Key takeaways
- If an AI phone agent creates, receives, stores or sends protected health information for you, the vendor is a business associate and needs a BAA.
- Map every place call data goes: audio, transcripts, summaries, integrations and the vendor's own subcontractors.
- Ask directly whether call data trains any model, how long each data type is kept and how it is deleted when you leave.
- The safest design for many practices is an agent that avoids collecting health details and hands patient-specific calls to staff.
When HIPAA applies to an AI phone agent
HIPAA does not regulate a technology; it regulates what happens to protected health information. An AI phone agent that answers a caller's question about parking handles no health information. The same agent hearing a patient's name, date of birth and reason for a refill does. Once a vendor creates, receives, maintains or transmits protected health information on your behalf, it is a business associate. A covered entity may let a business associate handle that information only with satisfactory assurance, in a business associate agreement, that the vendor will safeguard it[1].
That gives practices two legitimate paths. One is a vendor that signs a business associate agreement and has the safeguards to match. The other is an agent configured not to collect health information at all, which answers general questions and hands every patient-specific call to staff. Many practices start with the second and move to the first only when a specific task needs it.
Concern is high, rules are thin
Patients and physicians are both wary. 77% of the US public say they are concerned about the privacy of personal medical information provided to AI tools[2]. About 40% of physicians say they are equally excited and concerned about AI, naming patient privacy among their top concerns[3]. Yet most practices have no written rules: in January 2026, only 20% of medical groups had AI governance or a formal AI-use policy, and 56% had none[4].
Map the data flow first
Before comparing vendors, ask each one to draw where call data goes. You are looking for every place a copy exists, who can see it and how long it stays.
Audio
Is the call recorded? Where is the audio stored, and for how long? Can recording be switched off?
Transcription
Which service turns speech into text, and does that service keep a copy?
Language processing
Which systems process the text to understand and reply? Are they run by the vendor or by subcontractors?
Outputs
Where do transcripts and summaries go: a vendor dashboard, email, your EHR, a messaging tool?
Integrations
Does the agent read from or write to your PM or EHR, and with which permissions?
Deletion
What is deleted, when, and how is deletion confirmed, during the contract and after it ends?
The questions to ask
| Area | Question | Red flag |
|---|---|---|
| Agreement | Will you sign our business associate agreement, or yours, before any patient call? | We are HIPAA compliant, so you do not need one |
| Subcontractors | Which subcontractors handle call audio or text, and do they sign agreements with you? | Cannot list them |
| Model training | Is any of our call data used to train or improve models, yours or anyone else's? | Only anonymized data, without saying how |
| Retention | How long are audio, transcripts and summaries kept, and can we set shorter periods? | Kept indefinitely by default |
| Access | Who at your company can see our calls, and is that access logged? | Support staff can see everything |
| Staff access | Do our staff get individual logins with multifactor authentication and role limits? | One shared login |
| Audit logs | Can we see who viewed or exported each call record? | No user-level logs |
| Encryption | Is data encrypted in transit and at rest? | Vague or marketing-only answer |
| Incidents | How quickly will you tell us about a security incident involving our data? | No commitment in the contract |
| Exit | What do we get back when we leave, and how do you confirm deletion? | Data deleted on request only, no confirmation |
Model training deserves its own question
AI vendors often improve their systems using customer data. For a medical practice, that is a business associate using protected health information for its own purposes, which your agreement should either forbid or tightly define. Ask the question in plain words and get the answer in the contract: is our call audio, transcript or summary data used to train, fine-tune or evaluate any model, and if so, which data, after what de-identification, and can we opt out? A vendor that cannot answer clearly is not ready for patient calls.
Recording and consent
Call recording is regulated separately from HIPAA, by federal and state wiretap and consent laws, and some states require every party's consent. If the agent records calls, the greeting should say so, and your counsel should confirm the wording for the states your patients call from. Consider whether you need audio at all. A transcript and summary may meet your operational needs with less risk, and audio can often be kept for a short quality-review window, then deleted.
