The AI press release from a major hospital system in 2023 described “comprehensive clinical intelligence that will transform the patient journey from intake to discharge.” The system spent $22 million over two years. What is running in production today is a documentation transcription tool used by 400 physicians who dictate clinical notes, and a scheduling optimization system that reduces appointment no-shows by 11%.
The executives involved will tell you privately that this is not a failure. They will say — and they are right — that the transcription tool saves each physician approximately 90 minutes per day that was previously spent on documentation, that this has measurably reduced physician burnout (their satisfaction scores went up for the first time in four years), and that the scheduling optimization generates real revenue recovery. These are genuine wins.
What the press release described and what actually shipped are connected only by the word “AI.” The announcement promised transformation; the deployment delivered optimization. Optimization is less interesting to write about. It is more valuable to patients.
The Narrowing Principle
Healthcare AI has converged, through trial and expensive error, on what might be called the narrowing principle: the probability of a successful deployment is inversely proportional to the scope of the initial ambition. The AI projects that have actually reached production and delivered measurable clinical value are almost uniformly ones that started with a very specific problem, for a very specific patient population, in a very specific clinical context.
Sepsis prediction algorithms — systems that flag patients at elevated risk based on vital signs and lab values — have been deployed successfully at several large health systems, including Epic’s integration-based approach and standalone products from companies like Dascena and Somatix. What makes these deployable is their narrow scope: predict one outcome, for one patient population, using one set of inputs, with one alert output reviewed by a human clinician. The AI is not trying to synthesize a patient’s entire clinical picture. It is doing one thing repeatedly and well.
Radiology AI is the other success story, and for the same reason. Products from Viz.ai (large vessel occlusion detection in stroke), Aidoc (incidental finding prioritization), and Nuance’s AI-assisted reporting (structured findings from radiology reports) are running in genuine production environments at hundreds of hospitals. They succeed because they address specific, bounded tasks: detect this finding, prioritize this study, extract these data elements. The scope limitation is not a product design choice imposed by conservative product managers — it is what regulators required and what clinical validation demanded.
The FDA’s 510(k) and De Novo pathways for AI-based medical devices have, paradoxically, been beneficial for deployment quality because they forced vendors to define specific intended uses, specific patient populations, and specific performance standards for each AI function. A product cleared for “detection of large vessel occlusion on CT angiography” cannot be quietly expanded to “general neurological AI.” The regulatory scope limitation mirrors the clinical scope limitation that makes these systems trustworthy.
The Administrative AI Window
While clinical AI has progressed carefully and narrowly, the most interesting current deployment frontier is the administrative and operational layer — an area with significant potential, less regulatory complexity, and a track record that is more mixed but genuinely encouraging in specific cases.
Clinical documentation has become the clearest administrative AI success story. Ambient documentation tools — systems that listen to physician-patient conversations and generate structured clinical notes — have been adopted by a meaningful number of practices and health systems, with Microsoft’s acquisition of Nuance (and the DAX Copilot product line) being the most prominent example. The clinical and commercial rationale is compelling: physician documentation burden is a primary driver of burnout, and the AI reduces it substantially without changing the physician-patient interaction.
The quality limitations are real and worth naming honestly. Ambient documentation AI produces notes that are accurate for the documented conversation but miss clinical context that the physician holds in their head but does not verbalize — prior visit context, implied reasoning, the significance of what was not said. A physician who relies on the AI note without reviewing and editing it will sometimes produce a clinical record that is technically accurate but clinically incomplete. The systems that work best are the ones that treat the AI output as a high-quality draft, not a finished note.
Prior authorization processing is a potentially large deployment opportunity that is moving much more slowly. The US prior authorization system requires submitting clinical documentation to insurers who then review it against coverage criteria — a process that costs the healthcare industry roughly $30 billion per year, according to AMA estimates, and requires a significant fraction of clinical staff time. AI that could automate the documentation assembly and initial submission review could recover a substantial portion of that cost. The technical capability exists. The deployment is complicated by payer-side data requirements, variable insurer processes, and the adversarial nature of the authorization workflow (insurers have an incentive to deny; providers have an incentive to approve).
The EHR Integration Bottleneck
Almost every healthcare AI conversation eventually arrives at the same constraint: Epic.
Epic Systems runs the electronic health record for approximately 37% of US patients, and its architecture decisions determine what AI can and cannot access at most large health systems. Epic has a commercial AI program (AI Now) and a partnership-based ecosystem (App Orchard) that allows third-party AI products to integrate — but the integration standards, data access controls, and workflow integration requirements are entirely on Epic’s terms.
This creates a market structure that looks nothing like other enterprise AI sectors. In healthcare, the AI vendors are not primarily competing with each other for enterprise contracts. They are competing for position within Epic’s ecosystem, or building their businesses around the smaller market of Epic competitors (Cerner, owned by Oracle; Meditech; athenahealth) who have fewer integration constraints but smaller market shares.
Vendors who have figured out the Epic integration requirements — Nuance, Viz.ai, Aidoc — have significant advantages that are essentially structural: they have done the expensive work of building and certifying integrations that their competitors would need to replicate from scratch. This is one reason why enterprise healthcare AI is consolidating faster than enterprise AI in other sectors: the integration bottleneck creates natural barriers to entry that reward incumbents.
What Is Not Deployed Yet and Why
The most significant gap in healthcare AI — the use case that has been piloted more times than any other and deployed at meaningful scale less than almost any other — is clinical decision support for complex diagnoses.
The vision: an AI that synthesizes a patient’s complete history, current symptoms, lab values, imaging findings, and medication list to suggest diagnoses the physician might have missed, flag drug interactions, identify unusual presentations of common conditions, and detect rare diseases that require pattern recognition across large patient populations. IBM’s Watson for Oncology pursued this vision in the 2010s and failed badly. The current generation of LLMs has demonstrated more impressive pattern recognition but has not solved the fundamental problem: clinical decision support in complex cases requires AI that can be trusted at the point of care, where errors have direct patient harm consequences.
The trust problem is not simply that the current AI is insufficiently accurate. It is that the accuracy of current AI in complex clinical reasoning is sufficiently variable — excellent in well-represented conditions, poor in rare or atypical presentations — that a physician who trusts the AI’s output uniformly will be misled in the cases where it matters most. Building appropriate trust calibration — helping clinicians know when to trust the AI and when to be skeptical — is an unsolved user experience and AI transparency problem that the technology industry is only beginning to engage with seriously.
Until that problem is solved, the healthcare AI that matters most will continue to be the kind that is narrow enough to be reliably trustworthy — and the kind whose failure modes are visible enough to be caught before they harm patients.
One email a week with practical Mac workflows, useful comparisons and honest tech recommendations.