Edge cases you never think to test: that's where AI output breaks. I deal with this daily in document processing.