Then came the moment to sell it to someone who would use it every day, and the gap between the two became the problem.

A prototype like this works because whoever demos it knows which questions to ask and which documents to load. In production, badly formatted documents arrive, questions come in unexpected shapes, and volumes are unknown in advance. The document-processing pipeline has to become reliable and traceable: when an answer is wrong, you need to be able to trace back what produced it.