HIPAA-Compliant AI Assistants for a Medical Office: BAA by Plan

Few AI tools will sign a BAA for a small US medical practice, and the plan matters more than the brand. Claude signs on Enterprise only; ChatGPT on Enterprise or Healthcare through sales, the API, and individual ChatGPT for Clinicians accounts; Gemini once a Workspace super administrator accepts Google's BAA; Copilot through Microsoft's standard data protection terms. ChatGPT Business has none.

The second thing is that a BAA covers named products and features, not everything with the logo on it, and the exclusions differ by vendor. OpenAI excludes improved memory and event-triggered scheduled tasks; Anthropic excludes connectors, MCP servers, skills and Claude in Chrome on Enterprise; Google excludes third-party add-ons and Gemini in Chrome. Sign the agreement, then switch off or avoid whatever sits outside it.

Follow me on Instagram@sagnikteaches

The plan-by-plan answer, as the vendors' own pages put it

I read the vendor pages on 4 November 2026. OpenAI's own BAA help article is blunt on the plan most managers would buy: "We do not offer a BAA for ChatGPT Business." Plans and terms move fast, so treat the table as a checklist of what to confirm, not as a contract.

Connect on LinkedInSagnik Bhattacharya

The one-line version: the cheapest plan that signs a BAA is not always the one a practice can buy. Claude's Enterprise seat minimum puts it out of reach for a five-person office. Gemini and Copilot are the natural fit if you already run Workspace or Microsoft 365 and an administrator can accept the terms.

Subscribe on YouTube@sagnikteaches

What each vendor's page says, in detail

ChatGPT: three different products under one name

ChatGPT for Healthcare is described as an enterprise version for clinicians, administrators and researchers, with a BAA available. The BAA covers only the eligible products and functionality listed on the HIPAA eligible products page. Improved memory is off by default and not covered, and the help page warns that PHI should not be entered when it is on. ChatGPT for Clinicians is a separate product, free at launch to verified US physicians (MD/DO), nurse practitioners, physician assistants and pharmacists. Its help page says it is designed for individual use: the clinician signs the BAA in ChatGPT under Settings, then Agreements, and should not share PHI unless a BAA is in place and they are authorised to sign it for that account. A BAA that covers several users, with central admin controls, needs ChatGPT for Healthcare instead. If you want the longer analysis of that individual BAA, ChatGPT for Clinicians: does the individual BAA cover your practice? covers it.

For a practice, the practical reading is: individual clinicians may have a route; the office as an organisation needs Enterprise or Healthcare through sales, or a vendor whose product sits on the API with a BAA of its own. ChatGPT Business, the plan a manager would naturally buy, has no BAA, and OpenAI's own BAA article says so.

Claude: Enterprise only, with a switch

Anthropic's support article says HIPAA can be enabled in organisation settings on an Enterprise plan, and that Team and individual plans (Free, Pro, Max and Ultra) cannot enable it. Only the Primary Owner can accept the BAA: Organization settings, then Data and privacy, then HIPAA Compliance, download the BAA and Implementation Guide, accept. The change is irreversible from organisation settings and resets some settings to defaults. Features are in three groups: covered, available but not covered, and disabled. On Enterprise, chat, projects, voice, web search and research are listed as covered; connectors and MCP servers, Enterprise Search, skills and Claude in Chrome are not. Claude Code and Cowork in local mode are covered only once the Primary Owner applies a separate HIPAA configuration, which Anthropic's account team has to switch on. A BAA signed for the API before 2 December 2025 does not extend to the HIPAA-ready Enterprise plan. More in is Claude HIPAA compliant: which plans offer a BAA.

Gemini: an admin acceptance, not a plan upgrade

Google's administrator page says the BAA is accepted electronically in the Admin console by a super administrator, under Account settings, then Legal and compliance, and that administrators must accept it before using PHI in Google services. It names help me write, contextual smart replies and side-panel features as covered under Workspace with Gemini, and says third-party apps and add-ons are not covered. Google's list of covered Workspace services, updated 28 October 2026, includes Gemini in Google Workspace, the Gemini app (excluding Gemini in Google Chrome) and the Gemini Mac app, so check that list again for any newer Gemini feature before staff use it with PHI. Detail in is Gemini HIPAA compliant: what a Workspace BAA covers.

