Invoices arrive from national retailers in every format there is, and at ScanApps my job was to make a model read them anyway. I tuned the OCR and extraction models in Azure Document Intelligence and Azure AI Foundry to classify each document and pull the structured fields out of it.
The whole difficulty is the input. A clean vendor PDF and a phone photo of a handwritten invoice sitting on someone's desk are the same task to the business and a completely different one to the model. Tuning across that range, from print to handwriting and from structured to photographed, is the work. Chains like Kroger, Albertsons, and Safeway each bring their own layouts on top of it.