What It Actually Means
In conversation the word covers several distinct things, and the ambiguity causes real confusion in commercial discussions.
The Three Things People Mean
A general-purpose model from a provider. GPT, Claude, Gemini and similar. You rent access, you do not own it, and it can change or be retired on the provider’s schedule. Most AI proposals are built on one of these.
A model trained on your data. A classifier or predictor built specifically for your business, usually much smaller and narrower. You own it, it runs on your infrastructure, it costs nothing per request, and it does exactly one thing.
A fine-tuned version of somebody else’s model. A middle case that inherits the provider’s dependency while carrying your customisation.
Those three have completely different answers to who owns it, what it costs, what happens when it is wrong, and what happens when the provider changes something. When a supplier says “the model”, establishing which of the three is meant is a reasonable first question.
What Follows From The Distinction
Ownership and continuity. A model you trained continues to work indefinitely. Access to somebody else’s is a commercial relationship, and providers do retire versions. Anything depending on precise behaviour needs testing when that happens, and needs a plan for when it is withdrawn.
Cost shape. Your own model has a build cost and negligible running cost. A rented one has little build cost and a permanent per-request cost, which is inference, and it scales with success.
Explainability. Smaller purpose-built models can often show which factors drove a result. Large general models cannot in any practical sense. Where a decision has to be justified to a customer, a regulator or a court, that difference decides the architecture.
Data. Training your own means your data stays yours. Using a provider means it leaves your systems, which needs checking against what you have promised your own clients.
What To Ask
- Which of the three is this? Everything else follows.
- What happens when the provider changes or retires it? There should be an answer that is not a shrug.
- Could a smaller purpose-built model do this job? For classification and prediction it frequently can, more cheaply and more explainably.
- Where does our data go? Different by a wide margin across the three.
Why It Is Worth Being Precise
Because “we built a model for you” and “we wrote a prompt against a rented model” are both said the same way, and they are very different products at very different prices.
Neither is wrong. Renting a general model is the right answer for most language tasks, and there is nothing to apologise for in it. The problem is only when the pricing implies bespoke engineering and the substance is configuration, which is a question you can settle by asking what happens if the provider disappears tomorrow.
More terms are in the glossary.