Copilot: covered by the tenant, with web search to check

Microsoft's HIPAA page says its BAA is available through the Microsoft Online Services Data Protection Addendum by default to customers who are covered entities or business associates, and lists Microsoft Copilot and Microsoft Copilot Chat among the in-scope commercial services. Its Copilot privacy page adds that prompts and responses are not used to train foundation models, and that web search generates a query sent to the Bing Search service, documented separately. For a practice, that means Copilot answers from your own tenant data sit inside your Microsoft agreement, while a web-grounded answer is a separate data path to understand before patient details go near it. Is Copilot HIPAA compliant: Microsoft 365 and web search limits goes further.

Choosing for a six-person practice

An illustrative practice has two physicians, a nurse practitioner, an office manager and two front-desk staff. They already use Google Workspace for email. What to pick depends on the job, not the brand.

1. Drafting patient letters and recall messages with patient details. Needs a covered tool. Gemini in the Workspace account works after the administrator accepts the BAA and checks the covered list.
2. Clinical research and documentation by the physicians. ChatGPT for Clinicians for each verified clinician is an option; tell the practice which clinician holds which agreement.
3. Policy documents, job adverts, social posts, the website. Any business plan, no patient information. ChatGPT Business or Claude Team is fine here.
4. Phone answering. A separate product with its own BAA. See the checklist in a BAA and retention checklist for an AI receptionist.

This keeps the number of covered tools small. The fewer places patient information can go, the shorter the risk assessment.

Four jobs, sorted by whether patient information is involved

The test is whether the text you paste, taken with the date and your practice, could identify a patient. If yes, it goes to a covered tool; if no, any approved business plan will do. Keep the sorted list on the wall by the desks.

An email that gets a straight answer from a vendor

Subject: BAA for [product and plan] at a small medical practice

Hello,

We are a [number]-person practice and would like to use [product and plan] with patient information. Please confirm in writing:

1. Will you sign a BAA for this exact plan? If not, which plan will you sign for?
2. Which features are covered by the BAA and which are not (memory, web search, plug-ins, agents, file uploads, voice)?
3. Are chats stored, for how long, and can we set retention and deletion?
4. Who are your subprocessors for this plan?
5. Is the BAA with the practice, or with the individual user who signs up?
6. What is the seat minimum and the billing term?

Thank you,
[your name], [role], [practice]

Question 5 catches the individual-account problem, and question 6 catches Claude's seat minimum before you plan around it.

A tracker that shows what is covered, by whom

Keep one row per tool in a shared sheet. A filled-in example for the practice above (illustrative):

The sheet answers the question a regulator or insurer would ask first: which tools touch patient information, and what is the paper for each. It also tells a new employee in one glance which tool is safe for what.

What to read inside the BAA before you accept it

Accepting a BAA is a click in some products, so it is easy to skip reading it. Five clauses are worth five minutes each time.

• Scope. Which products and functions are covered. OpenAI ties its BAA to a list of eligible products and functionality; Anthropic sorts features into covered, available but not covered, and disabled.
• Permitted uses. Whether the vendor may use your information for anything other than providing the service. A training exclusion in the privacy settings is not a substitute.
• Subcontractors. Whether the vendor flows the same obligations down to the companies that process your data.
• Breach notice. How quickly the vendor tells you of a security incident and what it tells you.
• Return or deletion at the end. What happens to stored conversations when you leave.

Save the signed version, the date and the person who accepted it. If the product changes and a feature moves outside the covered list, the version you saved is your evidence of what you agreed to.

A cost sum for a six-person office, using list prices

Price drives some of this choice, so run the sum before you shortlist. Illustrative, from the list prices quoted on the Claude pricing page at the time of reading. Claude Team Standard is $20 a seat a month billed annually for 2 to 150 seats, so six seats cost $120 a month, but Team cannot enable HIPAA. Self-serve Enterprise is listed at $20 a seat a month, billed annually, with a 20-seat minimum and usage billed separately at API rates, so the floor is 20 seats at $20, which is $400 a month or $4,800 a year before usage. A six-person practice would pay for 14 seats nobody uses, unless the practice has other staff or partners who would also use it.

