OpenAI said Tuesday that clinicians can now use ChatGPT to read and summarize patient records stored in Epic Systems’ electronic health records software, a step that puts the AI company inside the systems where much of American medicine is documented.
The integration, announced as part of OpenAI’s ChatGPT for Healthcare offering, gives authorized doctors read-only access to records covering more than 325 million patients whose information flows through Epic-based systems, the company said. Nothing the model produces is written back into the chart.
UCSF Health in San Francisco is the first health system to pilot the integration. Clinicians there can pull up a patient’s history and ask for a summary within ChatGPT or inside Epic’s own workflows, according to OpenAI.
The pitch to hospitals is built on a familiar pain. Clinicians spend a large share of their day on chart review and documentation, and health systems have struggled to hire enough staff to keep up; any tool that compresses the reading part of that work is worth testing. Ambient documentation tools that write notes from a conversation have already won a foothold in hospitals, and OpenAI is aiming at the reading side of the same workload.
The service can also draw on public health sources including CMS and PubMed, which the company said lets clinicians check trial-eligibility criteria, drug identifiers and versions of insurance policies against current standards without leaving the chart.
The read-only design answers the question health systems ask first: whether the AI can alter the medical record. By keeping ChatGPT from writing to the chart, the company addresses concerns about the model changing records it is only meant to help read, while still putting it in the path of daily clinical work.
Epic’s reach gives the integration unusual distribution. The company’s records software is used by many of the largest U.S. hospital systems, and a tool that works inside it reaches clinicians without asking them to abandon the workflow they already use.
Epic has built AI features of its own and works with a range of partners, so the OpenAI integration is one of many inside its software. But ChatGPT brings a name recognition that most vendors cannot match, and hospitals say familiarity matters when asking clinicians to trust an AI with a patient record.
The move extends OpenAI’s strategy of building specialized offerings for specific industries. The company has marketed ChatGPT for Healthcare as a service bound by health-privacy rules, and the Epic connection turns it from a tool clinicians seek out into one that meets them in the chart.
OpenAI is not the only AI company courting hospitals. Microsoft, through its Nuance subsidiary, already sells AI that drafts clinical notes inside Epic’s software, and a cluster of startups has raised large rounds to put documentation and chart-review assistants into exam rooms. Google has also marketed clinical AI to health systems. OpenAI’s pitch is a general model that can handle the reading and reasoning, not just the note-taking.
The companies competing in clinical AI have tended to specialize: one drafts the note, another reads the scan, a third schedules the follow-up. OpenAI’s argument is that a general model can do more of these jobs through a single interface, and that its consumer reach means clinicians already know how to talk to it.
The pilot will test more than accuracy. Health systems operate under strict federal privacy rules, and contracts with AI vendors are being scrutinized as hospitals weigh what their data is worth and who may see it. UCSF’s experience will show whether patients and regulators accept an AI reading records as a routine part of care.
Data handling is part of the negotiation. OpenAI has said its enterprise offerings do not train models on customer data, and health-system contracts are expected to carry similar commitments, with the records themselves staying under the hospital’s control. Those terms, not the technology, will decide how quickly other health systems follow UCSF.
Read-only access narrows the risks, but it does not eliminate them. A summary can still miss or misstate details, and clinicians are expected to verify what the model reports before acting on it. The pilot is designed to test the summaries against clinicians’ own reading of the chart before any broader rollout.
For hospitals, the appeal is workflow economics. Chart review consumes staff time, and summarizing years of a patient’s history is exactly the kind of task administrators hope AI can compress. For OpenAI, the appeal is access: health systems are large, sticky buyers that pay for software that slots into daily work, and Epic’s footprint offers a channel that no consumer app could match.
OpenAI’s earlier health efforts sold an assistant; the Epic integration sells a reader that sits where the work happens. If UCSF’s clinicians find the summaries trustworthy, the model’s next step into medicine will be measured one hospital system at a time.
The integration is also a test of distribution. OpenAI has one of the largest consumer AI audiences in the world, but hospitals do not buy consumer products; they buy software that fits their workflows and their compliance rules. Epic’s footprint gives ChatGPT a path into clinical practice that no amount of consumer awareness could buy, and the pilot will show whether the model performs well enough to keep that path open.
