AI models change quickly in price, quality and terms. A model-agnostic design keeps prompts, business logic, data access and tests separate from any one provider's API, so a workload can move to another provider, or to a model hosted in your own cloud account, when there is a good reason.
For example, a document workflow might use a larger model for complex contracts and a cheaper one for simple forms, with both running through the same interface and test set. The limit is that switching is never entirely free. Different models behave differently, so every change still needs to be checked with evals before it goes live.