AI Agents

HIPAA-Compliant AI Chatbots: What It Takes and Who Signs a BAA

An empty, bright clinic waiting room with blue chairs and a reception counter A signed BAA is only the start.

A HIPAA-compliant chatbot is not a product you can buy. HIPAA applies to organisations, so the question is whether your use of a chatbot meets the rules. That depends on what the bot sees, who else handles the data and what safeguards sit around it. If a chatbot handles protected health information for a covered entity, every vendor in the chain needs a signed Business Associate Agreement, called a BAA.

Several AI providers will sign one, on some products and not others. This article explains the key terms in plain English, shows what each major provider says on its own pages, lists the safeguards you still have to build, and covers what Florida adds for Florida practices. It is written by a software studio that builds AI agents, not by a law firm. It is not legal advice, so take it to your compliance lead or counsel before you build.

Key takeaways

  • No chatbot is “HIPAA compliant” on its own. Your organisation’s use of it is, which means the contracts, the safeguards and the data flow all count.
  • If the bot touches patient data, the model provider, the hosting and the chatbot platform each need a BAA. A consumer chat app is the wrong tool.
  • Providers differ by product. OpenAI, Anthropic, Microsoft, Amazon and Google all have a route to a BAA, each with conditions, and none of them makes your bot compliant by signing it.
  • Identity checks, audit logs, no patient data in analytics tools and a hand-off to a person matter as much as the contract.
  • The safest first chatbot answers opening hours, directions and general questions and never touches patient records.

Is a chatbot HIPAA compliant on its own?

No. The rules attach to covered entities and their business associates, so a chatbot becomes a HIPAA matter when a covered entity uses it with protected health information. The federal regulation defines a covered entity as a health plan, a health care clearinghouse, or a health care provider that transmits health information electronically in connection with certain transactions (45 CFR 160.103).

A practice using a chatbot to book appointments is in that position if it falls into one of those groups. A wellness app that is none of them may not be covered by HIPAA at all, though other privacy laws can still apply. Your counsel can tell you which side you are on.

Marketing pages for chatbot products often say “HIPAA-compliant” as if it were a feature. The regulation sections quoted in this article describe duties for the organisation and for each business associate, not a label that software can carry.

HIPAA terms in plain English

Five terms cover most of the conversation, all taken from the federal regulation.

  • Protected health information (PHI) is individually identifiable health information that is transmitted or maintained in any form, with a few exclusions such as employment records held by an employer (45 CFR 160.103). A name next to a symptom is PHI. An anonymous question about opening hours is not.
  • Covered entity is the health plan, clearinghouse or provider described above.
  • Business associate is a person who, on behalf of a covered entity, “creates, receives, maintains, or transmits protected health information” for a regulated function. The definition also includes a subcontractor that does the same on behalf of a business associate (45 CFR 160.103).
  • Business Associate Agreement (BAA) is the contract. A covered entity may let a business associate handle PHI only if it obtains satisfactory assurance that the information will be safeguarded (45 CFR 164.502(e)). The contract must set out what the business associate may do with the data and require appropriate safeguards and reporting of any improper use (45 CFR 164.504(e)).
  • Minimum necessary means making reasonable efforts to limit PHI to the minimum necessary to accomplish the purpose (45 CFR 164.502(b)). For a chatbot, that is a design rule: do not pull a full record into a conversation that only needs an appointment time.

For a chatbot, the covered entity is usually the practice, the chatbot platform is a business associate, and the company that runs the language model is a subcontractor business associate. Each link in that chain needs its own agreement. Hosting counts too: HHS says a cloud provider that stores electronic PHI is a business associate even if it holds only encrypted data and has no key (HHS guidance on HIPAA and cloud computing).

Which AI providers will sign a BAA?

All five providers buyers ask about have a route to a BAA, and none of them offers it on every product. The table is a summary of what each vendor’s own pages said when we checked on 17 November 2026. Vendor terms change, so confirm in writing before you send patient data anywhere.

ProviderProductBAA available?
OpenAIAPIYes, with conditions
OpenAIChatGPT Enterprise, EduYes, via sales
OpenAIChatGPT BusinessNo
AnthropicClaude APIYes, with conditions
AnthropicClaude EnterpriseYes, owner accepts
AnthropicTeam, Pro, MaxCannot enable HIPAA
MicrosoftAzure servicesYes, via terms
AmazonBedrockYes, AWS BAA
Google CloudAgent PlatformYes, Google BAA

The detail behind each row matters more than the row.

OpenAI. Its help centre says that to use the API with PHI, “your organization must first enter into a Business Associate Agreement (BAA) with OpenAI”. It adds that an enterprise agreement is not required, and that self-serve enrolment requires an established history of API usage (OpenAI Help Center). A separate page says the API needs an executed BAA and “Modified Retention”, unless OpenAI specifies otherwise (OpenAI HIPAA eligible products). For ChatGPT, a BAA for a sales-managed Enterprise or Edu account goes through sales, and the same page states: “We do not offer a BAA for ChatGPT Business.” Some features, such as event-triggered scheduled tasks, are listed as not covered.

