The Care Stack

Build healthcare AI that can survive a merger

Victor Wilson · · 4 min read

Health systems change shape all the time. Advocate merged with Aurora. Intermountain combined with SCL Health. Smaller deals, divestitures, joint ventures and EHR migrations happen every year.

When that happens, the integration layer is the first thing to break. Today that layer includes something new: AI agents.

I like a phrase I’ve seen used for this: a transplantable skeleton. Last year a lot of health systems gave their AI a brain, with models that summarize discharge notes, sort prior auth requests and read clinical narratives. This year the job is to give it a skeleton. That means a structure that can move into a new organization and keep working.

How agents get stuck to one organization

Most agent setups I see are built the same way. They call one model provider directly and read from one EHR instance directly. A vendor’s workflow engine runs the steps. The data comes from a warehouse with schemas shaped for this organization only.

Every layer assumes the current org chart is permanent.

That works until the board announces a deal. Then the agents become legacy systems almost overnight. The models are fine and the logic is fine. The problem is that every piece was built for one place.

Picture a revenue cycle agent that took more than a year to build and works well. It reads custom Z-segments from one Epic instance’s interface feeds, plus a warehouse schema that exists only in that system. A few weeks into an integration, the team learns the combined organization won’t keep either one. The agent gets shelved. The engineering was good. It just couldn’t move.

Four parts of a portable skeleton

A skeleton gives structure and allows movement. Your agent infrastructure needs both. I’d focus on four layers.

The data layer. Agents should never read raw EHR tables. Have them read from FHIR as a shared model instead. FHIR helps with interoperability, but for agents its bigger value is portability. When you move from Epic to Oracle Health, or merge two Epic instances that were configured differently, agents that read FHIR resources keep working. Agents tied to Clarity tables or custom ADT feeds don’t.

In practice, build your warehouse landing zone around FHIR resources. Use dbt models to turn each source system’s extracts into standard, FHIR-aligned tables. When a source changes, you rewrite the staging layer. The agents never notice.

Orchestration. If your agent workflows live inside a vendor product, like an RPA tool or a proprietary automation platform, they end when that contract ends. Define the workflows as code instead, in a tool like Prefect or Dagster, or in Snowflake’s own tasks. The test is simple. Can you lift the workflow out, keep it in version control, and run it on a different backend without rewriting the business logic?

The model. The model you use today won’t be the one you use in two years. Put a clear boundary between your agent and the model it calls. Write down the inputs, the outputs and the behavior you expect, and keep the model call behind that boundary. Swapping models should be a configuration change, not a rewrite. Every direct call to one provider’s API, with nothing in between, is debt you’ll pay later.

Identity and access. This is the one teams forget. Agents need patient data, so they need HIPAA-scoped access controls. If those controls are role mappings tied to one organization’s identity provider and department structure, a merger means rebuilding all of them. Define agent permissions by data sensitivity and clinical context instead. “May read medication lists for patients in an active encounter” can move to a new organization. “Member of the North Region pharmacy group” can’t.

If you already run on Snowflake, much of this comes naturally. Secure data sharing, secure views and cross-account replication let you set up a data layer that can move between accounts. Add dbt’s declarative models, and your definitions describe what the data should look like rather than where it came from.

You can go a step further by running the agent’s intelligence inside the platform too: Cortex functions, Snowpark Container Services, Streamlit apps. Then the agent moves with the data platform instead of with the org chart. That matters the day the acquiring system’s CIO decides to consolidate onto one Snowflake account.

The same idea holds on other platforms. Keep the agent logic close to governed data, and define it in code you can redeploy.

Build for the deal you can’t see yet

Most data teams build AI for this organization and this quarter. That’s natural. But leadership turns over, and health systems restructure on a regular cycle. Your AI infrastructure will probably outlive the org chart it was built for.

So build the skeleton first. Make it indifferent to which model sits on top and which organization surrounds it. When a merger comes, the AI that survives is the AI whose infrastructure can move.

Here’s a simple test for any agent you have in production or in a pilot. Draw a line between its business logic and your current organization’s infrastructure: the EHR instance, the schemas, the workflow tool, the model provider and the identity groups. If you can’t draw that line cleanly, fix that before you build the next agent. Start with the data layer, because every other layer depends on it.

Discussion

    Join the discussion