The industry has spent billions on healthcare artificial intelligence. Almost none of it has reached production. Architects who have spent their careers in the gap say the failure is not technical. For a sector that has spent the last decade declaring itself transformed by artificial intelligence, U.S. healthcare has remarkably little of it in production.
The investment has been there. Venture funding into healthcare AI has run into the billions of dollars annually for several years. Hospital systems have built innovation labs, signed enterprise pilots, and named chief AI officers. Major academic medical centers each maintain portfolios of dozens of in-house models, covering sepsis prediction, deterioration scoring, readmission risk, imaging triage, and ambient documentation. The clinical literature is large enough that meta-analyses of clinical AI meta-analyses are now routine.
The Missing Layer Can Mean the Difference Between Mere Model Performance and Real-World Impact
What is missing is the layer where any of it touches a patient.
The pattern has acquired a name inside the industry. People who work in healthcare technology call it the Scale Gap. Hospital systems describe long lists of completed AI pilots and short lists of production deployments. Vendors describe long lists of signed pilots and short lists of paid renewals. Chief information officers describe, at industry forums, a recurring failure mode in which the model performs in development, the institution prepares to deploy, and the project ends quietly. The dominant explanation is that the science is not yet good enough. The data is dirty. The models are brittle. The bias is unresolved. Each of those statements is often true. None of them, taken together, accounts for what is going wrong.
Models that have been independently validated, audited for bias, and deployed in narrowly scoped settings still do not scale. Models that scale at one institution rarely transfer to a second. The technical objections are real and downstream.
A small set of architects who have spent their careers inside the gap have a different diagnosis.
“The Scale Gap is a leadership failure, not a technical one,” Rahul Awasthy said in an interview. “The primary hurdle to AI adoption in healthcare is enterprise architecture and governance, rather than data quality or model accuracy.”
Awasthy is the Lead (Principal) Solution Architect at enGen, the technology operating company of Highmark Health, the Pittsburgh-based integrated health and finance organization with about 6.9 million insured members. He has spent 22 years architecting enterprise AI delivery in healthcare and life sciences, including portfolios in the $15 million to $30 million range at Accenture and Tata Consultancy Services and more than 100 enterprise solutions shipped across that span. At enGen, he leads a Blueprint team responsible for the architectural standards behind 25-plus digital health solutions a year, a function that has tracked a 40 percent revenue increase across his most recent two-year period. His credentials include a Ph.D. in data science, an MBA, and an AWS Solution Architect certification. He has watched the same failure pattern repeat for two decades, in different industries, under different leadership, with different vendors, on different stacks.
The Governance Issue is at the Core of Why Healthcare AI Constantly Dies in the Last Mile
The pattern he describes is monotonous. A health system identifies a clinical or operational problem. A vendor or internal team builds a model. The model performs well in a development environment. The institution prepares to deploy. Procurement, security, compliance, clinical informatics, and operations meet for the first time. The architectural questions that should have been answered in week one surface as objections that no one has the authority to resolve. Where does the model run? Who owns it on call at 2 a.m.? Which compliance regime governs its outputs? What audit trail is generated when it is wrong? The pilot ends. The model goes nowhere.
What kills these projects, Awasthy argues, is the absence of a governance structure the institution can defend to its regulators, its clinicians, and its board. When an AI system makes a decision that touches a patient or a payment, somebody has to be willing to put their name on it. If enterprise architecture has not answered that question before the model is built, the institution is dealing with a leadership problem rather than a deployment one.
A second argument, increasingly common among architects in his cohort, concerns how the work is run. The management orthodoxy of modern software development is some version of Agile, with short iteration cycles and working software prioritized over documentation. In consumer software, the orthodoxy is largely correct. In regulated healthcare, the architects who have lived through the Scale Gap argue that it is dangerous.
“Pure Agile is dangerous in healthcare and must be replaced by a Hybrid Governance Model,” Awasthy said. “In a sector where patient lives, and fifty-million-dollar portfolios are at stake, the rigor of the architect is more important than the speed of the developer.”
The hybrid he describes is a synthesis of three frameworks: SAFe Agile to maintain delivery velocity at scale, Lean Six Sigma to enforce process rigor on the parts of the system that touch claims and care decisions, and ITIL service management to govern the operational handoff. The novelty is not the list. The novelty is the explicit hierarchy. Architecture decisions outrank the sprint. The release schedule is subordinated to the risk register. The product manager does not override the compliance lead.
The pattern is uncomfortable for the Silicon Valley healthcare entrants that have driven most recent investment in the sector. The cultural assumption that velocity is the metric to optimize collides, in regulated healthcare, with the operational reality that velocity without architectural rigor produces models that pass the demo and fail the audit.
The Scale Gap diagnosis applies, today, to advisory clinical and administrative AI. The next generation of useful systems will be agentic, acting within governance on the institution’s behalf. Agentic systems require more architecture than advisory ones, not less: explicit accountability frameworks, defended override pathways, audit trails detailed enough to reconstruct any single decision. Institutions that have not built the architecture for advisory systems will not build it in time for agentic ones.
Appropriate Guardrails are Inherently Necessary As Healthcare Enters a New and Uncharted Era
A regulatory window remains open, narrow, and closing. Signals from the Office of the National Coordinator, the FDA’s digital health center, and the Centers for Medicare and Medicaid Services suggest the federal regulatory regime around clinical AI is already hardening. Once it does, the architectural choices will be made for the institutions that have not made them for themselves.
“The organizations positioned to benefit from AI are not the ones who built the safest AI,” Awasthy said. “They are the ones who built the most governed, scalable, and production-hardened agentic systems before the regulatory window forced it.”
The institutions that have spent the most on healthcare AI are not, in general, the institutions that have deployed it. The difference is whether the institution has decided, before building the model, who is willing to put their name on the decision the model makes.
Photo Details: Rahul Awasthy delivering a lecture at The NorthCap University
