I’m seeing some applications for it as an easy-to-implement context-aware OCR scanner for things that have consistent content in wildly inconsistent formats. Think lists of open invoices from suppliers in a thousand different for-printing layouts, plus some just being screenahots from their systems. The AI system can process those and output a consistent list, which can then be matched against a database the old-fashioned way.

That’s the only practical use I’ve found.