General information, not legal advice — confirm with your compliance officer or attorney.
There is no such thing as a HIPAA compliant AI tool you can buy off the shelf. HIPAA governs how your practice and its vendors handle protected health information (PHI), so "HIPAA compliant AI" really means an AI tool, under a signed business associate agreement (BAA), configured correctly and used for the right tasks. This hub gives you a vendor-neutral test, a dated verdict for each popular tool, and a one-page policy your staff can follow.
It's a common question. "hipaa compliant ai" is searched about 6,600 times a month, "hipaa compliant chatbot" 5,400, "is chatgpt hipaa compliant" 880, "is claude hipaa compliant" 720, "is copilot hipaa compliant" 480, "hipaa compliant ai tools" 390 and "is gemini hipaa compliant" 320 (Google Ads Keyword Planner data via DataForSEO, US, 12-month average to Sep 2026).
Why 'HIPAA-compliant AI' is a setup, not a label
Most practices already have AI in the building. A receptionist drafts a reply in ChatGPT, the office manager's Gmail suggests whole sentences, and a note-taker joins the Monday huddle. Each of those can be fine or a reportable problem, and the difference is rarely the tool itself.
Three things decide it:
- The contract. Does the vendor sign a BAA for the exact product and plan your staff use?
- The configuration. Are training, retention, sharing and add-on features set the way the BAA assumes?
- The behavior. Do staff enter only what a task needs, and only into approved accounts?
Even the biggest vendors say so. Microsoft's HIPAA page answers "Does having a Business Associate Agreement with Microsoft ensure my organization's compliance?" with "No" (Microsoft (opens in a new tab)).
It also helps to separate two kinds of AI, because the risks differ:
- Staff productivity AI: assistants your team uses to write, summarize and take notes.
- Patient-facing AI: a chatbot on your website that talks to the public.
Five checks for any AI tool that touches patient data
Run these five checks on any tool before PHI goes near it. Each one maps to a HIPAA requirement.
| Check | What to confirm | The HIPAA hook |
|---|---|---|
| 1. BAA for the exact product and plan | Signed, names the product, covers the features you'll use | Business associates need a BAA before handling PHI |
| 2. Data-use terms | No training on your data; retention limits; deletion at the end | The BAA must limit uses and require return or destruction of PHI where feasible |
| 3. Access controls, MFA, audit logs | Practice-managed accounts, two-step sign-in, a log of who saw what | Security Rule access control, authentication and audit controls |
| 4. Encryption and data location | Encrypted in transit and at rest; you know the country | Encryption is an addressable safeguard; location affects risk |
| 5. Minimum necessary | Staff paste only what the task needs | The minimum necessary standard |
1. A BAA for the exact product and plan. HHS defines a business associate as anyone that creates, receives, maintains or transmits PHI for a covered entity, and lists cloud providers and app vendors as examples (HHS business associates (opens in a new tab)). Vendors draw coverage tightly: OpenAI covers ChatGPT for Healthcare but not ChatGPT Business, and Anthropic covers Claude Enterprise only after the HIPAA configuration is switched on. Read the plan name, not the logo. The basics are in business associate agreements.
2. Data-use terms. HHS says a BAA must stop the vendor from using or disclosing PHI beyond what the agreement allows, and must require it to return or destroy PHI at the end where feasible. Training a model on your chats is a use. Find the vendor's training and retention terms for your plan, in writing.
3. Access controls, MFA and audit logs. The Security Rule requires access control, person or entity authentication and audit controls (45 CFR 164.312 (opens in a new tab)). In practice: practice-owned accounts, multi-factor sign-in, and an admin who can see and revoke access.
4. Encryption and data location. Encryption at rest and in transit are "addressable" specifications, which HHS says is not the same as optional (HHS Security Rule summary (opens in a new tab)). A vendor that stores only encrypted PHI is still a business associate, and HHS notes that risks to PHI stored overseas "may vary greatly depending on its geographic location" (HHS cloud guidance (opens in a new tab)).
5. Minimum necessary. Covered entities must take reasonable steps to limit PHI to the minimum necessary for the purpose (HHS (opens in a new tab)). A BAA doesn't make it fine to paste a whole chart when one line would do.
The same five checks work for non-AI marketing software; see HIPAA checks for marketing software.
This is general information, not legal advice. Confirm how it applies to your practice with a qualified advisor.
Staff AI tools: ChatGPT, Claude, Gemini, Copilot and Otter
Each verdict below comes from the linked page, which was checked against the vendor's own documentation in October 2026. Plans and coverage change, so follow the link before you decide.
| Tool | BAA available for | Not covered | Full check |
|---|---|---|---|
| ChatGPT (OpenAI) | ChatGPT for Healthcare, Enterprise with Regulated Workspace, ChatGPT FedRAMP, ChatGPT for Clinicians, API with Modified Retention | Free, Plus, Pro, ChatGPT Business | Is ChatGPT HIPAA compliant: plans, OpenAI's BAA and tasks that are safe without PHI |
| Claude (Anthropic) | First-party API, Claude in Microsoft Foundry, Claude Enterprise with the HIPAA configuration | Free, Pro, Max, Team | Is Claude HIPAA compliant: covered features, files, projects and retention |
| Gemini (Google) | Gemini in Workspace and the Gemini app as a Workspace core service, under Google's BAA | Personal accounts, Gemini in Chrome, users without a qualifying license | Gemini and HIPAA: reading Google's included functionality list |
| Copilot (Microsoft) | Microsoft Copilot and Copilot Chat on commercial Microsoft 365, under the BAA in the DPA | Consumer app on personal accounts; web search queries | Microsoft Copilot and HIPAA: oversharing in Outlook, Teams and SharePoint |
| Otter.ai | HIPAA listed as an Enterprise add-on; BAA process not published | Basic, Pro, Business | Otter.ai and HIPAA: meeting bots, auto-share and recording consent |
The pattern is consistent: the free and low-cost tiers your staff are most likely to sign up for themselves are the ones without a BAA.
