Exactly. Testing an API in isolation doesn't help you a thing when the bug arises because of, say, subtle inconsistencies between API implementations of interacting services on a long chain.

This sounds a lot like the code coverage fallacy. (to which I usually answer "call me when you have 100% data coverage").