Skip to main content

Beyond the Hype: What It Takes to Actually Deploy AI in Clinical Workflows

April 30, 2025
Image: [image credit]
Photo 147140056 © Denisismagilov | Dreamstime.com

Mark Hait
Mark Hait, Contributing Editor

It’s become a predictable cycle in healthcare IT: a high-profile partnership between a hospital and an AI vendor is announced, often accompanied by a flurry of LinkedIn posts, conference panels, and phrases like “revolutionizing care.” Six months later, the project quietly disappears—no outcomes reported, no clinician adoption, no operational integration. In the rare cases where AI does survive implementation, it’s typically relegated to a pilot status, siloed from real workflows and unsupported by the infrastructure required to keep it clinically meaningful. We don’t have a shortage of AI models. We have a failure to operationalize them.

The disconnect is structural. A 2022 review in The Lancet Digital Health found that while hundreds of AI models are published annually, less than 2% ever make it into production use, and fewer still are prospectively validated in live settings (Topol, 2022). Why? Because deployment requires more than technical sophistication—it requires alignment with real clinical constraints, governance from interdisciplinary teams, and a system-wide tolerance for iteration.

Even models that have cleared regulatory hurdles, like FDA-approved sepsis or stroke detection tools, struggle with utilization. Take the well-documented example of Epic’s sepsis prediction model. In a retrospective study at Michigan Medicine, it was found that the model missed two-thirds of sepsis cases and generated alerts on almost one in five hospitalized patients—contributing significantly to alert fatigue and being largely ignored by clinicians (Wong et al., JAMA Intern Med, 2021).

Worse yet, the AI field is plagued by a kind of deployment theater—a performative adoption of AI tools for reputational gain. Institutions eager to appear innovative push AI tools into workflows with little regard for downstream consequences. The result is what Stanford’s Dr. Nigam Shah calls the “model-to-workflow gap”—a scenario where models function in isolation, outputs are unread or ignored, and patient outcomes remain unchanged (Shah et al., NEJM AI, 2023).

The roadblocks aren’t just technical—they’re political and cultural. AI systems that alter clinical decision-making tread into deeply professional territory. They challenge physician autonomy, complicate liability, and raise questions about accountability. At UCSF, the AI governance team includes not just data scientists but also ethicists, frontline clinicians, and operational leaders—a recognition that AI must be managed like any other clinical intervention: with oversight, documentation, and continuous review (UCSF Health Hub, 2023).

Meanwhile, vendors continue to promise what they cannot deliver: turnkey AI integrations that “plug into your EHR” and magically surface the right insights at the right time. The reality is far messier. True integration requires not only FHIR or HL7-based interoperability, but also thoughtful decisions about where in the clinician workflow a prediction appears, how it’s explained, and what actions it triggers. This is why organizations like Mayo Clinic and Geisinger are investing in internal “AI ops” teams focused on deployment architecture, model retraining, and change management—because without that scaffolding, AI systems degrade, clinicians disengage, and projects die (Mayo Clinic Platform, 2024).

The most dangerous myth in healthcare AI right now is that technical success equals clinical success. It doesn’t. We need to stop measuring AI projects by model metrics and start measuring them by workflow adoption, clinician trust, and patient outcomes. Until health systems embrace this reality, the AI hype cycle will continue to produce more disappointment than disruption.

The lesson is simple but hard-earned: AI doesn’t transform healthcare unless it transforms care delivery. That means building systems, not just models. And until we start funding the boring, infrastructural work that turns AI from an idea into an asset, we’ll keep celebrating announcements instead of outcomes.