HIPAA-Compliant AI: Which Tools Sign a BAA
No AI model is HIPAA compliant on its own. A deployment is: a vendor that signs a Business Associate Agreement, a product tier and feature set that agreement actually covers, and settings your practice controls. Here is where OpenAI, Anthropic, Google, Microsoft and AWS stand as of October 2026.
FIG · 01The tool can be eligible. Only your use makes it compliant.
HIPAA-compliant AI takes three layers. Miss any one and every prompt holding patient data is a disclosure to a vendor with no agreement.
1. A signed BAA
The vendor is your business associate the moment PHI reaches it.
45 CFR 164.502(e)
2. A covered product and feature
BAAs list tiers and features. Anything off the list is outside it.
Vendor HIPAA scope page
3. Your safeguards
Risk analysis, access control, a written policy and trained staff.
45 CFR 164.308(a)(1)
Never for PHI: an ordinary personal chatbot account. Apart from OpenAI's ChatGPT for Clinicians, which verified US clinicians can sign individually, every BAA below attaches to a business account: a Workspace domain, a Microsoft tenant, an enterprise workspace or an API organization.
What Makes AI HIPAA Compliant
Start with the definition, not the marketing. A business associate is anyone who “creates, receives, maintains, or transmits protected health information” on a covered entity's behalf (45 CFR 160.103). When a biller pastes a denial letter with a patient's name and diagnosis into a chatbot, the AI company receives and processes PHI. It is your business associate whether or not anyone signed anything. If you are unsure your practice is a covered entity at all, check who HIPAA applies to first.
OCR settled the edge cases years ago in its cloud computing guidance. A cloud provider that processes or stores ePHI is a business associate even if it cannot read the data, because the conduit exception is limited to transmission-only services like the postal service. And using one without a BAA puts the covered entity “in violation of the HIPAA Rules” under 164.308(b)(1) and 164.502(e). OCR cites a resolution agreement with a covered entity that kept ePHI on more than 3,000 people on a cloud server with no BAA. An AI vendor is a cloud provider with a chat window on the front.
“HIPAA certified AI” is a sales phrase. HHS does not endorse, certify or recommend any technology product, so no badge replaces the BAA and your own risk analysis.
The BAA is the floor, not the finish. Microsoft says it plainly on its own HIPAA page: having its BAA does not ensure your compliance, and how you use the service is your responsibility. The same holds for every vendor in the next section.
Takeaway: if a prompt, an upload, a recording or an output could identify a patient, the AI vendor needs a signed BAA before anyone on staff uses it. If a vendor refuses one, you can propose your own with the BAA template generator; if it still says no, it never sees PHI.
Which AI Tools Sign a BAA
Every row below comes from the vendor's own HIPAA page, read on 2026-10-22. These pages change often (OpenAI's and Anthropic's were both revised within days of this check), so re-read the scope page on the day you sign, and keep a dated copy with the agreement.
| Vendor | Covered with a BAA | No BAA |
|---|---|---|
OpenAI API: an org admin accepts the standard BAA in API Platform settings (needs API usage history). ChatGPT: sales-managed Enterprise, or a verified clinician's in-product BAA. |
|
|
Anthropic (Claude) Enterprise: the Primary Owner activates HIPAA and accepts the BAA under Data and privacy. API: for HIPAA-ready API organizations; ask Anthropic to enable it. |
|
|
Accept Google's BAA for the Workspace or Cloud account; coverage follows Google's list of included services. |
|
|
Microsoft Included by default through the Data Protection Addendum for covered entities and business associates. |
|
|
AWS Sign the AWS BAA before any PHI; eligibility is per service. |
|
|
Two patterns stand out. First, the cheapest business tier is usually the trap: ChatGPT Business has admin controls and still gets no BAA, and admin controls without an agreement protect nothing. Second, if your practice already runs on Microsoft 365 or Google Workspace under a signed BAA, the AI built into those suites is the shortest path, because the agreement you have already reaches Copilot or Gemini. Our guides to Microsoft Teams and Google Drive walk through those two BAAs.
