Respect for the work, but recommend a pivot to Epic IT integrators as your target customer...

(1) Don't confuse medical errors with claims errors. Your claims-amplifying customers don't really care about medical errors; they're mainly just optimizing their extraction from government and insurance payment systems. (And the vast majority of medical errors take significant skill to detect - beyond even complicated decision support systems.)

For claims errors, I would rather the system provided feedback to the EHR Epic engineers than trying to block providers. The Epic IT should be getting regular reports that prompt them to fix their UI issues.

But then I care more about fixing the Epic UI than claims.

(2) Epic/EHR's are an epic UI failure (not surprising since it was not driven by user need but via top-down requirements). It has random and super-complicated form UI's forcing users into complicated multi-step workflows to say something trivial.

Today in medicine the logistics of interfacing with Epic and other EHR's takes longer than the actual care. (Just imagine having to use a compiler that took longer than you did to write the code.) It's the scourge of medical care today.

In that context, imagine: you want to build a system that argues with providers when they're done, based on AI logic completely separate from the Epic system logic? It's hard to imagine a better way to make a bad situation worse.

What would be huge benefit instead is an AI tester for Epic. Something where you could generate all the ways users might see their UI and need to use it, and quantify all the confusing+unnecessary visuals and steps, to actually measure usability. Think user modeling and fuzzing coupled with progressive pruning for workflows, with actual metrics of system and workflow complexity.

That usability testing would probably be useful in other domains, too. But starting with Epic would be good because it has so many UI errors (high signal to noise) and saving time for highly-paid, highly-blocked users translates directly to dollars. You could sell it to every Epic integrator in the US. Those customers are easy find and target, they have strong needs, they can work with you as your technology evolves (and filter the inaccuracies). By giving them objective measures of usability/complexity, you simplify their design space and give them a clean way to measure system improvement, reducing the level of politics they endure.

Along the way you would build AI models over user interaction (instead of tokens). Then you could build interactive auto-completing UI's that work based on session observation and voice alone.

Unless I'm searching or researching, I don't want AI to replicate what's been said. I want AI to anticipate what I'm doing, and afford me the choices I need to make. That's exactly the model of diagnosis and treatment.

Most of that seems rather pointless. Epic is the worst EHR — except for all the others. There aren't a lot of Epic IT integrators out there as potential customers: Epic does most customer implementation and support in house. They do offer an extensive set of APIs for third-party developers including SMART on FHIR but there aren't many IT integrators focused on that either.

As for usability feedback to Epic engineers, they generally aren't interested in what you have to say. I mean Epic does make product changes based on customer feedback but they don't listen to input from random other companies and they certainly don't pay for it. Their culture is more that they know the correct way to do things and customers should change their processes to fit the software.

https://www.acquired.fm/episodes/epic-systems-mychart

Very insightful, thank you! A comment on claim-amplifying customers - optimizing for the insurance payouts actually forces providers to adopt certain standards of compliance for medical records - here we are not talking about mistakes but about depth of CPT coding (the biggest for value-based care providers), etc. And it is not about claim but the underlying documentation of care