"HIPAA compliant" in a product page is not enough to approve an AI tool for patient data. You need to know which product and features are covered, what data they process, and how the deployment is configured.
What HIPAA governs
HIPAA applies to covered entities, business associates, and relevant subcontractors when they create, receive, maintain, or transmit Protected Health Information. PHI is individually identifiable health information held in that regulated context, such as a patient name paired with a diagnosis or appointment record. HHS explains the relationship and contract requirements in its business associate guidance.
If an AI vendor creates, receives, maintains, or transmits PHI on behalf of a covered entity or business associate, the parties generally need a Business Associate Agreement. A BAA is necessary in that relationship, but it does not make every feature or configuration safe to use.
Where the product check breaks down
Product names alone are not enough. OpenAI lists specific HIPAA-eligible products and features; Anthropic says its BAA covers qualifying services but not ordinary Claude Free, Pro, Max, or general Claude for Work chat products in its BAA guidance; and Google documents which Workspace Gemini configurations can support HIPAA workloads.
Pasting a patient note into a consumer or otherwise uncovered AI product can create an impermissible disclosure. Check the current contract and eligible-services list rather than assuming a vendor's consumer, business, and API products have the same status.
This doesn't make AI off-limits for clinical or administrative work. It means the specific service must be approved for the intended data and use. Confirm the signed BAA, eligible features, retention settings, subprocessors, access controls, and whether the tool sends data to another service.
Some administrative tasks, such as drafting generic training material, can be done without PHI. Other administrative work contains patient-specific data even when no clinical decision is involved. Classify the data, not just the task name.
The practical checklist
Before using an AI tool with PHI: Is your organization covered by the right BAA? Is the exact product and feature eligible? Are retention, logging, connectors, and access configured correctly? Does the workflow send data to any uncovered service? Have you completed the required risk analysis?
If the vendor will not cover the service under a BAA, do not send PHI to it on behalf of a regulated organization. De-identification or a different eligible service may be an option, but that decision belongs in the organization's privacy and security review.
A signed agreement is only one control. HHS treats risk analysis as foundational, so the review also has to cover the actual architecture and operating process.
Build notes: a review you can repeat for each new tool
Create a one page review that every new AI tool must pass before staff use it with patient information. Record what the tool does, what data it receives, where that data is stored, whether it is used to train models, who at the vendor can see it and whether the vendor will sign a business associate agreement that covers the specific service you plan to use.
Check the agreement, not the marketing page. A vendor may sign a BAA for one product tier and not another, and some features within a product may be excluded. Write down which features are in scope and set the tool up to use only those.
Limit the data. Use the minimum necessary information for the task, remove identifiers where the task does not need them, and prohibit staff from pasting patient details into tools that have not passed review. Train staff with two or three real examples of acceptable and unacceptable use. Keep a list of approved tools and review it at least annually or when a vendor changes its terms. This is operational guidance and not legal advice, so have your privacy officer or counsel confirm the decisions that matter.
