What actually goes into a context layer
I get asked one question more than any other: what actually goes into a context layer? Fair question. The term is suddenly everywhere, and most definitions stop at the diagram. Here is the plain answer, from building them.
A context layer holds four kinds of knowledge.
Facts
The things your systems of record already know: accounts, contacts, products, pricing, pipeline, tickets. This is the easy part, and it is where most projects stop. Facts alone get you a chatbot that can look things up. That was never the bottleneck.
Language
How your company talks. The approved claims and the banned ones. The way you describe what you sell, the words your customers actually use on calls, the difference between what marketing writes and what wins deals. Feed a model your language and the output stops sounding like a press release written by a stranger, because it is no longer being written by one.
Rules
The judgment your business runs on, written down. Pricing floors and who can approve an exception. What legal has cleared. Which customers get the white-glove treatment and why. Most of this lives in people's heads and surfaces only when somebody breaks it. Encoding it is tedious, and it is the difference between an AI you can trust with real work and one you have to babysit.
History
What was promised, what was decided, and why. The deal that fell apart and the reason. The campaign that worked and the one that did not. This is the layer Gartner is pointing at with the term context debt: companies remember what happened and lose track of why. History is what lets the system learn, because every action it takes feeds the next one.
What holds it together
Two properties make these four things a context layer instead of a data lake. It is governed: somebody owns each piece, it stays current, and there is one version of the truth. And it is connected: it supplies all of this to whatever model you use, at the moment the work happens, through the tools your team already lives in. A data lake stores everything and explains nothing. A context layer exists to be drawn on.
You do not build all four layers for the whole company on day one. You build them for one team and one workflow, prove the payoff, and expand. The facts take days. The language and rules take weeks. The history compounds forever, which is exactly why it should start now.