Frameworks / The Intelligence Control Stack
The Intelligence Control Stack
What an organisation actually controls when it deploys intelligence it did not build. Working through the layers separates the parts that are genuinely yours from the parts you are renting under someone else’s terms.
Ask most enterprises whether they “have” AI and the answer is yes. Ask which parts of it they actually control, and the confidence drops. The intelligence in a modern product is a stack of layers, and ownership is different at every layer. Some you own outright. Some you rent on terms that can change without your consent. The danger is not renting — renting is often correct — it is renting without knowing you are doing it.
The layers, top to bottom
From the surface down: the experience your users touch, the application logic that orchestrates the intelligence, the model that produces it, the data it was trained and grounded on, and the infrastructure it runs on. Most organisations own the top confidently, and get progressively vaguer the further down you go — until, at the model and infrastructure layers, they are almost entirely dependent on someone else and often have not said so out loud.
The gradient is the point. Control is strong and teal at the top, where you built it, and weakens toward amber as you descend into layers someone else provides. A healthy organisation can draw this diagram for its own estate and defend the colour of every layer. An unhealthy one assumes the whole stack is teal because the top of it is.
How to see it in the field
The tell is an organisation that cannot quickly say what happens to its product if a single model provider changes terms, raises prices, or withdraws a model. If that question produces a long silence, the model layer is rented and the rent is unexamined. The dependency was always there; only the awareness was missing.
A second tell: strategy documents that describe the AI capability entirely in terms of the top two layers — the experience and the clever application logic — with the model, data, and infrastructure treated as an undifferentiated “platform” someone else worries about. That silence about the lower layers is exactly where the control questions live.
Getting out of it
Draw the stack for each significant AI product and colour every layer honestly — owned, or rented, and if rented, on whose terms and with what exit. The goal is not to own everything; that is neither possible nor wise. The goal is for every rented layer to be a decision you made on purpose, with a known cost of switching, rather than a dependency you backed into and would discover only in a crisis.
Where a rented layer is both critical and hard to exit, that is where to invest in optionality — a second provider, an abstraction layer, a fallback. Sovereignty at the stack level is not about owning the most. It is about never being surprised by what you rent.
Developed in the book Who Controls Your Intelligence? Related: the Workload Sovereignty Matrix.