A demo on a clean invoice proves nothing. A document extraction pipeline meets scans of scans, merged cells, drawn forms and a corpus nobody counted.
Every document extraction demo works. Someone drops a clean invoice into a web form, a model returns supplier, date, total and VAT number, the room nods, and a budget gets approved. Six months later the same pipeline is running at 71% straight through on the real archive, an analyst is opening 3 documents out of 10 by hand, and nobody can explain the gap because nobody measured the thing that caused it.
The gap is not the model. The gap is the corpus. A demo runs on a document somebody chose because it was readable. Production runs on whatever twenty years of scanners, mail rooms, ERP exports, legal departments and Dropbox folders happen to have left behind, and that material is hostile in ways that have nothing to do with intelligence and everything to do with the file format.