This is probably a question more focused at the development team and developers using the solution to backend their systems.
I have been playing around with the Form Recognizer and building a test front end that uses the extracted data for later processing. On completion of the AI Process, the idea is that a user will compare and correct the receipt data recognised against the image of the receipt. (Mismatches/Incorrectly recognised data seems to be higher based upon quality of the initial image scan.)
The process of using the AI to do the hard work and following with a user check should produced quite accurate results and might well be akin to a double entry approach. Any corrections that are made by the user are stored in the database as well as the initial scanned data. I am sure that this approach will ultimately be deployed by many organisations in the future given we are still stuck on receipts for now.
One of the benefits here might be that there is an opportunity to further train the pre-built models using this data.
Is there a feedback loop available or planned where this corrected data can be pushed back for further AI model training to built greater accuracy for all users?