Inside a Working Intelligence System: How to Build a Company Knowledge System

Most writing about company knowledge systems is theoretical. Someone describes an architecture they have read about, draws a diagram with four boxes and some arrows, and stops before the part where you find out whether it actually works. I want to do the opposite here, so this post is a tour of ours: the real one, the one this business runs on, with the real numbers attached.

As I write this, our intelligence system holds 666 documents across the business vault and the project memory pools that sit alongside it. A question put to it comes back with a cited answer in about 50 milliseconds. It was not a project. It accumulated, and it is still accumulating, and the whole point of showing you the inside of it is that I am not describing something I read about.

If you have not come across the term, an intelligence system is an owned, maintained store of what your business knows, structured so people and AI can ask it questions and get current, cited answers. This post is the level below that definition: what one actually contains, how the answers stay right, and who keeps it alive.

What does an intelligence system actually contain?

It contains four layers: dated raw sources, the entities your business revolves around, a synthesis layer that turns those into one current answer per question, and a retrieval layer that returns answers with their sources attached. Most failed attempts have one or two of these and call it a knowledge base.

Take them in turn, because the interesting part is what each layer is deliberately not.

Raw sources are the immutable record: call notes, emails, signed documents, meeting transcripts, each with the date it happened. Ours are filed by date and never edited after the fact. That last part matters more than it sounds. A source you can quietly revise is not evidence, and the moment you start tidying history you lose the ability to answer "what did we actually agree in March?"

Entities are the things the business revolves around: customers, people, projects, decisions. Ours currently holds 16 customers, 32 people, 15 recorded decisions. These are the nouns your business thinks in, and giving each one a page is what turns a pile of documents into something you can navigate. When a customer page links to the meetings, the decisions, and the case study, you stop having to remember which folder anything lives in.

Syntheses are the answer layer, and this is the one nearly everybody misses. A synthesis is one current page for one recurring question, whether that is what our positioning actually is, what we learned from a particular engagement, or how our tone of voice is meant to sound. Without this layer, asking a question means reading eleven documents and forming your own view, which is exactly the senior-hours cost that corporate amnesia charges you over and over.

Retrieval is the ability to ask in plain language and get an answer with its source attached, fast enough that you actually bother. Ours answers in about 50 milliseconds, which sounds like a vanity metric until you notice that speed is what decides whether people use it or go back to asking a colleague.

The anatomy, in real numbers

Here is the actual shape of ours today, because rounded claims are easy to make and easy to inflate.

The business vault holds 261 documents. Alongside it sit the project memory pools, another 405, giving 666 in total that a single query searches. Of the vault's 261, the largest group is raw sources at 77, then syntheses at 28 and meetings at 23. Customers, people, projects, decisions, processes, and case studies make up most of the rest.

That ratio is worth dwelling on. There are nearly three raw sources for every synthesis, which is roughly the balance you want. Too few sources and your syntheses are assertions. Too few syntheses and nobody can get an answer without doing the reading themselves.

Every document carries the same small set of fields, and these do most of the work:

• created and updated, so staleness is visible rather than assumed
• sources, linking a synthesis back to the raw records it was built from
• confidence, so a well-evidenced page and a working assumption do not look identical
• confidentiality, so the system knows what can leave the building

Those five fields are the entire governance layer. There is no workflow tool, no approval chain, no separate platform. A document that cannot say when it was last touched and what it was based on is not part of the memory, it is just a file.

How an intelligence system compares to what you probably have

Dimension	Shared drive of documents	Company wiki	Intelligence system
What it holds	Whatever was saved, in whatever state	Pages someone wrote once	Dated sources plus a synthesis per question
Finding an answer	Search by filename and hope	Browse the tree someone designed in 2023	Ask in plain language, get a cited answer
Staying current	It does not	Decays quietly after the launch push	Dates and sources make decay visible
Safe for AI to use	No, it cannot tell current from obsolete	Partly, if anyone kept it honest	Yes, because every answer is traceable
When someone leaves	Their context leaves with them	Their undocumented reasoning leaves	The reasoning was the thing you recorded

The part that keeps it alive

The honest answer to "what stops it going stale" is nothing, automatically. Every abandoned wiki in the world was built by people who meant well. What you can do is make decay visible and cheap to fix.

Two habits do almost all of that work. Sources land as the work happens, not in a documentation quarter, so the raw layer grows by itself as a by-product of doing business. And syntheses carry a visible updated date, so a page that has not been touched in eight months announces itself rather than quietly misleading someone.

