ChatGPT Health Connects Epic Records for Clinicians

The Chart Comes to the Chat, Read-Only

OpenAI announced on September 1, 2026 that healthcare organisations can connect their Epic electronic health record environments to ChatGPT for Healthcare. Alongside it came a Healthcare Public Data plugin that pulls from a defined set of public medical sources. The design decision that matters most is the one about direction of travel: the connection is read-only. ChatGPT can read an authorised patient record. It cannot write anything back into the chart.

• Epic integration lets clinicians ground questions in an authorised patient's record instead of searching across notes, labs, medications and specialist documentation
• Read-only by design — the model does not write to the chart
• Healthcare Public Data plugin covers ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed and PubMed
• Safety evaluation: OpenAI reports physicians rated 99.1% of responses as safe across 4,363 ratings spanning 27 clinical use cases with connected EHR context

What Clinicians Can Actually Do With It

The core use case is synthesis across a fragmented record. A patient's history lives in appointment notes, lab results, medication lists and documentation from multiple specialists, often entered over years by people who never spoke to each other. Answering a straightforward clinical question can mean reading all of it.

With the Epic connection, a clinician can ask a question grounded in the authorised record instead. OpenAI describes pre-visit review and building a clinical timeline as primary scenarios, and in some deployments ChatGPT is reachable directly inside EHR workflows without leaving the patient chart. Epic's platform holds records for over 325 million patients, which is the reason this particular integration is the one that moves the needle.

The Healthcare Public Data plugin covers the other half of a clinician's search burden: trial eligibility criteria from ClinicalTrials.gov, medication identifiers from RxNorm and DailyMed, coverage policy versions from CMS, and literature from PubMed.

How Should the 99.1% Safety Figure Be Read?

Carefully, and in both directions.

OpenAI reports that physicians rated 99.1% of ChatGPT responses as safe across 4,363 ratings covering 27 clinical use cases with connected EHR context. That is a substantial evaluation with a real denominator, and publishing the count rather than only the percentage is the right practice.

It is also a vendor-run evaluation of a vendor's product, which is a primary source for its own methodology and not an independent result. And the residual matters: at 4,363 ratings, roughly 39 responses fell outside the safe category. In a clinical setting the interesting question is always what the failures looked like, not what the pass rate was. That analysis is what independent evaluation would add.

Why Read-Only Is the Right Constraint

Keeping the integration read-only draws a clean line. The model can inform a clinician's reasoning; it cannot alter the legal record of care. Every action that changes a patient's chart still passes through a human who is accountable for it.

That constraint also makes the deployment auditable in a way write access would not. If a recommendation turns out to be wrong, the chart shows what a clinician decided and when — the model's contribution sits in the conversation, not in the record. For a first production integration at this scale, that is the conservative choice and the correct one.

Where This Fits in the Broader Picture

Clinical AI has been converging on the same pattern: retrieval grounded in an authorised, permissioned data source rather than a model answering from parametric memory. It is the same architectural instinct behind Gemini Enterprise for Legal bringing AI agents into law firms and behind Claude for Scientists opening 10,000 free research seats — connect the model to the authoritative corpus, constrain what it can change, and let domain experts keep the decision.

The open questions are the ordinary enterprise ones rather than exotic AI ones: how organisations handle consent and access scoping, how audit logging works across the boundary between EHR and model, and how this layers onto AI tooling health systems already run. Our AI coverage will follow how the first deployments handle those.

Sources: OpenAI — Healthcare organizations can now connect EHR and additional industry data to ChatGPT — September 1, 2026; TechCrunch — September 1, 2026; Fierce Healthcare — September 1, 2026.