[jmmcd]

I disagree. If we consider a single customer with a particular application, we probably imagine an iterative process like this: the customer supplies some sample images, the system gets a few wrong, they add special cases, fine-tune the hyperparameters, whatever, and after all this they literally get 100% on the customer's holdout data. If that happens for several customers then the GP statement is justified (no other number is possible for these customers).

Possibly you're thinking of a single error rate across all customers? For other customers, as stated, it's not 100%. But taking an average across multiple customers is not meaningful when some are counting pearls and some are counting crops.