AI in your practice
The plain answer for ChatGPT, Microsoft Copilot, Google Gemini and Claude, what a Business Associate Agreement does and doesn't cover, and what is safe to paste. Every vendor fact is linked to the vendor's own page.
By Sohel Siddique, who builds Lydoh and VoiceboxMD. Updated November 4, 2026.
The short answer
The ChatGPT most people use (Free, Go, Plus and Pro) is not: OpenAI does not sign a Business Associate Agreement for it, so a practice can't put patient information into it. ChatGPT Business has no BAA either. OpenAI does sign BAAs for ChatGPT for Healthcare, ChatGPT Enterprise with a regulated workspace, ChatGPT for Clinicians and its API, and then only for the features on its covered list (OpenAI eligible products; OpenAI BAA).
The same pattern holds for the other tools practices use. Microsoft 365 Copilot on a work account, Gemini in a Google Workspace with a signed BAA, and HIPAA-ready Claude Enterprise can be covered. The free and personal versions of all of them cannot. And a signed BAA is the start, not the finish: the vendors themselves say it does not make your use compliant on its own.
Vendor terms checked November 4, 2026. They change often; recheck before you buy. General information, not legal advice.
Vendor by vendor
What each vendor's own HIPAA page says today. If a product isn't listed by the vendor as covered, treat it as not covered.
| Product | BAA? | What the vendor says |
|---|---|---|
| OpenAI (ChatGPT) | ||
| ChatGPT Free, Go, Plus, Pro | No BAA | Not on OpenAI's HIPAA-eligible list. OpenAI says the consumer Health feature "does not offer a Business Associate Agreement" (eligible list, Health in ChatGPT). |
| ChatGPT Business | No BAA | "We do not offer a BAA for ChatGPT Business" (OpenAI). |
| ChatGPT Enterprise, Edu | BAA, with conditions | Sales-managed accounts only, through OpenAI sales. The eligible list names "ChatGPT for Enterprise with Regulated Workspace" (OpenAI, eligible list). |
| ChatGPT for Healthcare | BAA available | Enterprise product sold through OpenAI sales; improved memory is off by default and not covered (OpenAI). |
| ChatGPT for Clinicians | BAA, with conditions | Free for verified US physicians, NPs, PAs and pharmacists; an individual BAA in Settings, not one that covers staff (OpenAI). |
| OpenAI API | BAA, with conditions | Self-serve BAA for organizations with an established history of API usage; requires Modified Retention unless OpenAI says otherwise, and covers listed endpoints only (OpenAI, eligible list). |
| Microsoft (Copilot) | ||
| Microsoft 365 Copilot and Copilot Chat (work account) | BAA available | "Microsoft Copilot" and "Microsoft Copilot Chat" are on the Office 365 in-scope list. The BAA comes by default through Microsoft's Data Protection Addendum for covered entities and business associates (Microsoft). |
| Copilot with a personal Microsoft account | No BAA | Not on the in-scope list; Microsoft's BAA covers its enterprise services (Microsoft). |
| Google (Gemini) | ||
| Gemini app, Gemini in Workspace, Gemini Mac app (Workspace account) | BAA, with conditions | Covered once a Workspace administrator accepts Google's BAA. Gemini in Chrome is excluded. Without a BAA, Google says customers "must not use PHI" (Google Cloud, Workspace Help). |
| Gemini with a personal Google account | No BAA | No BAA. Google says human reviewers read some chats and asks users not to enter confidential information (Gemini Privacy Hub). |
| Anthropic (Claude) | ||
| Claude Free, Pro, Max, Ultra, Team | No BAA | "Team plans and individual plans (Free, Pro, Max, and Ultra) can't enable HIPAA" (Anthropic). |
| Claude Enterprise (HIPAA-ready) | BAA, with conditions | The Primary Owner turns on HIPAA and accepts the BAA. Chat is covered; connectors, skills, Cowork in the cloud and beta features are not (Anthropic, BAA scope). |
| Anthropic API, Claude in Microsoft Foundry | BAA, with conditions | Covered by Anthropic's BAA, but not every API feature (Anthropic). |
"BAA, with conditions" means a BAA exists but depends on the plan, a setting an owner must turn on, eligibility, or a limited feature list. Disclosure: Lydoh's phone assistant runs on Anthropic's models, so we are a customer of one vendor in this table (our subprocessors).
Our take for a small practice
Start with the office suite you already pay for, not a new AI subscription. If the practice runs on Microsoft 365 or Google Workspace and has signed that vendor's BAA, Copilot Chat or Gemini may already be covered for the whole staff (Microsoft; Google). Check that before anyone signs up for anything.
