Knowledge accumulates in an organisation over time, and it has never been evenly accessible. Knowledge has always been power, and often an asset people hold onto. Efforts to democratise it have been underway for years, and they have mostly failed.
They failed for two main reasons. First, the old tools put the burden on you. You had to store the knowledge in a rigid structure, and know the exact path to get it back, and the information flowing through the organisation exists in neither form. It lives in a mix of slides, spreadsheets, half-finished documents, chats and emails and verbal discussions that were never documented. For many, SharePoint is the very last place where information is stored, due to its usability.
Second, and more fundamental reason is that those tools asked people to document as a separate task on top of their real work. And documentation is the part that always gets skipped. Manually curating the tacit or scattered knowledge into a rigid structured knowledge base just won’t work.
That is the antipattern the company brain has to avoid. The knowledge has to be captured as a byproduct of the work already happening (not as an extra chore nobody has time for). Let the model work over the mess as it is, help you curate and store the highest-value context as you go, rather than trying to structure everything up front. And that work is already happening: every day people assemble context, logic and reasoning inside general-purpose AI tools to get a good answer. And when the session closes, it’s left there in the chat history. The organisation never sees it. Catching that byproduct, with a human reviewing what’s worth keeping, is the opening.
There have always been people who know how things really work: why customers behave the way they do, which processes work in practice, and where the margins come from. The problem is that when these people leave, the knowledge leaves with them. Experienced professionals from the big generations are retiring in every industry, taking decades of context with them.
In the “Company brain” model, knowledge behind a conversational user interface stops living only in individual heads. A new employee can ask the same questions a seasoned veteran would ask and get useful answers immediately.
It also shifts power. Information that used to require knowing the right person or having the right access becomes available to everyone entitled to see it. For some, that’s a sore spot. Losing the leverage of being the gatekeeper is one reason it meets resistance. For many, information is power. It is an asset. The best and most up-to-date information is often stored in the minds of individual people, without any documentation.
It also raises the question of information security. “Available to everyone” can’t mean everyone sees everything. The point isn’t to remove access controls but to enforce them consistently. It’s the same access, made usable through a new interface. If you can’t open a record from a source system, your agent can’t either. And it arguably improves the governance picture: knowledge sitting in scattered files is genuinely ungoverned state – a managed context layer is where access control can be applied in one place.
Build the ownership
In principle, it’s quick to get started with all of this. The problem is that, over time, your business logic, your context, and everything the system has learned become embedded in the product. If the model changes – or the pricing, or the terms – your foundation moves out from under you.
So, build the layer that carries your meaning yourself. Your context layer describes your business concepts in your own language, records your decisions and reasoning behind them, and stores rules for how that information is interpreted. It binds to your existing semantic and data layers you already have rather than replacing them, so you don’t maintain manually what already is done in a system of record. The language model underneath is a replaceable component that you can swap anytime to another one or host your own local model. What you build on top of it is your company’s capital.
The learning loop is where the real value shows up
Every question, every output, every validated observation adds context – a feedback loop that improves what the next answer can rely on.
At the start, the system doesn’t understand your context; users are still learning what to ask, and there are gaps in the layer you are building. Over time, it gets good at the retrieval a senior does in their head (knowing what exists, where and why) and connects that information across silos that never talked to each other. The tacit judgment of what a thing means in this exact situation still belongs to your people; the system’s job is to put the right context in front of them, fast.
Microsoft’s Satya Nadella calls this “token capital”: AI capabilities built and owned by the company itself. One that grows alongside the human capital rather than replacing it. The point holds either way: capability grows over time only if you own it. If it remains locked in a vendor’s closed system, the growth benefits the vendor, and you end up paying for access to your own insights.
Are you looking for incremental or structural change?
Every company uses AI, to some degree. The move from shared context in a few workspaces and teams to something organisation-wide is already happening. The question is whether you do this deliberately or accidentally. In other words, do you own what you build, or are you renting intelligence from someone who can change the terms whenever they want?
The same owned-context approach that lets a manager get a straight answer also changes how the work itself gets built. When software and data teams stop re-deriving the same context on every task and draw it from a shared, owned layer instead, delivery starts to compound too. That’s a subject for a separate post.
Those who build their own learning ecosystem on an architecture they control gain an advantage that competitors can’t simply buy off the shelf.
Prefer to start with the why before the how? Read: Your organisation has the answers. So why the roughly right decisions?
