Healthcare AI at scale: Why the operating layer matters
Published: Sept 24th, 2026
Healthcare AI is moving past early experimentation into routine enterprise use. Recently, the American Medical Association reported that 81% of physicians surveyed in 2026 said they were using health AI, up from 38% in 2023. Use is also expanding into documentation, care planning, diagnosis support and other activities.
That shift changes the technology problem for health systems.
As AI becomes embedded across more clinical, administrative and operational workflows, organizations are no longer managing one AI capability at a time. They are building environments in which multiple models, applications and data sources need to work together.
The challenge is therefore becoming less about acquiring capable models and more about creating the infrastructure that allows those capabilities to operate as part of a coherent healthcare technology environment.
More models create a new layer of technology fragmentation
Consider a health system using an AI tool in radiology, an ambient documentation application in primary care, a patient-risk model and an operational forecasting model.
Each may solve a legitimate problem. Each may also have its own data requirements, interfaces, access controls, monitoring processes and workflow dependencies.
At the individual application level, that complexity can remain largely invisible. At enterprise scale, it compounds.
A model can perform well in isolation while still creating friction if clinicians have to access it through a separate interface, if it cannot access the right patient context, or if its output arrives outside the workflow where a decision is being made. Multiple systems can also create duplicated integrations, inconsistent approaches to monitoring and different expectations around how users interact with AI.
This creates an important distinction between model-level performance and system-level performance.
For technology leaders, the question is no longer only whether a particular model works. It is whether the broader environment can manage multiple AI capabilities consistently, securely and efficiently.
An AI operating layer connects models to systems and data
This is where the idea of an AI operating layer becomes important.
Rather than allowing every AI application to build its own connections into enterprise systems, an operating layer provides common infrastructure between the underlying data environment and the applications and workflows where AI is used.
That can include reusable APIs, healthcare interoperability solutions, model registries, access controls, context management and orchestration services. The purpose is not to create another application for clinicians to navigate. It is to make the infrastructure underneath AI more consistent.
Imagine a clinician working within an existing workflow. The operating layer can help determine what information an application is permitted to access, assemble the relevant context, and use FHIR integration services to connect that information to the appropriate model and return the resulting output to the system where the clinician is already working.
The individual model becomes one component of a broader architecture rather than a standalone implementation.
This matters because every new AI capability does not have to mean another bespoke integration project. The underlying infrastructure can provide reusable patterns for connecting data, models and workflows.
This is fundamentally a product engineering challenge. The opportunity is to design the connective layer that allows healthcare organizations to introduce and manage multiple AI capabilities without repeatedly rebuilding the infrastructure around each one.
AI governance needs to extend into production
Getting an AI system approved for deployment is only one point in its lifecycle.
Once a model is operating in the real world, its inputs, usage patterns and surrounding workflows can change. A model may encounter situations that were not adequately represented during evaluation. Upstream data sources can change. Models themselves can be updated. The way clinicians use an AI capability can also evolve over time.
That makes production visibility an architectural requirement, not simply a governance exercise.
The World Health Organization has emphasized the importance of transparency, expert supervision, rigorous evaluation, human autonomy, safety and accountability in the use of AI for health. Its guidance also highlights the need to evaluate risks throughout the design, development and deployment of AI systems, rather than assuming that an initial review is sufficient.
For technology teams, those principles need to translate into capabilities that operate alongside the AI itself.
That can mean monitoring how systems are being used, maintaining clear version histories, recording relevant activity, validating performance over time and creating controlled escalation paths when an AI system encounters circumstances outside its intended use.
The objective is not to create governance around AI after the fact. It is to build governance into the environment in which AI operates.
Human oversight should be part of the architecture
Healthcare AI also introduces a design question that cannot be answered by the model alone: where should human judgment sit within the workflow? The AMA found that 84% of physicians surveyed wanted AI to be integrated into the EHR workflow, while 88% wanted a dedicated channel for providing feedback when issues arise.
The implication is important. AI cannot be treated as a separate destination that clinicians have to visit. It needs to fit into the environments where clinical work already happens.
That also means being deliberate about where AI recommends, where it assists and where human review remains necessary.
A well-designed workflow might allow an AI system to handle a structured, lower-risk task with minimal intervention, while presenting a clinician with relevant context and supporting information when the decision carries greater clinical significance. The important point is that these boundaries are designed into the product rather than left to users to work out for themselves.
Human oversight becomes part of the architecture.
The operating layer must be designed for a changing AI landscape
Healthcare organizations will not deploy one model and leave it untouched for years.
New models will emerge. Existing models will improve. Organizations will evaluate different vendors and approaches. Some applications will prove valuable, while others will not deliver the expected value.
An architecture tightly coupled to individual models can make each of those changes expensive. If a model is directly embedded into a workflow, replacing it can mean rebuilding the surrounding integration as well.
A more adaptable architecture separates the workflow from the underlying model.
Standardized interfaces, reusable services and consistent evaluation approaches can make it possible to introduce or replace AI capabilities without redesigning the entire workflow around them. The clinical experience can remain relatively stable even as the intelligence underneath it evolves.
That is an important shift in how health systems think about AI architecture.
The goal is not to predict which model will dominate. It is to build an environment that can accommodate change.
The future of healthcare AI will depend on the environment around the models
Healthcare AI is becoming less about individual applications and more about how those applications operate together.
As more models enter clinical and operational environments, the infrastructure around them becomes increasingly important. Data access, orchestration, monitoring, governance and human oversight are not secondary considerations that can be addressed once AI adoption has already taken place. They are part of what makes that adoption sustainable.
The organizations that approach AI this way are not necessarily trying to deploy the greatest number of models. They are creating an environment in which the right capabilities can be connected to the right data, introduced into the right workflows and managed appropriately as those capabilities evolve.
That is the role of an AI operating layer.
At Reveal HealthTech, the opportunity sits in engineering that environment: connecting AI capabilities with healthcare data through healthcare data engineering, enterprise systems and real-world workflows so organizations can scale AI without simply creating another layer of fragmentation.
About us
Reveal HealthTech partners with leading health systems, integrated delivery networks, and healthcare technology organizations to design custom workflow-native platforms, modernize health data pipelines, and architect clinical-grade AI execution environments.
Are you ready to move past disconnected point solutions and build a secure, unified operating layer for enterprise AI? Connect with our technology directors at hello@revealhealthtech.com or visit our Contact Us page to schedule a strategic architecture briefing.
