AI Vocabulary
Model, layer, application: which of these three storeys is yours?
The foundation model is the generic engine, trained by a laboratory on general corpora. The applicative layer is what frames it for a given profession. The application is what you handle. None of the three belongs to you, but what you deposit in them does — and that depends entirely on which storey holds it.
The three storeys
On the first, the foundation model. A handful of laboratories worldwide produce them, at a cost counted in hundreds of millions. It knows neither your profession nor your matters, and it is the same for everyone who uses it.
On the second, the layer that frames it: what decides what to give it, how to divide a task, what to check before returning a result, and what to keep afterwards. This is the storey where professional specialisation lives.
On the third, the application: the interface, the integration with your tools, access management. This is what your teams see and judge.
Why the distinction is poorly known
Because market communication blurs it. An application provider will readily say it uses the best model, which is true and says nothing about what it does on top. A laboratory will say its model understands law, which is true in the sense that it has read a great deal of it.
The practical consequence is that products are compared on the name of their engine, when that is often the only thing they have in common.
How to tell which storey you are talking about
Three questions place an interlocutor, and they require no technical knowledge.
Who trained the engine you use? If the answer is a third-party laboratory, your interlocutor works on the second or third storey, which is the normal situation and in no way troubling.
What do you do between my request and the call to the engine? A detailed answer describes a real layer; a vague one indicates there is little between the interface and the model.
And what happens if that engine disappears tomorrow? A provider that has built a layer answers that it plugs in another. One that has not answers that it will not happen.
What it does not solve
Knowing the architecture does not tell you where your data is held. An application may keep none, or keep everything. A model may be hosted by its publisher or in dedicated infrastructure. These questions arise storey by storey, and the answers differ.
The distinction says nothing about quality either. An excellent model badly framed gives poor results on a complete task; a middling model well framed gives usable ones. The intermediate storey decides, and it is the least visible of the three.
Why it matters to a lawyer
Because reversibility arises differently at each storey. What is attached to the model disappears when it changes, which happens every 6–12 months. What is attached to the layer survives that change, if the layer returns it to you. What is attached to the application disappears with it.
The question to ask is therefore: at which storey are our decisions, our terminology, the history of our matters held? If the answer is the first, you will lose everything at the next change of engine.