Epic’s electronic health record system holds data on more than 325 million patients. As of this week, a clinician working inside one of those charts can ask ChatGPT what changed since the last visit and get an answer drawn from the record itself.
OpenAI announced the integration on 1 September, connecting ChatGPT for Healthcare to Epic environments so that clinicians can pull appointment notes, laboratory results, medication lists and specialist documentation into a conversation. In some deployments the assistant sits directly inside the EHR workflow, which means a pre-visit review or a clinical timeline can be assembled without leaving the patient chart.
Read-only, by design
The single most important detail in the announcement is a limitation. Access runs one way. ChatGPT can read an authorised record, but it writes nothing back into the chart. No orders, no notes, no amendments to the medication list.
That constraint matters more than it might sound. The failure mode everyone in health IT worries about is not an AI giving a doctor a bad summary, which a doctor can catch. It is an AI quietly depositing a hallucinated detail into a permanent record that follows a patient across providers for decades. Keeping the connection read-only removes that entire category of risk, at the cost of most of the efficiency gains vendors have been promising from ambient documentation.
Alongside the Epic link, OpenAI added a Healthcare Public Data plug-in that pulls from official sources including ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed and PubMed. The company says it is aimed at work like checking trial eligibility criteria, resolving medication identifiers, and tracking which version of a coverage policy applies. That is unglamorous administrative labour, and it is also where a large share of clinical staff time actually goes.
Organisations that have signed a Business Associate Agreement can now also run ChatGPT Work, Codex, apps and connectors inside their workspace for compliant workflows, which pushes the product from a clinical assistant toward a general-purpose enterprise tool for hospital systems.
The safety number, and what it leaves out
OpenAI says it gathered more than 4,300 physician ratings across 27 clinical use cases, covering things like pre-visit review, clinical timelines, medication review and handoff summaries. Physicians rated 99.1 percent of responses as safe.
Read that figure carefully. It is a strong result by the standards of AI evaluation, and it is also an admission that close to one percent of responses were not rated safe. Apply that rate to a large hospital system running thousands of queries a day and the absolute number of unsafe outputs stops being a rounding error. The people running these deployments know this, which is part of why the read-only boundary exists.
The context around the launch is not comfortable either. In July, a Florida pastor sued OpenAI alleging that ChatGPT gave him a near-fatal recommendation. In May, family members of a user sued over advice related to dosage. OpenAI has consistently maintained that its products are not suitable for diagnosis or treatment, a position that sits awkwardly beside a product now wired into the record systems where diagnosis and treatment happen.
Why Epic is the one that counts
Plenty of AI companies have announced health partnerships over the past two years. Most involved pilot programmes, research collaborations or single hospital networks. Epic is a different order of thing. Its market position in large US health systems is close to structural, and an integration at that level means the assistant reaches clinicians who never asked for it and may have no particular interest in AI.
That is the real test. Tools that succeed in medicine tend to be the ones that disappear into an existing routine rather than demanding a new one. A chatbot that a doctor has to remember to open in another tab loses to the five minutes they do not have. A summary that appears where they are already looking has a chance.
OpenAI reported in July, when it opened ChatGPT for health to all US consumers, that people were sending 300 million health-related queries a week. The consumer side of that has always run on a disclaimer. The clinical side now runs on a real record, an audit trail and a hospital’s legal exposure, which is a considerably harder environment and a considerably more useful one.
What to watch next
The interesting question is whether read-only holds. Every efficiency argument in clinical documentation points toward eventually letting the model draft into the chart with a human approving it, and that is where the regulatory and liability fights will happen. Watch for the first health system that asks for write access, and watch what OpenAI says when they do.
For more coverage of AI in healthcare and enterprise technology, visit Mylistingo.