Anthropic. It provides a BAA for its HIPAA-ready services, such as the first-party API, Claude in Microsoft Foundry and Enterprise plans (Anthropic Privacy Center). For the API, the organisation’s Primary Owner signs and then contacts Anthropic’s sales team to have it turned on. On Enterprise, the Primary Owner accepts the BAA in the organisation settings. Anthropic’s support page says Team plans and individual plans (Free, Pro, Max and Ultra) “can’t enable HIPAA” (Claude Help Center). Features such as connectors to third-party tools are marked as available but not covered.

Microsoft. Microsoft’s HIPAA page says there is no separate contract to sign. The BAA comes through its Product Terms and Data Protection Addendum, by default, to customers that are covered entities or business associates, “for use of such in-scope Azure services” (Microsoft Learn). That page does not name Azure OpenAI. The list of in-scope services sits in a separate compliance workbook, so ask Microsoft to confirm in writing that the specific model service and region you plan to use are covered.

Amazon. Amazon Bedrock is on AWS’s HIPAA Eligible Services list, and AWS says customers must not use those services with PHI “without first entering into an AWS business associate agreement” (AWS HIPAA Eligible Services). The page also says customers still must configure these services consistently with HIPAA.

Google Cloud. Google’s page says it “enters into Business Associate Agreements as necessary under HIPAA” and lists “Gemini Enterprise Agent Platform” and “Generative AI on Agent Platform” among covered products. It does not use the older name Vertex AI, so confirm in writing that the product you mean is the listed one. It also says not to use pre-GA offerings with PHI unless a notice says otherwise (Google Cloud).

Is ChatGPT or Claude HIPAA compliant?

Neither is, in the sense the search question usually means. The consumer apps and the smaller business plans are the wrong place for patient data: OpenAI offers no BAA for ChatGPT Business, and Anthropic says its Team and individual plans cannot enable HIPAA. Both companies do sign agreements for their API and enterprise products, with the conditions above.

OpenAI’s own help page makes the second point plainly: accepting a BAA and enabling HIPAA support “do not, by themselves, make your application HIPAA compliant.” Microsoft’s page describes the same split, listing each HIPAA control as the customer’s, Microsoft’s or shared. The vendor covers its side. You still own your side: how the chatbot is built, who can reach it and what it stores.

If you are choosing between the two model families for a non-clinical job, our ChatGPT vs Claude comparison covers the differences. For patient data, the contract and the configuration usually decide the choice before the model quality does.

What a compliant chatbot needs beyond the BAA

A signed agreement is one input. The Security Rule asks the organisation to build safeguards around the system, and the eCFR text gives a clear checklist.

  • A risk analysis. The regulation requires “an accurate and thorough assessment of the potential risks and vulnerabilities” to the confidentiality, integrity and availability of electronic PHI (45 CFR 164.308). Do this before the build, on the actual data flow.
  • Access control and unique identification. Access is limited to people or software programs granted access rights, and each user has a unique name or number (45 CFR 164.312).
  • Person authentication. The same section requires procedures to verify that a person seeking access to electronic PHI is the one claimed. For a chatbot, that means verifying who is typing before the bot discusses anyone’s details. A bot that answers anyone who types a name has failed this one, whatever its contract says.
  • Audit controls. Record and examine activity in systems that contain or use electronic PHI. Also review those records regularly (45 CFR 164.308).
  • Encryption and transmission security. Encryption of stored PHI is listed as “addressable”, and the text calls for encrypting transmitted PHI “whenever deemed appropriate”. Treat both as decisions to make deliberately and write down (45 CFR 164.312).
  • Sanctions and trained staff. Apply sanctions to workforce members who break the policies. Staff who configure or supervise the bot need to know what it may and may not do.

Some practical rules follow from those. Keep PHI out of general logs, analytics tags and error-reporting tools, because any tool that receives it becomes part of the chain and needs its own agreement. Set a retention period, then delete on schedule. Give patients a way to reach a person, and hand over cleanly when the bot cannot answer safely.

If the bot retrieves answers from a document store, the store holds PHI too. Our explanation of retrieval-augmented generation shows where that data sits and why access rules must apply at retrieval time. For the customer-facing side of the same idea, see AI for customer service.

Low-risk and high-risk chatbot uses

The easiest way to stay inside the rules is to decide up front what the bot is allowed to touch. Keep the first release on the left of this table.

Lower riskHigher risk
Opening hoursSymptoms
Directions, parkingTest results
Services offeredMedical records
General FAQsBilling, insurance
Booking, no health detailAnything tied to a name

A bot on the left has no PHI to protect, provided it never asks for it. Add a clear instruction and a filter so that a patient who starts typing symptoms is told to call the practice. The right-hand column is where the full set of safeguards applies. It can still be done, and some practices need it, but it should follow a risk analysis and not come first. Our health and wellness industry page shows the wellbeing and fitness products we have built, none of them a HIPAA project.

What Florida adds for Florida practices

Two Florida statutes sit beside HIPAA. This section describes what they say. Whether they apply to you is a question for counsel.

