It depends on whether your API provider is HIPAA compliant and what sort of disclosures are provided. If you're running inference in house then it's not an issue provided you're following all the other guidelines.

With FormKit you'd still be using a predefined deterministic flow where all the conditionals are accounted for in advance. FormKit wouldn't be able to generate a truly non-deterministic UI that changes based upon how the user answers. If you're wanting to sue your neighbor for their tree falling onto your house, an AI powered UI for a law firm could be giving you form fields on what type of tree it was, how old the tree was, hot tall the tree was, how close it was to the property line, and what room in your house the tree landed in... All generated from how the user responded to an initial "what's your grievance?" question.

I'm not advocating for or against this type of approach - I'm just pointing out a few potential use-cases for it since that's what you asked for.