What to settle before a chat assistant reads the patient record

On September 1, 2026, OpenAI announced that clinicians can pull patient records from Epic into ChatGPT, TechCrunch reported. More than 325 million patients have a current record in Epic, according to Epic. Before clinicians connect a chat assistant to the record, five things need settling: what the business associate agreement covers, who gets access, a test on made-up charts, who reviews the logs, and how clinicians check answers against the chart.

This guide for hospitals, health centers and practices takes the five in order, using OpenAI's own setup guide and the HIPAA rules that apply. The same checks hold for any chat assistant that reads the record. It is not legal advice, so have your compliance lead or counsel review your plan.

What did OpenAI connect, and what can it see?

OpenAI's guide to the Epic plugin says clinicians can review patient information they already have permission to see, including clinical notes, medications, conditions, encounters and lab results. For now the connection is read-only. It cannot update medical records, place orders, message patients or override a chart's existing permissions.

The plugin is available in approved ChatGPT for Healthcare and HIPAA-enabled Enterprise workspaces, and not with individual ChatGPT for Clinicians accounts. Your Epic team requests and enables OpenAI's app in Epic. A workspace administrator then creates the plugin and publishes it for the roles that should have it, and each clinician signs in with their own Epic account.

Does your business associate agreement cover it?

It has to cover it before any patient information reaches the vendor. Under HIPAA, the federal health privacy law, a vendor that creates, receives, maintains or transmits patient information on your behalf is a business associate (45 CFR 160.103). You need the vendor's assurance that it will safeguard the information, documented in a written contract or other written agreement (45 CFR 164.502(e)). That contract is the business associate agreement, or BAA.

The vendor's guide makes the same point. It says to confirm, before anyone uses the Epic plugin with patient information, that your organization has an applicable BAA, appropriate Epic authorization, an approved workspace configuration and approval to handle patient information in the product you're using. It adds that an enabled plugin, a connected app or an accepted warning does not automatically mean that every product, service or workflow is covered by your BAA.

So ask the vendor to confirm in writing which products, features and settings your agreement covers, and check that this plugin, in this workspace, is one of them. The first of the five data questions to ask an AI vendor asks the same thing, whether the agreement covers the AI features specifically.

Who should be able to use it?

Only the roles with a clear need, and you decide that in two places. The plugin uses each clinician's own Epic sign-in and cannot override chart permissions, so the access rules you already set in Epic carry over. In the workspace, an administrator chooses the roles that get the plugin. Start with one group, for example the clinicians who prepare for visits, and widen it later.

HIPAA asks for the same decision. Its minimum necessary rule requires a covered entity to identify the staff, or groups of staff, who need access to patient information to do their jobs, and the kinds of information each of them needs (45 CFR 164.514(d)). The plugin guide tells users the same thing in plain terms: use only the patient information needed for the task.

Clinicians who get access should also know the data rules that apply to every AI tool, such as the five data rules for AI training, which start with keeping patient data out of any tool that is not on the approved list.

How should you test it?

Test it on made-up patients first. The setup guide says to start in a non-production Epic environment and have test users validate the connection with synthetic test records, then set up production as a separate plugin with its own name.

Use the non-production test to check the answers as well as the connection. Build a few made-up charts with details you know, ask the questions your clinicians will ask, and compare each answer with the chart you built.

HIPAA's Security Rule also requires an accurate and thorough assessment of the risks to the electronic patient information an organization holds (45 CFR 164.308(a)(1)). Add the new connection to that risk analysis before it goes live.

Who reviews what the assistant looked at?

A named person, on a set schedule. HIPAA's Security Rule requires audit controls that record and examine activity in information systems that contain or use electronic patient information (45 CFR 164.312(b)). It also requires procedures to regularly review records of that activity, such as audit logs and access reports (45 CFR 164.308(a)(1)).

Ask the vendor, ahead of go-live, what its workspace records about each request, and ask your Epic team how the app's access shows up in your own audit logs. Then write down who reads those records and how often.

Does a clinician still check the chart?

Yes. The plugin guide says clinicians remain responsible for reviewing the underlying record and making care decisions, and it tells users to review the source record and relevant dates before relying on a response.

Language models can also repeat a mistake they are handed. In a 2025 study in Communications Medicine, six large language models repeated or elaborated on a fake lab value, sign or disease planted in text written like a clinical note, in up to 83% of cases.

The evidence map of healthcare AI lists this as a limit that applies to every tool built on a language model. Treat a summary as a guide back to the chart, and check anything you will act on, such as a medication or a lab value, against the record and its date.

What about patients who connect their own records?

That connection is the patient's, and HIPAA treats it differently. In July 2026, OpenAI said it was making ChatGPT's health feature available to all US users over 18 on every plan, and users can connect their medical records from hospital systems such as Epic and Oracle Health, TechCrunch reported.

The Department of Health and Human Services (HHS) says in its guidance on health apps that once a patient directs their records to an app that is neither a covered entity nor a business associate, HIPAA no longer protects that information. If the provider did not supply the app, the provider is not liable under HIPAA for what the app later does with the records. If the patient's electronic health record can be readily produced in the form and format the app uses, HIPAA also generally bars the provider from refusing to send it because of concerns about how the app will use it.

The answer changes when the app is the provider's own. If a patient picks an app that the provider uses to deliver its services, HHS says the provider may be liable if that app discloses the information improperly. So the five checks above are for the connection your clinicians use, and a patient's own connection is a records request to answer.

Where does this fit in an AI policy?

A chat assistant that reads the chart is one entry in an organization's list of AI tools, and it needs more review than a tool that drafts emails. The AI Governance Starter finds every AI tool already in use and writes an AI use policy that places each tool in a risk tier, with the review that tier needs. It also delivers a vendor and BAA checklist, a committee charter and a one-page board summary. It takes three weeks and costs $8,500, fixed. The work uses no patient data.