Finish the Pilot, Don't Multiply It
Picture the demo: an agentic AI system that triages referrals, drafts notes, flags a deteriorating patient, surfaces a care recommendation — all in one flawless pass. The committee approves it. Six months later, the system is technically live. It no longer touches a single real clinical decision.
That failure mode has a name now, and it isn't rare. Mykhailo Hentosh, writing in MedCity News, cites the widely circulated MIT finding that 95% of enterprise AI pilots never deliver measurable ROI — and names the four predictable ways hospitals get there: the all-department Big Bang that runs out of budget before a user sees value; Pilot Purgatory, where a system that dazzles in a demo can't survive real workflows; Technology-First builds that go looking for a problem after the money's already spent; and Governance Paralysis, where the oversight model gets perfected while momentum quietly dies.
Four different failures. One shared cause: systems engineered to look intelligent instead of being trustworthy inside an actual clinical workflow.
The sentence that matters most
Buried in the operational detail is the line that should anchor every hospital AI strategy: an agent surfaces the recommendation. The specialist makes the decision. Not as a compliance checkbox — as the actual design.
That's not a limitation on the technology. That's the technology working correctly. An AI system that quietly makes the call and routes around the clinician isn't more advanced — it's less accountable, less legible, and, per Hentosh's data, less likely to survive contact with a real floor nurse's Tuesday.
He offers a test for this: if you can't explain to a floor nurse, in one sentence, what the agent does — you don't have one project. You have several, wearing a trench coat.
Discipline as the sovereignty mechanism
The fix Hentosh proposes isn't a better model. It's sequence: a four-week assessment, a twelve-week pilot scoped to one clearly defined task, a change-management phase that actually measures adoption instead of assuming it, and only then scale — with a real decision gate, and the willingness to stop, at every stage.
Read past the project-management language and this is a sovereignty argument, whether or not Hentosh would frame it that way. Pilot sprawl doesn't just waste budget — it diffuses authority until no one can say who's accountable for what the system is doing. A narrow pilot with a one-sentence explanation and a human making the final call keeps that authority exactly where it belongs: visible, traceable, and held by the clinician in the room.
The hospitals that pull ahead over the next few years, per Hentosh, won't be the ones running the most ambitious pilots. They'll be the ones running the fewest — and actually finishing them.
That's not a modest goal. In a field this crowded with impressive demos, finishing might be the most sovereign thing a hospital can do.
Source: Mykhailo Hentosh, "Why Agentic AI Works in Hospitals When You Strip It Down," MedCity News, October 14, 2026.