Buyers ask which model a product is built on. It is close to the least informative question available.
Because the models are the visible, branded, impressive component, and because vendors talk about them constantly.
Capable models are broadly available and the gap between the leaders narrows and reopens continually. A product built on one can usually be moved to another.
Anything that can be swapped in a sprint is not a durable difference between two products.
The data the product has access to and the rights to use. The workflow it sits inside. What it does when the model is wrong. And whether the output can be explained.
Those are all product decisions rather than model decisions, and they are what determine whether it works in a regulated firm.
Ask what happens when it is wrong. That question separates products more reliably than any capability question.
What specifically does this system do that a well-written rules-based system could not.
Sometimes the honest answer is "not much, but it does it without a year of rule-writing." That is a legitimate answer and it tells you what you are buying.
Cost, latency, and whether it can run in an environment that satisfies your data requirements. Those are real constraints and they are operational rather than about capability.
Because the underlying capability is broadly available, you are not choosing between access and no access. You are choosing between products.
That means the ordinary questions — support, data rights, exit terms, what happens when it is wrong — matter more here than in most technology decisions, not less.
Read next
What AI-native should mean →