OpenAI Connects ChatGPT Health to Epic, Reaching Records for 325 Million Patients

OpenAI announced on September 1, 2026 that healthcare organizations can now connect their Epic electronic health record environments directly to ChatGPT for Healthcare, giving authorized clinicians access to Epic's records covering more than 325 million patients. Alongside the Epic integration, OpenAI introduced a Healthcare Public Data plugin connecting the workspace to nine official healthcare data sources, including ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed, and PubMed.

What the Integration Actually Does

• Authorized clinicians can access and summarize patient information — pulling together clinical notes, lab results, medications, and specialist documentation — from within ChatGPT or directly inside supported Epic workflows
• The integration is explicitly read-only: it does not write information back to the patient record, limiting its role to information retrieval and summarization rather than clinical documentation
• Physicians evaluated the system across 27 distinct clinical use cases, including pre-visit reviews, medication reviews, clinical timelines, and handoff summaries

The Safety Numbers Behind the Launch

Across 4,363 physician evaluations spanning those 27 use cases, 99.1% of responses were rated safe. That's a meaningfully specific, published figure rather than a general safety claim, and it's the kind of number worth holding any competing clinical AI tool's own safety claims against directly — ask specifically for a comparable per-use-case evaluation count and safety rate, not just a general assurance.

What This Means for Healthcare IT Decision-Makers

The read-only design is a deliberate and sensible constraint for a first integration of this scale — it limits the blast radius of any AI error to information retrieval rather than record-keeping accuracy, while still addressing a genuine clinician pain point (assembling scattered patient information across a large record). For healthcare IT teams evaluating this or a similar integration, the published per-use-case safety evaluation methodology is worth reviewing directly rather than accepting the headline percentage alone — ask what the 0.9% of unsafe-rated responses actually looked like and in which use cases they clustered.