4 options to Integrate Copilots with EHRs

Every major EHR will support some form of chat-based interface sooner rather than later. The real question is not whether copilots will exist, but how they will be architected.

There are four primary integration options emerging. The key differentiators are:

  • What data can the interface access

  • What actions can it perform

  • What large language models can it support

  • How open the ecosystem is to innovation

Below is a breakdown of the four dominant options.


Option 1: EHR-Native Chat Interfaces

In this model, the EHR vendor builds the copilot entirely within its own platform. The interface, the orchestration, and often the LLM integration are controlled by the EHR.

This is the most vertically integrated and closed model.

Pros:

  • Full access to all available EHR data and services

  • Deep workflow integration

  • Strong control over compliance and governance

Cons:

  • User experience locked into the EHR’s design paradigm

  • Limited or no flexibility in choosing the LLM

  • Minimal room for third-party innovation

This model maximizes control and tight integration, but limits ecosystem flexibility.


Option 2: EHR-Supported MCP Server

In this model, the EHR exposes its data and capabilities through a Model Context Protocol server. External copilots can connect to the MCP endpoint to retrieve context and invoke tools.

Here, the EHR acts as a structured tool provider rather than the UI owner.

Pros:

  • Decouples the user interface and LLM from the EHR

  • Enables third-party innovation in UI, orchestration, and analytics

  • Preserves structured, governed access to EHR capabilities

Cons:

  • Only data and functionality exposed through the MCP server are accessible

  • Vendor-to-vendor variability in MCP implementation may create inconsistencies

This approach balances openness with governance, but remains dependent on what the EHR chooses to expose.


Option 3: Independent MCP Servers

In this scenario, the EHR does not provide an MCP server. Instead, third-party developers build their own MCP layer. These servers consume data from the EHR using public APIs, typically FHIR, or private interfaces, and then expose that data to LLM-driven applications.

This creates a modular architecture where apps, MCP servers, and LLMs are independently developed.

Pros:

  • Maximum modularity across UI, tooling, and models

  • Encourages innovation at every layer of the stack

  • Greater flexibility in model selection and orchestration

Cons:

  • Limited to data accessible via APIs

  • Write-back capabilities may be restricted

  • Dependent on API performance and completeness

This model is attractive for organizations seeking flexibility and faster innovation cycles, especially when real-time write access is not mission-critical.


Option 4: Bulk Export into an External System

In this model, EHR data is periodically exported in bulk into a separate infrastructure environment. That environment hosts its own FHIR server and MCP layer, which then connects to LLMs, analytics systems, and downstream applications.

This architecture is particularly strong for analytics-heavy use cases.

Pros:

  • Full control over database infrastructure, applications, and LLMs

  • Excellent for population health, cohort analytics, and retrospective analysis

  • Scales effectively for large datasets

Cons:

  • Not ideal for real-time, point-of-care use cases

  • Requires reliable extraction and synchronization pipelines

  • Adds operational overhead

This model excels when near-real-time interaction is not required and analytical depth is prioritized over immediate workflow integration.


Choosing the Right Model

These four approaches are not mutually exclusive. Large health systems may adopt multiple strategies simultaneously:

  • Native copilots for tightly integrated workflows

  • MCP-based integrations for third-party applications

  • Independent MCP layers for modular innovation

  • Bulk export environments for analytics and research

The architectural choice depends on the use case. Point-of-care clinical decision support requires low latency and deep integration. Population health analytics benefits from scale and data control. Administrative automation may fall somewhere in between.


What Comes Next

Over the coming weeks, we will share deeper dives into these integration strategies, particularly Options 3 and 4, where modular infrastructure and interoperability play a central role. At Darena Health, we have developed interfaces that support these models while addressing many of their inherent limitations around data access, performance, and workflow alignment. As this space evolves, new integration patterns will undoubtedly emerge.

Are there other strategies we should be considering?

If you are exploring how to bring LLM-powered copilots into healthcare workflows, whether through building, piloting, or early-stage ideation, we would welcome the opportunity to connect and exchange insights.


Previous
Previous

Let’s Talk About the Other Side of AI in Healthcare