Why Healthcare Interoperability Is the Real Prerequisite for Medical AI (And Most Systems Aren't Ready)

Every hospital vendor is selling AI agents right now. But a single stat buried in a JAMA Network Open survey should give any healthtech buyer pause: out of more than 2,000 physicians, only 8% said it was very easy to use information from different EHR systems. That is the foundation the industry is trying to build agentic AI on top of — and it is cracked. Interoperability isn’t the boring plumbing story anymore. It’s the gating factor for every AI roadmap a health system is about to sign a check for.

The Quiet Standard Doing the Heavy Lifting: HL7 FHIR

HL7 FHIR — pronounced “fire” and developed by Health Level Seven International — is the specification letting electronic health records exchange data regardless of how each system stores it internally. As Dr. Patrick Guffey, chief informatics and outcomes officer at UCHealth, puts it, FHIR “enables applications to communicate and know what they’re sharing with each other.”

Why it matters: FHIR is what turns a patient’s fragmented paper trail — hospital admission reports, lab results, outpatient follow-ups — into structured, queryable data that downstream systems can actually act on. Without a shared standard, every integration is a bespoke project, and every AI feature ships with a six-month data engineering tail.

Practical example: If you’re a digital health startup trying to plug a triage assistant into three regional hospital networks, FHIR is the difference between a two-week pilot and an 18-month integration slog. It’s also why federal officials keep pushing adoption harder. Teams building AI-integrated software solutions for clinical settings should treat FHIR fluency as table stakes, not a nice-to-have.

Our take: Vendors that ship FHIR-native from day one will win the next round of hospital RFPs. Everyone else will spend the next two years retrofitting.

The 8% Problem Is a Care Problem, Not Just an IT Problem

According to the CDW white paper cited in the original reporting, care is deeply fragmented for both patients and clinicians — and the JAMA survey of over 2,000 physicians quantifies just how bad it is. Only 8% called cross-EHR information use “very easy.”

Why it matters: That number translates directly into duplicate lab orders, delayed treatment while records are chased down, and clinicians making decisions with partial context. When interoperability works, the CDW paper notes, patients transition efficiently from admission to discharge to follow-up, and endpoint-integrated EHRs put accurate data in front of clinicians at the point of care. “To understand the full scope of a patient’s healthcare history, it’s critical to have seamless interoperability between systems,” Guffey says.

Practical example: Imagine a patient discharged from a Colorado hospital who visits an urgent care in Arizona a week later. With interoperable systems, the urgent care physician sees the discharge summary, the medication list, and the pending lab work — no phone tag, no faxed PDFs, no repeat imaging. Without it, the patient becomes the human middleware.

Our take: The next wave of healthcare litigation and payer pressure will target avoidable duplicate testing. Interoperability stops being a quality metric and starts being a cost-containment mandate.

Why AI Agents Need Interoperability Before They Need Better Models

The industry is racing toward agentic AI in clinical settings, and NVIDIA is openly building for it. David Niewolny, NVIDIA’s director of business development for healthcare and medical, describes an architecture where specialized subagents handle radiology, surgery, and other domains, then report to an overarching agent that behaves “more like a human who’s making a decision.” His framing is blunt: “AI agents no longer are just AI tools generating output, they’re generating output and taking action.”

Why it matters: An agent that takes action needs a complete, trustworthy view of the patient. Guffey notes that patient records now exceed what a human can reasonably process — which is exactly the workload AI is meant to absorb. But an agent reading from an incomplete record isn’t summarizing; it’s hallucinating clinical context. Interoperability is what gives the agent something real to reason over, and source links let clinicians audit the output.

Practical example: If a health system deploys a discharge-planning agent that pulls only from its own EHR, it will miss the specialist visit the patient had at a competing hospital last month. Route the same agent through a FHIR-based exchange and it can build a complete medication reconciliation before the patient ever leaves the bed. That’s the design pattern to look for when scoping AI agent development for clinical workflows.

Our take: Within two years, every serious clinical AI deployment will be judged first on data completeness and second on model quality. The model is the commodity. The pipe is the moat.

The Governance Gap Nobody Wants to Fund

Guffey identifies two structural blockers: too many vendors writing their own code, and inconsistent data definitions across providers that magnify over time. His fix is unglamorous — strict data governance and ongoing training on shared definitions. “We want to educate our care team members to make sure we’re all using the same up-to-date definitions and standards so that information flows clearly,” he says.

Why it matters: You can adopt FHIR and still ship garbage data if two departments define “active medication” differently. Governance is where interoperability projects quietly fail — not because the standard is wrong, but because the human process around it never gets funded. The same discipline applies in adjacent regulated industries: KYC and digital identity software fails for the same reason when definitions drift between systems.

Practical example: A hospital rolls out a sepsis-prediction model trained on one unit’s definition of “fever onset” and deploys it hospital-wide. Accuracy collapses in the units that record vitals differently. That’s not an AI problem — that’s a governance problem masquerading as one.

Our take: Expect the next generation of health IT hires to include “data steward” as a first-class clinical role, not a back-office IT function.

The Playbook Health Systems Should Steal From Contexture

Guffey chairs Contexture, described in the source as the largest health information exchange in Arizona and Colorado. His advice to peer institutions is direct: “Release information freely.” Niewolny’s parallel advice on agentic AI is that acceptance takes time, so systems should start piloting now.

Why it matters: Health systems that hoard data lose twice — they get worse AI outputs and they lose the trust of patients who increasingly expect access to their own records. Guffey argues for the maximalist position: “Release as much information to the patient as you possibly can. Patients will tell you if something in there is not accurate.”

Practical example: If you’re a mid-sized hospital network without a mature information exchange strategy, joining a regional HIE this year gets you further than any single vendor AI purchase. Pair that with a small agentic pilot in a low-risk workflow — say, prior authorization drafting — and you build both the data foundation and the organizational muscle at once.

Our take: Health systems that treat patients as auditors of their own records will produce cleaner training data than systems that treat records as proprietary assets.

FAQ

Q: What is HL7 FHIR and why does it matter for AI in healthcare? A: FHIR (Fast Healthcare Interoperability Resources) is a data-exchange standard developed by Health Level Seven International that lets different electronic health records share structured information. It matters for AI because clinical models and agents need consistent, complete patient data to produce trustworthy output — FHIR is what delivers that across systems.

Q: How does healthcare interoperability actually improve patient outcomes? A: Interoperability gives clinicians the full context of a patient’s history — labs, medications, prior visits — without having to reorder tests or chase down records. Per the CDW white paper, it enables coordinated transitions from admission through discharge and follow-up, and it lets patients see their own records, which makes patients partners in their own care.

Q: Can a hospital deploy AI agents without solving interoperability first? A: Technically yes, but the results will be limited to whatever single system the agent can read. NVIDIA’s David Niewolny describes a future where subagents take action across specialties, and that only works when the underlying data is complete and consistent. Interoperability is the prerequisite, not a parallel track.

Key Takeaways

• Treat FHIR fluency as a procurement requirement, not a vendor bonus — RFPs without it will age badly within 24 months.
• Budget for data governance headcount alongside every clinical AI purchase; models trained on inconsistent definitions will fail silently in production.
• Join a regional health information exchange before piloting agentic AI, so the agent has real cross-organization data to reason over.
• Expand patient access to their own records aggressively — patient corrections are one of the cheapest sources of data quality improvement available.
• Watch NVIDIA’s healthcare agent architecture closely; the subagent-plus-orchestrator pattern is likely to become the reference design other vendors copy within the next 18 months.