ChatGPT for Clinicians is a good tool for the doctor and the wrong answer for the office. It is free and comes with an individual BAA, but only verified physicians, NPs, PAs and pharmacists can join, and it is designed for one person (OpenAI). Your front desk and billing staff, who write most of the letters, are not covered by the doctor's account. Dentists are not on OpenAI's current eligibility list.
The bigger risk is the personal account, not the vendor. A practice can sign every BAA on this page and still leak patient information the first time someone pastes a note into ChatGPT on their phone. A one-page rule that names the approved tools, plus the checklist below, does more than any contract.
Reading the agreement
A Business Associate Agreement is the written contract HIPAA requires before a vendor may create, receive, maintain or transmit patient information for your practice (45 CFR 164.502(e)). It must limit what the vendor may do with the information, require safeguards and Security Rule compliance, require the vendor to report improper disclosures and breaches, bind its own subcontractors, and require return or destruction of the information at the end where feasible (45 CFR 164.504(e)).
What it does not cover:
The rules underneath
An AI vendor that handles patient information is a business associate. HIPAA defines a business associate as anyone who creates, receives, maintains or transmits protected health information on a covered entity's behalf (45 CFR 160.103). HHS's own examples include a third-party vendor chatbot on a provider's patient portal that handles symptom assessment, medical reminders and appointment scheduling. The conduit exception does not help: it is limited to transmission-only services, and anyone who accesses the information regularly to perform a service is not a conduit (HHS business associate guidance).
Minimum necessary applies to what you send. A practice must make reasonable efforts to limit protected health information to the minimum necessary for the purpose (45 CFR 164.502(b)), and must decide which staff need which information (45 CFR 164.514(d)). The rule's exception for disclosures to a health care provider for treatment doesn't fit an AI vendor, which is not treating anyone. Even with a BAA, paste the paragraph the task needs, not the chart.
De-identified data is outside HIPAA, but de-identifying is harder than deleting a name. HIPAA allows two methods (45 CFR 164.514(b)). Expert Determination needs a qualified expert to find, and document, a very small risk of re-identification. Safe Harbor needs 18 kinds of identifiers removed, for the patient and their relatives, employer and household, and no actual knowledge that what remains could identify them. The 18 include names; any place smaller than a state, including street, city and most of the ZIP code; every date element except the year, and ages over 89; phone and fax numbers; email; Social Security, medical record, health plan and account numbers; license, vehicle and device identifiers; web addresses and IP addresses; biometrics such as voiceprints; full-face photos; and any other unique number, characteristic or code. HHS says identifiers must come out of free text too, and that clinical narratives "are information rich and may provide context that readily allows for patient identification." Once data truly is de-identified, the Privacy Rule no longer restricts it (HHS de-identification guidance).
A slip is presumed to be a breach. A use or disclosure the Privacy Rule doesn't permit is presumed to be a breach unless the practice shows a low probability that the information was compromised, weighing what was disclosed, to whom, whether it was actually viewed, and how the risk was mitigated (45 CFR 164.402).
Print this
Ten questions for anyone in the practice, in the order they come up.
In practice
Each row is a real front-office or clinical task. "Personal account" means any consumer AI tool without a BAA; "covered tool" means a feature on the vendor's BAA list, used from the practice's account.
| Task | Personal account | Covered tool | Why |
|---|---|---|---|
| Rewrite the new-patient welcome letter or the cancellation policy | Fine | Fine | No patient information involved. |
| Draft a referral or prior authorization letter template with [PATIENT] placeholders | Fine | Fine | The template holds no PHI; staff fill in the details in the EHR. |
| Explain a payer's published policy you paste in | Fine if no patient details are added | Fine | Public documents aren't PHI. |
| Summarize a named patient's visit note | Not allowed | Allowed in a covered feature, with clinician review | Identifiable health information going to a vendor needs a BAA (45 CFR 164.502(e)). |
| Draft an appeal for a denied claim with name, birth date and claim number | Not allowed | Allowed, sending only what the appeal needs | Account and claim numbers are identifiers; minimum necessary still applies. |
| "A 52-year-old man seen March 3 in Wayne, NJ, with..." | Not allowed | Allowed | The date and the town are Safe Harbor identifiers, so this is not de-identified. |
| Upload a screenshot of tomorrow's schedule to draft reminder texts | Not allowed | Allowed but sends too much | Names, visit reasons and record numbers ride along. Send names and times only, or use a tool built for reminders. |
| Connect a third-party plugin or connector to patient data | Not allowed | Not covered by OpenAI's or Anthropic's BAA | Data sent to third parties through these features is outside the BAA, by the vendors' own terms. |
| Let the AI decide whether a patient's symptom is urgent, with no clinician check | Unsafe | Unsafe | Not a privacy question. A BAA doesn't make the answer correct; a clinician decides. |
Sources: 45 CFR 164.502, 45 CFR 164.514, OpenAI eligible functionality, Anthropic BAA scope.
