I think at a certain scale we're talking about switching to local trained models which don't have the same operating costs as running a frontier model for OCR. That would reduce the ongoing costs significantly. Might take longer than 30 seconds to read each receipt if you run multiple passes to ensure accuracy, but could run 24/7/365 without the same tax and administration overhead of humans.
Spherical cows aside though, I do agree with you that I should not consider scalability as a given.
