AI voice agents and HIPAA: what to ask a vendor.
A voice agent that answers a practice’s phone hears names, dates of birth, appointment reasons and insurance details. That is protected health information, and the vendor handling it is a business associate under HIPAA whatever the sales deck calls it. The rules are short and familiar. The questions to ask a vendor are about where the audio goes.
What HIPAA requires.
HIPAA has no clause about AI, and the existing rules already cover it. A vendor that creates, receives, maintains or transmits protected health information on a practice’s behalf is a business associate and must sign a business associate agreement before it touches a call. HHS’s cloud guidance settled the edge case: a provider that only stores encrypted data it cannot read is still a business associate. A voice agent that hears the call, transcribes it and books the appointment is well inside the line.
The Security Rule applies to the vendor directly: risk analysis, access controls, audit logs, encryption as an addressable standard today. The proposed update published in January 2025 would make encryption, multi-factor authentication and annual penetration testing mandatory; HHS has pushed the final rule to 2027, and the current rule is still enforced in the meantime. A vendor that already meets the proposed standard has nothing to change.
Minimum necessary still applies. The agent should hear what it needs to book the visit and nothing it does not, and the transcript should be kept only as long as there is a reason to keep it.
Fourteen questions, in writing.
- Will you sign our BAA, or send yours, before any call is routed?
- Where is the audio stored, for how long, and who can play it back?
- Where is the transcript stored, and is it written into our EHR or kept in your system?
- Is any call content, audio or text, used to train or tune a model, yours or a third party’s?
- Which subprocessors handle the call: the telephony carrier, the speech model, the language model, the hosting? Are they all under BAA?
- Is any of it processed or accessible outside the United States?
- Is the call encrypted in transit and at rest? With what?
- Who at your company can access our calls, and is that access logged?
- Is multi-factor authentication required for every account that can reach our data?
- What is in the audit log, and can we export it?
- How do you notify us of a breach, and within how many days?
- What happens to our data when the contract ends? Certificate of destruction?
- Have you had a third-party security assessment? Which one, and when?
- What does the agent do when a caller says something clinical, or is in distress?
Don’t put a vendor on a clinical line until it can answer all fourteen. Most of these questions apply to a live answering service too.
Consent and disclosure.
Three rules outside HIPAA reach the same phone line.
Outbound calls. The FCC ruled in February 2024 that a call made with an AI-generated voice is an artificial-voice call under the Telephone Consumer Protection Act. An agent that calls patients with reminders or results needs their prior express consent, in writing for anything that could be read as marketing, and must identify the practice and offer an opt-out. Inbound calls the patient places are not covered. Consent for reminders usually lives in the intake forms; check that it does.
Recording. Eleven states require every party’s consent to record a call: California, Delaware, Florida, Illinois, Maryland, Massachusetts, Montana, Nevada, New Hampshire, Pennsylvania and Washington, with Connecticut and Michigan as special cases. Decide whether calls are recorded at all, and if so how consent is captured, before the agent goes live.
Disclosure. California’s AB 3030, in force since January 2025, requires a generative-AI communication that contains clinical information to say it was AI-generated and how to reach a person, verbally at the start and end of an audio interaction, unless a licensed provider reviewed it first. Scheduling and billing calls are exempt. Whatever the law, an agent that says what it is at the start of the call is the right default. This is a summary, not legal advice; check your state’s rules with counsel.
What the agent should and should not do.
- Say it is an automated assistant, and how to reach a person, at the start.
- Verify the caller before reading anything back: two identifiers, every time.
- Book, reschedule, cancel, confirm, take a message, answer hours and directions.
- Not give clinical advice, not read results, not discuss medications. Anything clinical goes to a person with the context attached.
- Recognize distress and emergencies and hand off at once, with a scripted line for 911.
- Write what it did into the EHR, so the record is where the practice keeps records.
Our agents work in the practice’s own EHR and phone system. They run on our Google Cloud under a BAA, or in the practice’s own Google Cloud account if it prefers. The agent writes call content to the chart. No call trains a model, every action is logged, and we sign the BAA before any build starts. How we handle practice data.
Sources.
- Guidance on HIPAA and cloud computing. HHS Office for Civil Rights, 2016. A cloud service provider that creates, receives, maintains or transmits protected health information on a covered entity's behalf is a business associate and needs a business associate agreement, even if it only stores encrypted data it cannot read. The same logic reaches any AI vendor that handles PHI. Source
- HIPAA Security Rule to strengthen the cybersecurity of ePHI, proposed rule. HHS, Federal Register, 2025. Published January 6, 2025: would make encryption at rest and in transit, multi-factor authentication, network segmentation, annual penetration tests and an asset inventory mandatory rather than addressable. Not final; HHS's agenda moved the target to July 2027 after industry opposition. Current rule still enforced. Source
- AI-generated voices in robocalls are artificial voices under the TCPA. Federal Communications Commission, Declaratory Ruling FCC 24-17, 2024. Adopted February 8, 2024: a call that uses an AI-generated voice is an artificial or prerecorded voice call under the Telephone Consumer Protection Act, so outbound calls need the called party's prior express consent, written consent for marketing, with identification and opt-out. Inbound calls a patient places are not covered; outbound reminder calls are. Source
- Phone call recording laws by state. Rev and other vendor summaries, 2025. Eleven states are usually listed as requiring every party's consent to record a call: California, Delaware, Florida, Illinois, Maryland, Massachusetts, Montana, Nevada, New Hampshire, Pennsylvania and Washington, with Connecticut and Michigan treated as special cases; the rest need one party's. Vendor summaries agree on the core list; check the statute and the Reporters Committee for Freedom of the Press guide for your state. Source
- California AB 3030, generative AI in patient communications. California Health and Safety Code 1339.75, via Lengea Law, 2025. In force January 1, 2025 for health facilities, clinics and physician practices in California: a generative-AI communication containing clinical information must say it was AI-generated and how to reach a person, verbally at the start and end of an audio interaction, unless a licensed provider reviewed it first. Scheduling, billing and other administrative communications are exempt. Source
- Voice agent at a mental health practice. JRW case study, 2026. Booking rate for answered new-patient calls rose from 33% to 58% over six weeks at a practice with over 100 providers. One engagement. Liine's 54% across 325,000 inquiries and Patient Prism's 62% of real opportunities sit either side of it. Source
Questions.
- Is a signed BAA enough?
- You need one, and it covers only part of the risk. The BAA sets out the vendor’s duties in a contract; the fourteen questions tell you whether the vendor can meet them. Ask for the answers before the signature.
- Does HIPAA apply if the agent only schedules?
- Yes. A name and an appointment with your practice is protected health information. So is a phone number tied to a reason for the visit.
- Do patients have to consent to talk to an AI?
- For inbound calls, no federal rule requires consent, and California requires disclosure for clinical content. For outbound calls the agent makes, the FCC’s rule requires prior express consent. Disclose on every call regardless; patients who want a person should get one on request.
Put the fourteen questions to us.
Book a call and we’ll answer each one for our agent, and go through the BAA, before anything is built. Fixed price per automation. If the first one misses the scope we agreed, you don’t pay.
Book a call (opens in a new tab)Related: Voice agents for medical practices · Voice intake case study · What it costs to answer the phone · How to measure missed calls · AI vs live answering service for a law firm