If you had an internal API as opposed to using an external one you'd have some sort of A/B or multi-armed testing in place to check if the model you've deployed

If you had an internal API as opposed to using an external one you'd have some sort of A/B or multi-armed testing in place to check if the model you've deployed is performing as expected. At this point you don't think about the theoretical best you could perform at for the said model without computing some sort of complexity measure. I assumed one would have to use a similar automated process for external APIs. I don't see why they have to be different per se.