ChatGPT Enterprise and HIPAA: What to Verify
Verify the agreement, eligible workspace, covered features, and operating safeguards before using AI with protected health information. Keep approval work separate from patient records.
Updated 5 min read
ChatGPT Enterprise should handle protected health information only after your organization confirms the applicable Business Associate Agreement, eligible workspace, covered functionality, and required safeguards. The Enterprise name alone does not authorize every feature or use. OpenAI’s eligibility documentation specifically distinguishes ChatGPT Enterprise with Regulated Workspace.
Protected health information, or PHI, is identifiable health information protected by HIPAA in the relevant covered-entity or business-associate setting. Involve your privacy and security officers before testing with it. Tools like fenbs can hold nonsensitive policy decisions and approval tasks; keep PHI out of that shared work record.
Which agreement and workspace must you verify first?
A Business Associate Agreement, or BAA, defines responsibilities when a business associate handles PHI for a covered entity or another business associate. OpenAI’s BAA guidance (opens in a new tab) directs sales-managed ChatGPT Enterprise or Edu accounts to sales for a ChatGPT BAA. An API agreement or API configuration does not establish that your ChatGPT workspace is covered.
Ask the contract owner to identify the executed agreement, legal entities, workspace, and authorized use. Record who verified those facts and where the approved agreement is held. Staff should be able to recognize the correct workspace before uploading anything, especially when their account also offers a personal workspace.
Treat missing evidence as unfinished approval work. A colleague saying the company bought Enterprise is not the same as confirming the regulated setup and agreement. The broader ChatGPT for business guide covers organizational use; health information needs this additional, specific review.
Are the exact features in your workflow covered?
OpenAI’s HIPAA eligible products and functionality (opens in a new tab) lists ChatGPT Enterprise with Regulated Workspace among eligible products available with a BAA. It distinguishes covered functionality from excluded functionality. Availability in the interface does not, by itself, show that the use or a connected third-party service falls under your agreement.
The same documentation says improved memory is outside the BAA and disabled by default in regulated Enterprise workspaces. PHI should not be entered when using that functionality. Review the current list rather than relying on a saved list of feature names, because the distinction matters more than whether the button is visible.
Describe the full sequence for review: what the user uploads, which feature runs, whether it calls another service, who receives the result, and where the output is saved. A connection can introduce another organization and another agreement. Review that path explicitly before a pilot expands beyond its original boundaries.
| Area | Evidence to request | Unresolved means |
|---|---|---|
| Agreement | The executed BAA and service scope | Keep the workflow out of PHI use until reviewed. |
| Workspace | Confirmation of the eligible regulated setup | Do not infer coverage from the Enterprise label. |
| Features and connections | A reviewed data flow and covered-function list | Do not extend approval to an unreviewed path. |
| Operations | Named access, review, incident, and lifecycle owners | Assign the missing work before rollout. |
What does HIPAA require your organization to assess?
HHS cloud computing guidance (opens in a new tab) explains the responsibilities of covered entities, business associates, and cloud providers. A cloud provider that maintains electronic PHI can be a business associate even when it cannot read encrypted data. A BAA and an appropriate risk analysis are part of the assessment.
Have the security team review access, authentication, employee devices, downloads, connected systems, incident response, and the destination of generated documents. Assign the privacy team the proposed purpose and permitted disclosure questions. Assign clinical leadership any assessment of clinical use. One sign-off should not silently stand in for all three.
Test the process with fictional examples first. Can a user find the right workspace, recognize a prohibited feature, and send an error to the right team without copying health details into a support ticket? Keep nonsensitive findings and required corrections with the project’s Decisions and rules.
Keep health-data approval work free of patient data
Add tasks for agreement verification, feature review, staff instructions, and the fictional-data pilot. Record the policy decision a person approved, including its limits and review trigger. Use generic examples so the team and its connected assistants can follow the work without receiving PHI.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
Can you remove names and treat the material as de-identified?
Removing names is not enough to establish HIPAA de-identification. HHS de-identification guidance (opens in a new tab) describes two methods: Expert Determination and Safe Harbor. The relevant requirements include more than obvious name fields, and a distinctive narrative can carry identifying details.
Ask the privacy team to approve the method and review representative outputs from the preparation process. Dates, contact details, record identifiers, rare events, and details in free text deserve attention. Do not ask an unapproved AI service to de-identify the original record, because that would disclose the original material before the review is complete.
A fictional example is often a better starting point for learning a workflow. Write an invented scenario from scratch and label it clearly. That lets staff practice instructions and output review without claiming that altered real records have passed a formal de-identification standard.
How much information should an approved workflow use?
Use only the information needed for the approved task. HHS minimum necessary guidance (opens in a new tab) describes the general requirement to limit PHI uses, disclosures, and requests to the intended purpose, with stated exceptions such as certain treatment disclosures. Have your privacy officer determine how the standard applies to the workflow.
Build that decision into an input template. Specify required fields, excluded details, and who prepares the input. A request to improve the readability of generic discharge instructions may need a different input from a patient-specific task. Keep both the preparation procedure and the output review responsibility explicit.
An assistant’s polished answer is not evidence of clinical correctness. Name the qualified person who checks the output and the permitted destination before it is used. The AI governance hub covers broader oversight practices; a vendor comparison such as Claude versus ChatGPT cannot replace a health-data review.
What should staff do when approval is incomplete or something goes wrong?
Give staff a clear fallback: use fictional material or pause the proposed PHI workflow and contact the designated reviewer. If information is entered into an unapproved place, follow the organization’s incident procedure promptly. Preserve relevant evidence through the approved channel without spreading another copy of the sensitive material.
Review the approval when functionality, connected services, users, or purposes change. If evaluating ChatGPT alternatives, obtain a new service-specific assessment. The separate Enterprise overview covers the general product; the agreement and workflow remain the deciding evidence here.
How to do this with fenbs
- Create a free account, then add an AI policy review project using a nonsensitive description.
- Create tasks for each review gap, with generic acceptance criteria. Link only to approved references that do not expose patient information.
- Record human-approved usage rules, the responsible roles, and the circumstances requiring a new review in Decisions and rules.
- Connect an approved assistant through Integrations, ask it to read the rules, and save progress on the policy tasks before ending the session.
fenbs is not the right fit for storing PHI, handling patient-specific clinical work, or replacing a HIPAA review. Its useful role here is the nonsensitive work around approval: what has been checked, what remains, and who must decide. Begin with the missing evidence rather than uploading a patient example.