An eval pipeline raises an alert when quality starts to slip. That can happen when the model is updated, when the instructions to the model (the prompts) change or when the data shifts over time. In other words, it's to AI what automated tests (unit and integration tests) are to ordinary code, but it's built differently.
Our standpoint
Everything we build has an eval pipeline from the first pilot. No exceptions. Without it, neither you nor we would know if the solution still answers correctly after the next model update.