The maintenance cost is smaller than people expect, because the expensive part of documentation was never the writing. It was the deciding, and you were doing that anyway.

Why this has to come before the agents

There is a widespread assumption that AI agents will solve the knowledge problem. The evidence points the other way: agents make a fragmented memory more expensive, not less.

Salesforce's connectivity research found that 90% of IT leaders say data silos are creating business challenges in their organization, and that the challenge is greater at organizations using agents (94%) compared to those not using agents (87%), with organisations running an average of 897 applications. Adding agents to a business that cannot answer its own questions consistently just industrialises the inconsistency.

The value, meanwhile, is understood to sit in exactly the material an intelligence system organises. In IBM's 2025 study of 2,000 CEOs, 72% view their organization's proprietary data as key to unlocking the value of generative AI. Proprietary data is not a data lake. It is what you know about your customers, your decisions, and your own voice, which is the thing most businesses have never written down in one place.

And the barrier is not enthusiasm. DSIT's AI adoption research found the top barriers were a lack of identified need for AI (71%) and limited AI skills, expertise and knowledge (60%). A memory is a good answer to both, because it makes the need concrete and it lowers the skill required to get a trustworthy answer out of the business.

What it makes possible

The reason to build one is not tidiness. It is that a maintained memory changes what the automation on top of it can safely do.

Every post in our current content plan, this one included, is drafted against that memory rather than from a blank page, which is why the tone holds across months of publishing. When we built the content engine for Excellerate Services, the same principle applied on the client side: weekly production went from roughly 12 hours to about 2 hours of oversight, and it held its brand voice across regions because the voice was written down once rather than carried in several heads. An angel investment community got 4 hours a week back on the same basis.

None of those are automation wins on their own. They are what automation can do once it has something reliable to stand on. That is the whole argument of the maturity ladder: the memory is the rung most businesses skip.

Practical takeaways

1. Start with the sources you already generate. Call notes, signed documents, decisions. Date them and stop editing them. The raw layer costs almost nothing because the work is already happening.
2. Add a synthesis per recurring question, not a page per topic. If you find yourself answering something for the third time this quarter, that is the next synthesis. Ten good ones beat a hundred stubs.
3. Put updated and sources on everything. Those two fields turn a folder into a memory, because they let a reader judge whether to trust the page without asking anyone.
4. Give it one owner and a small rhythm. Not a librarian, not a project. A named person and a weekly habit. Nobody's job is how wikis die.
5. Build it before the agents, not after. An agent on top of a fragmented memory produces confident, inconsistent answers at scale, which is worse than no agent at all.

If you want the wider capability picture around this, the roadmap for building your own agentic intelligence covers where the memory sits in the build order, and if you would rather do it with a partner, that is the work we do.

Frequently asked questions

What does an intelligence system actually contain?

Four layers. Raw sources, which are dated records of what was actually said or signed: call notes, emails, signed documents. Entities, which are the things your business revolves around: customers, people, projects, decisions. Syntheses, which are the answer layer, where a question like what is our positioning gets one current page rather than eleven conflicting ones. And retrieval, the means of asking it a question and getting a cited answer back in seconds. Miss the synthesis layer and you have a filing cabinet; miss the sources and you have unverifiable opinion.

How do answers stay current in a company knowledge system?

By separating what was said from what is true now. Raw sources are immutable and dated, so a call note from March stays a March call note forever. Syntheses sit on top and carry their own updated date plus links to the sources behind them. When something changes you revise one synthesis rather than hunting through every document that mentioned it. That separation is the difference between a memory that compounds and a folder that rots.

What stops an intelligence system going stale?

A rhythm and an owner, not enthusiasm. Ours works because new sources land as the work happens rather than in a quarterly documentation push, and because every synthesis carries a visible updated date, so staleness is obvious rather than hidden. The honest answer is that nothing stops it going stale automatically. What you can do is make decay visible and cheap to fix, which is what dates, sources, and a small weekly habit achieve.

Who maintains a company knowledge system?

One named person owns the rhythm, and everyone who does the work contributes the sources. It does not need a librarian or a new hire. In our case the maintenance is minutes a day, because capture happens where the work already happens and the synthesis work is the part a person actually thinks about. The failure mode to avoid is making it nobody's job, which is how every abandoned company wiki started.

What does cited mean in practice for AI answers?

It means every answer comes back with the specific document it came from, so you can open the source and check it in seconds. Not a confidence score, not a plausible summary, an actual traceable page with a date on it. This is what lets a new starter or an AI agent act on an answer without a senior person re-verifying it first. An answer you cannot trace is a rumour with good grammar.