The Florida Information Protection Act. In this statute, a “covered entity” is any commercial entity that acquires, maintains, stores or uses personal information, which is not the same as HIPAA’s term. Under section 501.171, it must give notice to each affected individual in Florida “as expeditiously as practicable and without unreasonable delay”. The deadline is no later than 30 days after the determination of a breach or reason to believe one occurred, unless a delay is authorised (Fla. Stat. 501.171(4)). A breach affecting 500 or more individuals in Florida must also be reported to the Department of Legal Affairs within the same 30 days. Personal information under the statute includes information about a person’s medical history, condition, treatment or diagnosis.

There is a link to HIPAA. Notice given under the rules of your primary federal regulator counts as compliant for the individual-notice requirement. For the notice to the Department, the statute says a covered entity that “timely provides a copy of such notice to the department” is deemed compliant. In plain terms: following HIPAA’s breach rules does not remove the Florida step of sending the Department a copy.

Storage of patient information. Section 408.051(3) applies to a health care provider that uses certified electronic health record technology. The provider must ensure all patient information stored in an offsite physical or virtual environment, including through a cloud computing service, is physically kept in the continental United States, its territories or Canada (Fla. Stat. 408.051(3)). For a chatbot, that is a hosting question. Check the region of every service that holds patient information, including the model provider and any backups.

What happens if there is a breach

The HIPAA breach rules put timers on both sides. A covered entity must notify affected individuals “without unreasonable delay and in no case later than 60 calendar days after discovery” (45 CFR 164.404). For breaches affecting 500 or more individuals, it must notify the Secretary of Health and Human Services at the same time, and it logs smaller ones for an annual report (45 CFR 164.408). A business associate must notify the covered entity within the same outer limit of 60 calendar days (45 CFR 164.410).

In practice, the Florida 30-day limit is the tighter clock for Florida residents. Agree with each vendor in the BAA how fast they will tell you, so you are not learning of an incident on day 50.

Where we would start

We are CodeLeap, a UK software, AI and automation studio. We have no US healthcare client and have not delivered a HIPAA project, so nothing here is a claim of that experience. What we do is scope and build AI agents with the data flow designed first, and that is the right order for a patient-facing bot.

A sensible first step looks like this:

  1. List every place a conversation could touch patient data, from the chat window to logs and backups.
  2. Decide which questions the bot may answer without PHI, and build only those first.
  3. For anything that touches PHI, choose a provider route from the table above and get the BAA and the region confirmed in writing.
  4. Add identity checks, audit logs and a hand-off to a person before launch.

Our AI agents service starts with a first agent from £5,000: one named job, scoped, with clear limits on what it decides alone and human sign-off where it matters. If the real question is whether and where to use AI at all, AI strategy consulting comes first. If the data will live in your own cloud account, hosting, monitoring and backups are the work of managed cloud. Your compliance lead or counsel decides what HIPAA and Florida law require of you. We design the build to fit what they decide.

Frequently asked questions

Is ChatGPT HIPAA compliant?

ChatGPT is not compliant by itself, and the consumer app is not the product to use. OpenAI says it signs a Business Associate Agreement for its API, and for ChatGPT Enterprise or Edu through sales, but not for ChatGPT Business. Even with a signed agreement, your own configuration, access controls and staff habits decide whether your use meets HIPAA. Confirm the current terms with OpenAI in writing.

Is Claude HIPAA compliant?

Anthropic offers a Business Associate Agreement for its HIPAA-ready services: the Claude API, once the organisation's Primary Owner signs and Anthropic's team enables it, and Claude Enterprise plans. Team plans and the individual plans (Free, Pro, Max and Ultra) cannot enable HIPAA. Not every feature is covered. Signing the agreement does not make your chatbot compliant; your own safeguards still apply.

What makes a chatbot HIPAA compliant?

The chatbot does not carry the label; your use of it does. If it handles protected health information for a covered entity, every vendor that touches that data needs a Business Associate Agreement. The Security Rule safeguards also apply: access controls, audit logs, person authentication, encryption where appropriate, a risk analysis and trained staff. A person should be reachable when the bot cannot answer safely.

Do I need a BAA for an AI chatbot?

You need one if the chatbot, or any service behind it, creates, receives, maintains or transmits protected health information on your behalf. That includes the model provider, the hosting and the chatbot platform, because each handles the data. A chatbot that only answers opening hours and directions and never touches patient data can avoid the question, as long as it genuinely stays that way.

Can a chatbot ask patients about symptoms?

It can, but the moment a patient types symptoms that are linked to who they are, the chat holds health information. The high-risk rules then apply: a signed agreement with each vendor, verified identity, no copies in logs or analytics tools, and a route to a clinician. A safer path is to start with low-risk questions and add symptom intake only after the safeguards are in place.

Planning an AI chatbot for patients?

Book a free 30-minute consultation and we'll map where patient data would flow before anything is built.

Book a consultation Get a free quote

Keep exploring

Back to all insights

Book a free call with Vivian

COO and co-founder · London

Request a callback

Leave your name and number and we'll call you back.