differently from the others, why that policy exists, which decision made two years ago is still binding. That understanding lives inside the company. Which means a good company brain is always born from a collaboration: the agency brings the method (structure, ontology, RAG, compliance), the company brings the meaning. Anyone who promises to build your company brain “turnkey” without asking for your time to understand the business is selling you an archive, not a brain.

And that's exactly why the choice isn't really “build in-house or agency.” It's “who designs the structure” and “who brings the understanding.” The structure is almost always better handed to someone who does it for a living. The understanding stays yours, necessarily.