ChatGPT Health Hits Epic's 325M-Patient EHR With Read-Only Access

OpenAI announced on September 1, 2026 that ChatGPT Health is integrating with Epic's electronic health record system — a platform holding data for over 325 million patients — giving clinicians the ability to pull structured patient data directly into a ChatGPT session. The integration is strictly read-only: the AI cannot write, update, or annotate anything back into the EHR. In certain deployments, ChatGPT is embedded directly inside Epic's workflow UI, meaning clinicians can run pre-visit reviews and build clinical timelines without leaving a patient chart.

EHR-connected LLM deployments force access-control and compliance decisions that most enterprise AI integrations have so far avoided. The read-only constraint is the first load-bearing architectural choice, and it directly shapes what infrastructure governance frameworks are required for safe agent deployment in clinical contexts.

Epic Integration: What It Exposes

The data categories available through the integration are appointment notes, laboratory results, active medications, and specialist documentation. Clinicians can query across that corpus to identify changes in patient history and generate handoff summaries. The in-chart embedding — available in select deployments — presents ChatGPT as a native Epic interface element, which raises questions about permission scoping and session context isolation.

Healthcare Public Data Plugin

Alongside the Epic connector, OpenAI is shipping a Healthcare Public Data plugin that queries five external sources: ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed, and PubMed. The plugin lets clinicians cross-reference trial eligibility criteria, medication identifiers, coverage policy versions, and provider records without switching contexts. Architecturally, this is a retrieval-augmentation layer sitting parallel to the EHR feed — structured proprietary patient data on one side, public clinical knowledge bases on the other — a design choice with direct implications for reducing hallucination risk on drug-interaction and formulary questions.

BAA Coverage

Organizations that execute a Business Associate Agreement with OpenAI gain access to ChatGPT Work, Codex, apps, and connectors for HIPAA-compliant workflows. Codex inclusion is notable: it extends BAA coverage to code-generation workflows, with direct implications for clinical engineering teams building on the platform.

Safety Metrics

OpenAI collected over 4,300 physician responses across 27 clinical use cases — pre-visit review, clinical timelines, medication review, and handoff summaries among them — and found that 99.1% of responses were rated safe. OpenAI itself acknowledges that even a small number of unsafe responses can produce harmful outcomes.

The liability picture is already complicated. A Florida-based pastor filed suit alleging a near-fatal ChatGPT recommendation, and a separate suit filed in May names OpenAI in connection with harmful dosage advice. Both cases predate this integration, but the Epic rollout materially increases the surface area: a read-only connection does not prevent a clinician from acting on a response that pulls from a patient's own records and returns something clinically incorrect.

OpenAI is moving the enterprise AI debate in healthcare away from consumer wellness apps and toward credentialed clinical workflows with real PHI in the context window. A 99.1% safe-response rate may satisfy an internal threshold, but it is not a figure that regulators or plaintiff attorneys are likely to treat as a ceiling. How OpenAI governs the autonomy boundaries of embedded clinical agents will set the template the rest of the healthcare AI market follows or avoids.