Searches for “HIPAA-compliant chatbot” have taken off in 2026 as practices add AI chat to their websites, patient portals and text lines. Most guides stop at “make sure the vendor signs a BAA.” That’s necessary, but a patient-facing chatbot usually involves four or five companies, plus tracking scripts most practices forget are on the page. This guide covers what actually has to be in place.
Does your chatbot handle PHI?
Start here, because it changes everything else. Protected health information is individually identifiable health information. A chatbot handles PHI when it connects a person to their health or care, for example:
- Booking or changing an appointment for a named patient
- Asking about symptoms, conditions or medications
- Collecting insurance or date-of-birth details
- Answering questions about a specific patient’s bill, results or prescriptions
A chatbot that only answers general questions (hours, directions, which insurance you accept, how to prepare for a procedure) and never asks who the visitor is generally isn’t handling PHI. Many practices start there for that reason. The moment the bot books appointments or looks anything up, it’s in HIPAA territory.
The BAA chain: every company that touches the chat
Under HIPAA, any company that receives PHI on your behalf is a business associate, and that includes subcontractors of your vendors. For an AI chatbot, that chain usually looks like this:
| Link in the chain | Example | Needs a BAA? |
|---|---|---|
| Chat platform or widget | The chatbot product or your custom app | Yes |
| AI model provider | OpenAI API, Anthropic API, or a model on Amazon Bedrock or Azure | Yes |
| Hosting and database | AWS, Azure, Google Cloud | Yes |
| SMS or voice provider | Texting and phone carriers, speech-to-text and text-to-speech services | Yes, if they carry PHI |
| Logging and monitoring | Error tracking and AI observability tools that store prompts | Yes, if they store conversation content |
| EHR or scheduling system | Where the bot books or reads data | Covered by your existing agreement |
A packaged chatbot vendor should handle its own subcontractors and say so in its BAA. Voice platforms make the point clearly: Vapi’s HIPAA documentation requires that every downstream language model, voice and transcription provider is HIPAA compliant before the account can run in HIPAA mode. If you build your own, you’re the one assembling and checking the chain.
On the model side, the major options as of November 2026: OpenAI signs a BAA for the API with Modified Retention enabled; Anthropic signs one for the Claude API; and models on Amazon Bedrock fall under the AWS BAA. More detail is in our guide to HIPAA-compliant AI.
The part most practices miss: tracking scripts on the page
A chatbot usually sits on a website that also runs analytics, ad pixels and session-recording tools. If those scripts can see what a patient types, or the page they were on when they started a chat about a condition, PHI can leak to companies that will never sign a BAA.
HHS issued guidance on online tracking technologies in December 2022 and updated it in March 2024. A federal court in Texas vacated part of it in June 2024 (the part treating an IP address plus a visit to a public, unauthenticated health page as PHI), and HHS dropped its appeal that August. The guidance on authenticated pages wasn’t vacated, so the core point still stands: tracking on authenticated pages like patient portals, and any tracking that captures health information a patient enters, is a HIPAA problem. Before launching a chatbot:
- Keep the chat widget off pages that carry ad pixels, or load the widget in a way those scripts can’t read.
- Turn off session recording on any page with the chat or forms.
- Make sure chat events sent to analytics carry no message content or identifiers.
What the chatbot itself needs to do
Beyond agreements, a patient-facing bot needs some design choices that general chatbots don’t:
- Verify identity before sharing anything. A bot should never reveal appointment details, balances or results to someone who just typed a name. Use portal login, a one-time code or a callback.
- Collect only what the task needs. Minimum necessary applies to bots too. Booking a cleaning doesn’t need a medical history.
- Hand off to a person well. Clinical questions, emergencies and upset patients should go to staff, with the transcript attached. Emergency language should trigger an immediate “call 911” response.
- Tell people it’s AI. Several states now require it, and it builds trust.
- Log access to transcripts. You should be able to see who viewed or exported a conversation.
- Set a retention period. Decide how long transcripts are kept, and where the official record lives (your EHR, not the chatbot).
- Don’t let the bot give medical advice it isn’t designed for. Texas’s attorney general settled with an AI vendor in 2024 over claims about the accuracy of its generative AI in hospitals. Accuracy claims are a legal risk too.
Buy vs build
Buy a healthcare chatbot that signs a BAA if you want something running quickly and your needs are standard: FAQs, appointment requests and simple intake. Check which plan includes the BAA (it’s often only the higher tiers), whether the vendor reserves the right to use de-identified data, which EHR or scheduling systems it actually writes into, and who its AI subcontractors are.
Build on a HIPAA-eligible cloud if the bot has to work with your own systems in non-standard ways, if you want patient data to stay inside an account you own, or if you already pay for the cloud. A typical build runs a model on Amazon Bedrock or Azure under your existing BAA, stores conversations in your own database, and connects to your scheduling and EHR through their APIs. It costs more up front and less per month, and you control the whole chain.
Middle ground: a voice or chat platform that signs a BAA (Retell offers a self-serve BAA on pay-as-you-go accounts; Vapi sells HIPAA as an add-on), connected to your systems by a developer or integrator.
Building and connecting these systems is what we do. We set up chat and voice agents inside a client’s own cloud account and wire them into scheduling, EHR and messaging tools so patients get answers and staff get clean handoffs. See AI customer service automation and AI agent development, or start with the AI audit.
Launch checklist
- Decide whether the bot will handle PHI at all. If not, design it so it never asks.
- List every vendor in the chain and collect a signed BAA from each.
- Remove or isolate tracking pixels and session recording on chat pages.
- Add identity verification before the bot reveals anything patient-specific.
- Set up human handoff for clinical, emergency and complaint conversations.
- Add an AI disclosure to the greeting.
- Set transcript retention and access logging.
- Add the chatbot to your HIPAA risk analysis.
- Test it with real scenarios, including someone trying to get another patient’s information.
This guide is general information, not legal advice.
Sources
- 45 CFR 160.103 (business associate and subcontractor definitions): https://www.law.cornell.edu/cfr/text/45/160.103
- HHS guidance on HIPAA and online tracking technologies: https://www.hhs.gov/hipaa/for-professionals/privacy/guidance/hipaa-online-tracking/index.html
- American Hospital Association v. Becerra (N.D. Tex., June 20, 2024), Hunton summary: https://www.hunton.com/privacy-and-cybersecurity-law-blog/federal-judge-rules-that-online-tracking-technologies-guidance-exceeded-hhs-authority
- OCR withdraws its appeal (August 2024), Quarles: https://www.quarles.com/newsroom/publications/hhs-ocr-withdraws-tracking-technologies-appeal-in-aha-v-becerra
- OpenAI HIPAA Implementation and Configuration Guide: https://cdn.openai.com/osa/baa-hipaa-guide.pdf
- Claude API documentation, HIPAA readiness: https://platform.claude.com/docs/en/manage-claude/api-and-data-retention
- AWS HIPAA eligible services reference: https://aws.amazon.com/compliance/hipaa-eligible-services-reference/
- Vapi HIPAA documentation: https://docs.vapi.ai/security-and-privacy/hipaa
- Retell compliance documentation: https://docs.retellai.com/general/compliance
- Texas Attorney General settlement with Pieces Technologies: https://www.texasattorneygeneral.gov/news/releases/attorney-general-ken-paxton-reaches-settlement-first-its-kind-healthcare-generative-ai-investigation