Disclosure
Some AI tools handle patient information by design: the phone, and the note. We make one of each. Lydoh answers a practice's calls and signs a BAA with eligible practices; how it handles call data is on the HIPAA answering service page. VoiceboxMD, our sister product, writes clinical notes (voiceboxmd.com). Ask us, and any vendor like us, the same questions this page asks of ChatGPT: who signs the BAA, which subcontractors see the data, and which features it covers.
FAQ
Not the versions most people use. ChatGPT Free, Go, Plus and Pro are not on OpenAI's list of HIPAA-eligible products, and OpenAI says it does not offer a BAA for ChatGPT Business. Without a Business Associate Agreement, a practice can't put patient information into it. OpenAI does sign BAAs for ChatGPT for Healthcare, ChatGPT Enterprise with a regulated workspace, ChatGPT for Clinicians and its API, and even then only the features on its covered list are covered.
There are versions OpenAI will sign a BAA for: ChatGPT for Healthcare and sales-managed ChatGPT Enterprise or Edu (through OpenAI sales), ChatGPT for Clinicians (an individual BAA for verified US clinicians, accepted in the product), and the OpenAI API with Modified Retention. A BAA makes compliant use possible; it doesn't make the tool compliant on its own. OpenAI itself says accepting a BAA does not, by itself, make your use HIPAA compliant.
No government body certifies AI tools as HIPAA compliant. In practice it means three things: the vendor signs a Business Associate Agreement with your practice before any patient information reaches it, you only use the features that agreement covers, and your practice still applies its own HIPAA rules, such as sending only the minimum necessary and controlling who has access.
Usually not, if you mean a tool without a BAA. Removing the name is not de-identification. HIPAA's Safe Harbor method requires removing 18 kinds of identifiers, including all dates except the year, any geographic unit smaller than a state, phone numbers, record numbers and full-face photos, and you must not know of a way the rest could identify the person. HHS notes that clinical narratives are information-rich and can identify a patient even after the obvious identifiers are gone. Safer: write the prompt with placeholders and no patient facts at all.
No. OpenAI, Microsoft and Google all say their BAA supports, but does not by itself achieve, your compliance. The BAA covers only the listed products and features, often not memory, connectors, plugins or browser extensions, and only the organization that signed it. Your practice still decides who uses the tool, what goes into it, and who checks the output.
Microsoft lists Microsoft Copilot and Microsoft Copilot Chat among the Office 365 services in scope for its HIPAA BAA, which it makes available by default through its Online Services Data Protection Addendum to customers that are covered entities or business associates. That applies to work accounts on Microsoft's commercial services. Consumer Copilot used with a personal Microsoft account is not on Microsoft's in-scope list.
With a Google Workspace account whose administrator has accepted Google's BAA, the Gemini app (excluding Gemini in Chrome), Gemini in Google Workspace and the Gemini Mac app are on Google's list of covered services. Gemini on a personal Google account has no BAA, and Google says human reviewers read some chats and asks users not to enter confidential information.
Anthropic signs a BAA for its first-party API, Claude in Microsoft Foundry, and HIPAA-ready Claude Enterprise plans, where the organization's Primary Owner turns on HIPAA and accepts the BAA. Anthropic says Team plans and individual plans (Free, Pro, Max and Ultra) can't enable HIPAA, and connectors, skills and beta features are not covered even on Enterprise. Disclosure: Lydoh's phone assistant runs on Anthropic's models.
OpenAI says ChatGPT for Clinicians is free for verified clinicians in the United States, currently physicians (MD/DO), nurse practitioners, physician assistants and pharmacists, verified through an NPI. It is designed for individual use with an individual BAA. It does not cover your front desk or billing staff; OpenAI points organizations that need a BAA for multiple users to ChatGPT for Healthcare.
Tell your privacy officer the same day. Under HIPAA, a use or disclosure that the Privacy Rule doesn't permit is presumed to be a breach unless the practice shows a low probability that the information was compromised, based on a risk assessment of what was disclosed, who received it, whether it was actually viewed and how the risk was mitigated. Work through that assessment with your compliance advisor; this page is not legal advice.
HIPAA's definitions are about a person's health information, so a pet's medical record is generally not protected health information, and veterinary records are usually governed by state rules instead. Client names, phone numbers and payment details still deserve care, and the same habit of keeping identifiable data out of consumer AI tools is a good one.
Keep reading
Guides, a case study and checklists for using AI in a medical, dental or veterinary practice, safely.