AI Features Are Becoming the New EHR Lock-In

The lock-in has moved up the stack

For two decades, EHR switching costs lived in the interface layer. Hospitals stayed with Epic or Oracle Health because ripping out HL7 feeds and retraining staff was expensive, not because the software was irreplaceable. Modern Healthcare reports that AI is changing where the switching cost sits. Ambient scribes, embedded clinical decision support, and agentic retrieval tools are now tuned on years of institution-specific documentation patterns and workflows. That tuning does not travel. A health system that has spent three years feeding its own data into a vendor’s AI layer faces a migration cost that has nothing to do with interfaces and everything to do with starting the model relationship over.

This matters because the enterprise deals being signed right now assume the opposite. Abridge’s enterprise-wide agreement with the University of Texas System commits an entire health system to one AI layer across its care sites. That is a reasonable operational bet. It is also a decision that will outlast the CIO who signed it, and procurement teams are not yet writing these contracts the way they would write a core EHR agreement, with explicit terms for data export, model portability, and what happens to tuned behavior if the system leaves.

The regulatory backdrop makes this a compliance issue, not just a vendor issue

Information blocking enforcement under the 21st Century Cures Act began in September 2025, with penalties running up to $1 million per violation for health IT developers who impede access to electronic health information. HHS’s pending HTI-5 proposed rule signals an intent to let autonomous AI systems retrieve and share health data through FHIR APIs, which raises the stakes on who actually controls that pipe. ONC’s own work this year, from the EHIgnite challenge on EHI export to the 2026 SVAP standards, is explicitly aimed at keeping data liquid as systems change vendors. An AI layer that cannot be exported or re-platformed works against that goal even if the underlying EHR technically meets certification requirements.

Release notes like EverHealth’s Scribe AI transparency note show how granular this is getting. AI documentation features are now versioned, disclosed, and tied to specific FHIR API configurations inside the EHR itself. That level of integration is good for transparency. It also means the AI feature and the EHR core are no longer separable in any practical sense, which is precisely the condition that makes future migration expensive.

What this changes for procurement

Compliance and IT leaders evaluating AI-enabled EHR contracts should treat model and data portability as a named contract term, not an assumed byproduct of FHIR compliance. That means specifying what exportable artifact exists if the AI vendor relationship ends, who owns the tuned behavior built on institutional data, and how that maps to information blocking obligations if a competing system requests the same data access. ONC certification confirms the EHR can exchange data. It does not confirm the AI layer built on top of it will let you leave.

The Cures Act was built to stop vendors from using data as a cage. AI is quietly building a new one, one workflow at a time.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. The core thesis that AI tuning creates new switching costs is plausible and worth exploring, but the claim that this tuning ‘does not travel’ and creates migration costs distinct from interface costs
Source & Claim Verification	Qwen · hosted fallback (qwen-2.5-72b-instruct)	cleared. Most factual claims are supported by citations, but a few lines lack specific references, such as the discussion on the regulatory backdrop and the implications of the 21st Century Cures Act.
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately reflects key regulatory concerns under the Cures Act, ONC certification, and information blocking rules but lacks explicit mapping to ISO 42001, NIST AI RMF, CLIA, NERC CIP, an
Technical Accuracy	Llama	cleared. The technical premise regarding the non-portability of model tuning and the distinction between FHIR data liquidity and model weights is scientifically sound.
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively identifies and counters potential vendor hype by focusing on the practical implications of AI-driven lock-in and the regulatory landscape, though some sources are irrelevant.
Novelty & Non-Duplication	Grok	held. Core thesis and lead are a direct reframing of the already-wired Modern Healthcare piece on AI raising EHR switching costs, with only incremental procurement/regulatory synthesis that does not clear n
Validation	DeepSeek	cleared. The briefing’s central claim about AI becoming a new, workflow-driven lock-in mechanism is plausible and logically consistent with cited trends, but its specific assertions about enforceability and cu

Sources cited: 44. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.