Set that against the alternative: if you already pay for Google Workspace, accepting the BAA in the Admin console is a settings change, not a new subscription. For many small practices the cheapest covered route is the suite they already run, which is why the question "which plan" often becomes "which of the tools we already have is covered".

When the right tool has no BAA: two honest options

Sometimes the tool a clinician likes has no BAA for their plan. There are only two honest routes: use it for work with no patient information, or de-identify the text properly first. De-identification is stricter than deleting a name. The privacy rule's safe-harbour method removes a defined list of identifier types, including dates tied to the person, locations smaller than a state, phone numbers, record numbers and anything else that could identify them. If you cannot be sure you have removed everything, treat the text as patient information.

A before and after, illustrative.

Before (do not paste into an uncovered tool):
"[patient's full name], DOB [day, month and year], seen [visit date] at our [street name] clinic for a persistent cough and a medication review. Phone [mobile number]. Please draft a recall letter."

After (generic request, no patient information):
"Draft a friendly recall letter template for an adult patient who needs a follow-up appointment after a medication review. Leave placeholders for the name, date and phone number."

The second version asks for the template, and the clinician fills in the placeholders inside the practice's own record system. That route works for letters, scripts and checklists. It does not work for summarising a real note, which is the case where you need a covered tool.

A five-minute staff check after you choose

Once the tracker exists, test it on the people who will use it. Give each person three situations and ask which tool they would use, or whether they would use none. One: a template for appointment reminders. Two: a summary of a patient's visit note. Three: a request from a patient to explain a test result in plain words. The right answers are any approved tool, the covered tool, and the clinician, not an assistant. Anyone who gets one wrong gets a ten-minute conversation, not a reprimand, and the rule is repeated at the next team meeting.

Repeat the test when you change a plan or add a tool. A short test every quarter catches drift earlier than an annual training day.

Settings to check after the BAA is signed

The agreement is half the job. The settings decide what staff can do with it day to day. For whichever tool you choose, go through four settings with the administrator and write down the answers.

1. Who has an account. Remove leavers, give every user their own login, and switch on single sign-on or two-step verification if the plan offers it.
2. Memory and history. Decide whether the tool may remember anything between chats, and for how long chats are kept. OpenAI lists improved memory as not covered, and Anthropic says enabling HIPAA resets certain organisation settings to defaults, so check both settings again after you sign.
3. Connected apps and plug-ins. Every connection is a new place for patient information to travel. Allow only those the vendor lists as covered.
4. Sharing. Check whether chats, projects or files can be shared by link, and turn off public sharing.

An illustrative result for the six-person practice: single sign-on on, memory off, no plug-ins, link sharing off, chat history limited to the shortest period the plan allows. Those five lines go at the top of the tracker, with the date they were checked.

How patient data ends up on a plan with no BAA

A realistic one: the office manager buys a business-plan seat of a well-known assistant for the practice because it says "no training on your data". Staff start pasting patient letters. Nothing happens for months. Then an audit or a new hire asks for the BAA, and there is none. The fix is a rule that "no training on your data" is a privacy setting, not a BAA, and that the tracker is the only list of tools allowed with patient information.

Another: a clinician signs an individual BAA and starts sharing the account login with the front desk. That breaks the individual terms and puts uncovered users on a covered account. Give each person their own route.

A third: Gemini is enabled for the whole practice, the BAA is accepted, and then an administrator installs a third-party add-on that reads email. The add-on is outside the BAA. Review add-ons whenever the covered list changes.

Turning the table into a policy staff will follow

One page is enough: the approved tools and what each may touch, the rule against pasting patient information anywhere else, who to ask, and what to do after a mistake. A template and the surrounding rules are in an AI usage policy for a medical practice: template and HIPAA rules. After the policy, run a ten-minute test: ask each staff member which tool they would use to rewrite a referral letter, and fix any wrong answers.

This tutorial is practical guidance, not legal advice; confirm each answer with the vendor in writing and with your compliance adviser. If you want help sorting your own tools, that is what my AI implementation consultation covers. The AI guide collects the rest of the series.

Questions practice managers ask about AI and BAAs

Does a BAA make an AI tool HIPAA compliant?

