DEFINITION
What is a knowledge layer?
The layer beneath your AI that decides whether it works.
A model that knows nothing about your organisation gives answers that could be right anywhere except at your company. A knowledge layer is the place that records what is true at yours, readable by people and by agents.
A knowledge layer is a single shared source of truth beneath an organisation, recording what has been agreed, decided and learned, and readable and writable by both people and AI agents. W69 AI Consultancy in Amsterdam builds knowledge layers underneath AI platforms, so that every agent works from the same facts and every action stays traceable.
McKinsey, fieldwork November 2025: 88 per cent of organisations use AI, but in no single function had more than roughly ten per cent scaled it. PwC asked 4,500 chief executives in January 2026: 56 per cent said AI had yet to deliver revenue or savings.
Why your AI does not know what you know
The model is not the bottleneck. Models have improved several times over in the past two years and your results have not. What is missing is reliable context. The discipline that solves this has been called context engineering since July 2025, the term Gartner named as the successor to prompt engineering.
The knowledge sits in people’s heads
Why that client received a discount, why that process has an extra step, what went wrong with that supplier last year. None of it is written down. It lives with three people, and when one of them leaves it walks out of the door. An agent cannot reach it at all.
The knowledge is everywhere and nowhere
Spread across email, folders, chat threads, minutes and seven systems that do not know each other exist. Everyone searches and nobody finds. Put an AI tool on top of that and you get faster answers from the same mess. Being wrong more quickly is not progress.
The knowledge contradicts itself
Two versions of the same rate, two agreements about the same client, two dates for the same deadline. They live in different systems, so nobody puts them side by side and nobody notices. Until an agent acts on one of them on its own.
From scattered to an ordered library
Scattered knowledge goes in on one side. On the other side there is a single layer that people read from and agents both read from and write to.
Four things a knowledge layer must do
If your current solution does three of these, you have a good document system. Only with all four do you have a knowledge layer an agent can actually use.
People can read and search it
No export format and no black box. Someone must be able to ask a question and read the answer back, including where that answer came from and when it was established. Knowledge you cannot check is knowledge you will not trust.
Agents can read it and write to it
Reading is only half of it. An agent that works something out must be able to put the result back, so the next agent and the next colleague do not have to work it out again. Without write access the layer stays an archive instead of becoming a memory.
Contradictions surface
Two facts that cannot both be true should produce an alert, not silence. This is the part organisations underestimate and the part that pays off most in practice, because it brings to light the things everyone assumed were handled.
Something comes out of it every morning
A knowledge layer that only answers when asked will not be used a month from now. It should tell you what needs doing today, what has expired and what has been left undone. That is the difference between a reference work and a colleague.
What our knowledge layer found in its first days
We run this knowledge layer on our own company as well. This is what it brought to light there in its first days. Not an example from a brochure, but what actually came out.
Money left on the table
- A quote for additional work, ready to go and never sent
- A signed contract recorded in no file anywhere
- Four invoices that no longer existed as documents
- A referral fee nobody could remember
- Our own day rate, findable in exactly one quote
Risk nobody had spotted
- Fifteen files holding production keys, syncing to three devices
- A password sitting in plain text in a note
- A compliance claim on our own website that the website itself contradicted
- No second backup of any of the work
Quietly broken
- Google indexing that had been failing for weeks without a single alert
- An acquisition process that had not made an AI call in over a month
- A scan on our own site capturing zero leads, while being the main way in
- Nineteen references to a product name that no longer exists
Noise removed
- Four clients who had stopped being clients
- Two separate knowledge stores merged into one
- Eleven open files reduced to the three that actually matter
We build this underneath your organisation too. W69 AI Consultancy delivers the knowledge layer, the agents that run on it, the connections to your existing systems and the governance around them. One supplier, one point of contact, including after it goes live.
You do not have to take our word for it. We put this under our own company first, and then under TicketMatch, our own platform: live, with paying customers, five white-label storefronts and agents handling supply, bookings and delivery. On exactly this kind of shared knowledge layer.
And once it stands, we make sure it gets found as well, in Google and in AI answers alike. Designing, building, keeping it running and making it visible: it all comes from us.
Anyone selling AI architecture without running something that actually earns money is selling an opinion.
What people ask about this
An intranet is built to be read by people. A knowledge layer is built to be read and written by both people and agents. A document system holds a file. A knowledge layer holds a fact, with a date, a source and a record of where it came from. You only notice the difference once you put an agent on it: with a library of documents it can do very little, with a layer of recorded facts it can do everything.
No. RAG is how a model retrieves something and a vector database is where those pieces of text are stored. That is plumbing. A knowledge layer answers the question that comes before it: what is actually true here, who established that, and when. Without that answer, RAG simply retrieves the existing mess faster. Since July 2025 this discipline has been called context engineering, the term Gartner named as the successor to prompt engineering.
That is exactly the point. In most organisations contradictions stay invisible because the two versions live in different systems and nobody puts them side by side. A knowledge layer does put them side by side, so they surface. It feels uncomfortable in the first few weeks and it is the single most valuable thing it does. What you can see, you can fix.
The first working version takes days, not months. The clean-up afterwards is the real work, because that is when it becomes clear what is duplicated, out of date or contradictory. Budget a few days to build it and a few weeks to trust it. Start with the subject where the most mistakes are made, not the subject with the most documents.
You do. A knowledge layer belongs in your own environment, in readable files rather than a vendor’s closed format. You must be able to take it elsewhere tomorrow. For organisations with data sovereignty requirements this is not a luxury but a condition: by definition the knowledge layer contains the most sensitive thing you have, which is all of it in one place.
Want to know what sits beneath your AI?
The AI Navigator™ maps where your organisation stands, including the point where it usually comes unstuck: the knowledge your models are supposed to run on.