We expected Form Recognizer would reward the overhead of training and managing custom models with superior performance, but we were disappointed. In our testing, Form Recognizer's accuracy was poor and its response times were way too slow for synchronous use cases.
Training on just 26 documents seems woefully inadequate. I'm not a data scientist and have only cursory exposure to ML, but I'm not surprised to see terrible results with such a small training set.
I'm skeptical it's possible to achieve good performance with an unsupervised model with only 500 pages, unless those documents are very similar. In which case, why would you need a service like Form Recognizer at all?
From a product perspective, it just makes no sense to me.