For this task, general models will always perform poorly. My company trains custom gen ai models for document understanding. We recently trained a VLM for the German government to recognize documents written in old German handwriting, and it performed with exceptionally high accuracy.