Why do hospital AI pilots fail to scale?
·Updated ·3 source episodes
Most hospital AI pilots fail because the work never lands in the system of record, the success metric measures activity instead of resolution, and no executive owns the number the tool moves. Pilots that scale make staff days quieter, write back into the EHR, and report outcomes the buyer already tracks.
What happens when the work does not land in the EHR?
The most common failure is invisible: the model performs, and the staff still does the task. If the output does not write back into Epic, Athena, or whatever the system of record is, no time was saved.
“The important part is the write back into the EHR. If the work does not land in Epic or Athena, the staff still has to do it, and you have not saved anyone any time.”
Which metric tells you a pilot is real?
Deflection, containment, and volume handled are vanity metrics in patient facing AI. The honest measure is whether the person on the other end got what they came for, and whether the handoffs that should happen do.
“We report resolution rate, not deflection. Deflection just means a human did not touch the call. Resolution means the patient got what they called for.”
“Operators trust the system faster when you are honest about the calls you should not handle.”
Is the pilot a company or a consulting engagement?
Some pilots succeed clinically and still cannot scale, because the build was specific to one system's configuration. Investors who build companies from scratch screen for this explicitly.
“Can the same product serve a second and third system without a rewrite. If a problem fails the third filter, it is a consulting engagement, not a company.”
“There is too much money in thin wrappers around a model that any incumbent can ship as a feature.”
How do you keep clinical quality from being an afterthought?
When the clinical function is a review layer bolted on after the product, quality erodes quietly under commercial pressure. Guests who run clinical organizations build it into the product team instead.
“We built the clinical function as a first class part of the product rather than a review layer bolted on afterward. Our care navigation, our provider criteria, and our education content all come out of the same clinical team.”
Episodes this article draws on
More from the archive
This article is synthesized from Care Shift interviews and is updated as new episodes add material. Browse all episodes or read more about the show.