Buying a healthcare AI app instead? Scribes, chatbots and coding assistants usually run on one of the models above. Ask for two things: their BAA with you, and proof they hold a BAA with the model provider. HIPAA requires that chain of assurances for subcontractors (164.502(e)(1)(ii)).
FIG · 02What the BAA Leaves Out
A signed BAA covers features, not a brand. Inside the same workspace, one button can be covered and the one next to it not. These are the exclusions practices miss most often, straight from each vendor's scope page. The fix is the same everywhere: switch the feature off for anyone who handles PHI, or keep PHI out of it by written policy.
OpenAI (ChatGPT)
- Improved memory, which draws on past chats. Off by default in the healthcare workspaces; turning it on puts PHI outside the BAA.
- Browser use, network access and event-triggered scheduled tasks for Work in the cloud.
- Internet access inside Canvas and code blocks.
- Codex running on your own machines: the BAA covers what reaches OpenAI, not the local install.
Features not on OpenAI's covered list are not automatically covered, including new ones.
Anthropic (Claude)
- Batch API, Files API, Skills API, Code Execution and Web Fetch on the API.
- MCP servers, connectors and Enterprise Search: data sent to a third party is outside Anthropic's BAA.
- Claude Console, Cowork in the cloud and beta features such as Claude in Office.
- Some covered models need 30-day data retention, so plan for prompts being held that long.
The BAA covers only the one organization that accepted it, not sister orgs or personal accounts.
Google (Gemini)
- Gemini in Google Chrome, excluded by name from the Workspace list.
- NotebookLM, which does not appear on the Workspace list of covered services.
Coverage follows Google's list of included services; its Workspace help page says more products may be added.
Takeaway: print the vendor's covered-feature list, walk the admin console with it, and switch off every toggle that is not on the list. Then date and file the printout. When OCR asks how you limited the tool, that page is your answer.
Practice Use Cases: BAA or Not
The question is never “is this AI tool allowed?” It is “does this task put PHI into it?” Here is how the tasks we see most at small practices sort out.
Drafting a denial appeal with the patient's name, date of birth and diagnosis
BAA requiredIdentifiers plus health data is PHI. Covered tool, covered feature, minimum needed.
Summarizing a faxed referral packet or a prior authorization file
BAA requiredThe whole packet is PHI, including the cover sheet.
An ambient scribe recording and drafting visit notes
BAA requiredThe scribe vendor is a business associate, and so is its model provider. Audio retention belongs in your risk analysis.
A website chatbot that books visits or takes symptoms
BAA requiredPatients type PHI into it. BAA with the chatbot company and proof of its BAA with the model behind it.
Rewriting your privacy policy, a job ad or a staff memo
No PHI, no BAANo patient data, no BAA needed. Train staff not to paste a real case in as an example.
A coding question about a generic scenario with no identifiers
No PHI, no BAAFine if it is truly de-identified. A rare diagnosis plus a date can identify someone in a small town.
Staff pasting a chart note into a personal chatbot on their phone
NeverAn impermissible disclosure, presumed a breach unless your four-factor assessment shows low probability of compromise.
The last row is the one to plan for, because it happens without anyone asking first. A disclosure to an AI vendor with no BAA is not a permitted use, so the breach notification rules apply: you run the four-factor risk assessment in 45 CFR 164.402 and notify unless it shows a low probability of compromise. That incident belongs on your incident report form, and the pattern usually shows up on any list of common HIPAA violations.
De-Identify First: The No-BAA Route
Information that meets HIPAA's de-identification standard is not PHI, so it can go into any tool. There are two ways to get there (45 CFR 164.514(b)): Expert Determination, where a qualified statistician documents that the re-identification risk is very small, and Safe Harbor, where you strip the identifiers below from the patient and from relatives, employers and household members, and have no actual knowledge that what remains could identify someone.
Safe Harbor: remove all 18
- 01Names
- 02Geography smaller than a state
- 03Dates (except year), ages over 89
- 04Phone numbers
- 05Fax numbers
- 06Email addresses
- 07Social Security numbers
- 08Medical record numbers
- 09Health plan beneficiary numbers
- 10Account numbers
- 11Certificate and license numbers
- 12Vehicle identifiers and plates
- 13Device identifiers and serials
- 14Web URLs
- 15IP addresses
- 16Biometrics (finger, voice)
- 17Full-face photos
- 18Any other unique code
ZIP codes may keep their first three digits only where that area holds more than 20,000 people. Ages over 89 collapse into one “90 or older” group.
