Current tech [for reasonable versions of current] can be approved although not all configurations (basically you have to fix the model for a release cycle). It

Current tech [for reasonable versions of current] can be approved although not all configurations (basically you have to fix the model for a release cycle). It isn't good enough yet for most uses in the space you describe. Improving that state of affairs is more a data limitation than a technology one, probably, but medicine is also very conservative and will take it slowly to prove out. Doctors care about getting paid, they mostly don't care about coding per se, and they don't do it themselves. It's not at all clear that a NLP AI would be more cost effective at it than the people who do it now. Where you will probably see it is auto-rejects, from the insurance side, forcing more iterations. None of this stuff is off the radar of (at least some of ) the people who build medtech systems, they've been incorporating aspects for a long time. The idea that there is a sudden radical change to be made with todays tech is naive I think, lacking context of both the history of AI systems in medicine and the current techs' limitations...