No. A BAA is the contract that lets a vendor handle patient information for you. Compliance also needs the right plan and settings, covered features only, staff rules, access controls and a risk assessment. A tool can have a BAA and still be used in a way that breaks the rules.

Can we use a free chatbot if we remove patient names?

Only if the text is properly de-identified, which means removing all the identifiers the privacy rule lists, not just names. Dates, locations, record numbers and small details can still identify someone. When in doubt treat it as patient information and use a covered tool.

Our clinician signed a BAA on an individual account. Does that cover the practice?

Not automatically. An individual agreement covers that account's use, and the practice as a covered entity needs to know who the contracting party is, which staff can use it and whether front-office work falls inside it. Ask the vendor in writing before moving any practice workflow onto it.

What if the vendor says a feature is not covered?

Turn it off or never put patient information through it. Vendors list covered and uncovered features separately, and memory, connectors, browser extensions and third-party add-ons are common exclusions, though each vendor draws the line differently. Record the exclusion in your tracker and brief staff.

Further reads

• Is ChatGPT HIPAA Compliant? Free, Business and Clinician Plans — The ChatGPT plan-by-plan view for a small practice.
• Is an AI Scribe HIPAA Compliant? A BAA and Consent Checklist Before Go-Live — A BAA and consent checklist for scribes before go-live.
• ChatGPT for Chiropractors: What Is Safe Without a BAA? — What a practice can do with ChatGPT where no BAA exists.
• Best HIPAA-Compliant AI for Therapists: Which Vendors Sign a BAA? — The same BAA question for therapists.
• AI Scribe Security: Audio, Retention and BAA Vendor Questions — Audio, retention and BAA questions for scribe vendors.
• What to Check in an AI Tool's Privacy Policy and Terms — How to read an AI tool's privacy terms.
• AI Solutions for Medical Practices: Build, Buy or Hire Help? — A decision rule for build, buy or hire: by practice size and data sensitivity, with the cost and time of each route and when a consultation pays for itself.
• Cliniko for Doctors: Fit, Cost and the AI Tools That Sit Around It — Cliniko is built for allied health. What that means for a doctor's practice, what it costs, and the AI scribes and phone tools that connect to it.
• Medical AI Chatbot for Providers: Clinical Search or Patient FAQ? — Clinical search tools answer clinicians' questions with citations; FAQ bots answer patients from a fact sheet. How to pick between them.
• Medical Practice AI Budget: What to Fund First and What It Costs — A budget order for a small practice's first AI spend: a decision rule by job, cost bands per tool class from vendor pages, and payback inputs you can measure.
• Midwife Software Compared: Notes, Booking, Billing and AI Tools — Five midwife software options checked on vendor pages for notes, consent, portal, billing, data location and AI claims, with a scoring sheet.
• AI App for Medical Providers: Standalone, Add-On or Built-In? — Standalone app, add-on or built-in feature? A vendor-neutral look for small medical practices on cost, data flow, switching and fit, with a decision table.
• AI Assistant for a Medical Practice: Build a Staff Policy Q&A — Turn your protocols into a staff Q&A assistant in an afternoon, with a 20-question test set and no patient data anywhere near it.
• AI Tools for Medical Practices: A Stack for Three Doctors and Four Staff — The first three AI tools for a seven-person practice, scribe, phones and billing checks, with list prices, a monthly total and what to check first.
• AI Tools for Nurses Who Run a Team: Rotas, Handover and Audits — A nurse manager's guide to AI for rotas, handover templates and audit write-ups, with a clear list of what must never go into a prompt.
• AI Tools for Physicians: Seven Worth Paying For in Private Practice — Seven AI tools for a private-practice physician, scribes, clinical assistants, phones and claim checks, with what each does and what to verify.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Anthropic support articles on HIPAA-ready Enterprise plans and the HIPAA configuration, Anthropic's BAA article and claude.com/pricing; OpenAI Help Center articles on the API BAA, ChatGPT for Clinicians, ChatGPT for Healthcare and HIPAA eligible functionality; Google Workspace HIPAA administrator guidance and covered-services list; Microsoft Learn pages on its HIPAA offering and Copilot privacy; the safe-harbour rule at 45 CFR 164.514 (all read 4 November 2026).