Patient-facing AI: website chat and chatbots
A chatbot on your website talks to the public, so it raises different questions from staff tools. HHS lists a "third-party vendor Artificial Intelligence (AI) chatbot on a provider's patient portal" that handles services such as appointment scheduling as a business associate. Two guides cover it:
- HIPAA-compliant chatbots: the buyer's guide, with the data flow from chat window to inbox and a vendor checklist.
- What medical chatbots should do: the jobs a practice bot handles well, the hard limits, and a test plan before launch.
Rules that apply to AI with patients
A few rules reach beyond HIPAA, and both of these are verified:
- Section 1557. Covered entities must make reasonable efforts to identify and mitigate discrimination risks from "patient care decision support tools" (45 CFR 92.210 (opens in a new tab)). Compliance with those duties was due within 300 days of July 5, 2024 (45 CFR 92.1 (opens in a new tab)). This targets clinical tools, not a front-desk bot.
- California AB 3030. Since January 1, 2025, a California physician's office that uses generative AI for patient messages about clinical information must add an AI disclaimer and a way to reach a human, unless a licensed provider reviews the message (Health and Safety Code 1339.75 (opens in a new tab)).
Other states are passing their own AI and bot-disclosure rules. Check yours.
What never goes into a consumer AI account
A consumer account is any personal Free, Plus, Pro or similar login, or any account your practice hasn't covered with a BAA. Keep all of this out:
- Names, dates of birth, addresses, phone numbers or emails of patients.
- Medical record numbers, insurance member IDs, claim numbers.
- Screenshots of the schedule, the EHR, a fax or a portal message.
- Intake answers, referral letters, visit notes, lab or imaging results.
- Patient photos, including before-and-after images.
- Recordings or transcripts of calls and meetings where patients are discussed.
- "Anonymized" text that hasn't been de-identified properly. HHS recognizes two methods: an expert determination, or Safe Harbor, which removes 18 identifiers and requires no actual knowledge that the person could be identified (HHS de-identification guidance (opens in a new tab)).
What's fine: service-page outlines, plain-language policy rewrites, ad headline ideas, staff memos and job posts.
An AI use policy for a small practice
One page beats a binder nobody reads. Illustrative example for a fictional four-provider practice:
Illustrative example: Riverside Family Practice AI use policy (v1, reviewed every six months)
- Approved tools: Microsoft Copilot on practice accounts (BAA via Microsoft's DPA). No other AI tool may receive patient information.
- Never enter patient identifiers, screenshots, notes or recordings into any other AI tool, including personal accounts on personal phones.
- New tools need the privacy officer's approval and a signed BAA on file before use with PHI.
- Review: a clinician reviews any AI-drafted clinical content; the practice manager reviews anything sent to patients.
- Mistakes: if PHI goes into the wrong tool, tell the privacy officer the same day so the incident can be assessed.
- Training at hire and yearly; every staff member signs this page.
For marketing uses of AI that never touch PHI, see AI marketing for practices.
How Medfy uses AI with patient data
Medfy is AI marketing for independent practices, built HIPAA-compliant by design. Where patient data is involved:
- The HIPAA AI chat widget is a $49/month add-on, covered by the BAA Medfy signs with your practice before any patient data flows (chat widget pricing).
- Patient information from forms, chat, the inbox and the CRM is stored in separate, encrypted HIPAA databases in the US or Canada.
- Ad platforms never receive patient data. Marketing uses aggregate numbers only.
- SEO articles are drafted from MedlinePlus and PubMed and approved by the practice; they don't use patient data.
Details are in Medfy's HIPAA safeguards. To see where your practice stands today, including whether tracking pixels sit on your patient forms, start with a free marketing audit.
Frequently asked questions
Is there such a thing as HIPAA-compliant AI?
Not as a label. HIPAA governs how your practice and its business associates handle patient information, so an AI tool is only part of the answer. A tool can be used compliantly when the vendor signs a BAA for that exact product, it is configured correctly, and staff enter only what each task needs. Microsoft itself says its BAA alone doesn't make you compliant. Confirm your setup with your compliance officer.
Can staff use ChatGPT for patient emails or notes?
Only inside an OpenAI product covered by a signed BAA, such as ChatGPT for Healthcare, an Enterprise Regulated Workspace or, for an individual US clinician, ChatGPT for Clinicians. Free, Plus, Pro and ChatGPT Business are not covered as of October 2026. Without a covered workspace, use ChatGPT for templates only and add patient details in your EHR. Check your plan with your compliance officer.
Does an AI vendor need to sign a BAA?
Yes, if it creates, receives, stores or transmits patient information for your practice. HHS says that holds even for a cloud vendor that only stores encrypted data it can't read. If staff use a tool strictly for tasks with no patient information, no BAA is needed, but your written policy has to make that boundary clear and enforceable. Have your attorney review any BAA.
Is de-identified data safe to put into AI tools?
Only when it meets HIPAA's standard: an expert determination, or the Safe Harbor method, which removes 18 kinds of identifiers and requires no actual knowledge that the person could be identified. Removing the name is not enough, because dates, ZIP codes, photos and rare conditions can identify someone. When in doubt, treat the text as patient information and keep it out of uncovered tools.
What should a practice's AI use policy include?
Five parts fit on one page: the approved tools and account types, the data staff must never enter, who approves new tools and keeps the BAAs on file, which outputs need clinician or manager review, and what to do if patient information goes into the wrong tool. Add training at hire, a staff signature and a review date six months out.