Power Automate - AI Builder

Document processing AI Model used in power automate

Like (1) Share

Report

Posted on 11 Feb 2025 13:54:16 by KevinGador

560 Super User 2025 Season 1

Has anybody tried to use this AI model? How is your current model doing?

I've watched a couple of tutorials using this AI Model and secured proper licensing however it seems like my model is still very weak despite having 86 files. 66 english files and 20 in german files.

I don't understand why its still capturing things wrong. How do you train the model effectively? What is the literal minimum files and number of information does it take for this to be effective? What is the minimum amount of words for it to be effective looking for that specific information?

Do I need a separate model for each language?

Categories:

AI Builder

All responses (2)

Answers (0)

Building with Why 261 on 19 Feb 2025 at 01:39:14

Like (1)

Report

Document processing AI Model used in power automate

The number of documents is determined by the number of variations you have to the format of the document. How many fields are you attempting to model?

I have built a model for invoice processing. The invoices my company gets has many different formats. Logos, company names, basic invoice data all in a variety of locations. So you have to group them into sets that are similar to each other. I was looking at over 400 vendors, so doing a grouping per vendor would not work, but many of the vendors use the same format and are grouped together.

Also once you have built the framework of the groups, then the AI can figure out which group an invoice belongs in assuming it fits into a modeled group.
Suggested answer

takolota1 4,863 Super User 2025 Season 1 on 11 Feb 2025 at 15:06:41

Like (1)

Report

Document processing AI Model used in power automate

If you can’t get the trained models to work on AI Builder, you could try using OCR & GPT to extract the data like this template:

https://community.powerplatform.com/galleries/gallery-posts/?postid=31e67eea-3f73-47b4-95b7-fe4a7b646389

If your documents are well structured it should have 95% or more accuracy and at 4-7 credits per page it is less expensive than the trained models.