On paper it is a checklist. In a chat box it fails quietly, because free-text notes hide identifiers in prose: “her son drives for the county,” a visit date in the middle of a sentence, a scanned page with the MRN in its header. The rule strips even dates and ages over 89 for a reason: next to a rare diagnosis, they point to one person.
Takeaway: de-identifying by hand, one prompt at a time, is a policy that relies on every staff member getting every prompt right. Use it for de-identified datasets you prepare deliberately. For daily work with real charts, use a tool under a BAA.
FIG · 037 Steps Before PHI Goes In
The order matters: contract, then risk analysis, then configuration, then people. A practice that trains staff on a tool it has not configured yet has trained them on the wrong tool.
- 1
Confirm the exact product, then sign
164.308(b)(3)
Match the plan on your invoice to the vendor's eligible list, word for word. “Enterprise” on a sales call is not “Enterprise with Regulated Workspace” on the contract. The assurances must be in a written contract; file it with a dated copy of the scope page.
- 2
Add the tool to your risk analysis
164.308(a)(1)(ii)(A)
Record where prompts, uploads, outputs and logs live, and for how long (some of Anthropic's covered models need 30-day retention, for example). Our risk assessment tool has a row for each system that touches ePHI.
- 3
Switch off what the BAA excludes
Vendor scope page
Memory, connectors, plug-ins, web fetch, browser agents: off for every role that handles PHI. Re-check after each product update.
- 4
One workspace, company sign-in only
164.312(a)(1)
SSO with MFA, no personal accounts, departing staff removed the same day. A BAA covers the organization that signed it, not a doctor's home login.
- 5
Paste the minimum
164.502(b)
A covered tool still gets only what the task needs. Apply the minimum necessary rule to prompts: the denial reason and the codes, not the whole chart.
- 6
Write the AI rule and train on it
164.308(a)(1)(ii)(C)
List approved tools, ban personal chatbots for patient work, and name the sanction. Add it to your HIPAA policies and to the annual session in your HIPAA training.
- 7
A human signs what goes in the chart
Practice policy
An AI draft that lands in the record is the signer's note. Read it against the source before it is saved, and keep the audit log turned on.
Section 1557 and Clinical AI
HIPAA covers the data. A separate rule covers what the tool decides. Under the Affordable Care Act's nondiscrimination rule, a covered entity must not discriminate on the basis of race, color, national origin, sex, age or disability “through the use of patient care decision support tools” (45 CFR 92.210). The definition is broad: any automated or non-automated tool used to support clinical decision-making, from a risk score in the EHR to a triage chatbot.
92.210(b)
Identify
An ongoing duty to find tools whose inputs measure a protected trait.
92.210(c)
Mitigate
Reasonable efforts to reduce the risk of discrimination from each one.
92.1(b)
Since May 2025
Due within 300 days of July 5, 2024, which fell on May 1, 2025.
Who it reaches: “covered entity” means something different here. In Section 1557 it is a recipient of federal financial assistance from HHS, not HIPAA's provider, plan or clearinghouse. Ask your counsel whether your payer mix brings you in. If it does, keep a one-page inventory of every clinical AI tool, its inputs, and what you did about each.
Quick Reference: HIPAA-Compliant AI
No BAA, no PHI
An AI vendor that receives PHI is a business associate. Ordinary personal chatbot accounts never qualify.
Eligible tiers (Oct 2026)
ChatGPT for Healthcare, Enterprise with Regulated Workspace, ChatGPT for Clinicians, OpenAI API with Modified Retention; Claude API and Enterprise with HIPAA on; Gemini in Workspace; Microsoft Copilot; Amazon Bedrock.
Features, not brands
Memory, connectors, web fetch and beta features sit outside several BAAs. Switch them off for PHI roles.
No-BAA route
Safe Harbor de-identification: all 18 identifiers removed, no actual knowledge of who it is.