Breach terms and penalties
If a breach involves your patients' information, your practice carries the notification duty. Notices to affected individuals must be made without unreasonable delay and no later than 60 calendar days after discovery[5]. That means your business associate agreement should require the vendor to tell you about incidents quickly enough for you to investigate and notify in time.
Penalties are scaled. The inflation-adjusted HIPAA civil money penalty runs from $145 minimum per violation in the lowest tier up to a $2,190,294 calendar-year cap per identical provision[6]. The practical lesson is not fear; it is documentation. A practice that can show its risk assessment, its agreements and its configuration choices is in a far better position than one that cannot.
Designing for minimum data
- Answer general questions from approved facts: hours, location, insurance accepted, forms, portal help.
- Do not ask for dates of birth, symptoms or medication names unless a specific, covered task requires it.
- If a caller volunteers health details, do not repeat them back or include them in summaries beyond what staff need.
- Hand off results, refills, billing and appointments to staff lines.
- Give an approved emergency instruction to call 911, with no follow-up questions.
- Keep summaries short and operational: who called, the callback number and the type of request.
Example summary that avoids health details: 'Caller asked about a refill for an existing patient and requested a callback this afternoon. Callback number provided. Routed to the refill line at 2:14 pm.' The clinical detail is collected by staff on the callback, inside the EHR.
Access controls inside your practice
Vendor safeguards are half the picture. The other half is who in your own office can open call transcripts and summaries. Give each staff member an individual login with multifactor authentication, limit access by role so a biller does not browse clinical callbacks, and remove access on a staff member's last day. Check the vendor's dashboard for export and download controls, since a transcript exported to a personal laptop is outside every safeguard you negotiated. Review the user list every quarter alongside your EHR access review.
If the agent ever calls patients
Inbound answering and outbound calling are treated differently. On February 8, 2024, the FCC ruled that calls made with AI-generated voices are artificial under the Telephone Consumer Protection Act[7], which brings consent rules for artificial-voice calls into play. Under FCC rules, telemarketers must obtain prior express written consent before robocalling consumers[7]. Appointment reminders are not telemarketing, but they still need the right consent and an easy opt-out. Before switching on outbound reminders, recalls or follow-ups with an AI voice, have counsel review your consent wording and records.
Before you sign
- Data flow diagram received and reviewed, with every subcontractor named.
- Business associate agreement signed, or a written decision that no protected health information will be handled.
- Model training use answered in the contract.
- Retention periods set for audio, transcripts and summaries.
- Incident notice timing written into the agreement.
- Staff logins, roles and offboarding process agreed.
- Recording disclosure wording approved by counsel.
- Risk assessment written and filed.
What staff should be able to tell patients
Patients sometimes ask whether they are talking to a real person and what happens to their call. Give the front desk a short, accurate answer they can repeat: the practice uses an automated assistant to answer general questions and route calls, it does or does not record calls, transcripts are kept for a set period and then deleted, and any patient can ask to speak with a staff member at any time. Accurate, consistent answers do more for trust than any privacy statement on the website.
Document your decision
Write a short risk assessment for the agent: what data it handles, the safeguards, the agreements, the retention settings and who reviewed it. Store it with your HIPAA documentation; covered entities must retain required policy and compliance documentation for six years[8]. Review it when the vendor adds features, because a tool that starts by answering parking questions can quietly expand into handling patient-specific requests.
Example: a fictional five-provider OB-GYN practice chose an agent that took no health details for its first year, documented that choice and the data flow, and set transcripts to delete after 30 days. When it later wanted the agent to take refill requests, it reopened the assessment, signed a business associate agreement with the vendor and changed retention before switching the feature on.
Frequently asked questions
Is an AI phone agent HIPAA compliant?
No product is compliant on its own. Compliance depends on what data the agent handles, the safeguards in place, a business associate agreement where protected health information is involved, and how your practice configures and documents it.
Do we need a BAA if the AI only answers general questions?
If the agent truly handles no protected health information, a business associate agreement may not be required. Confirm that with your privacy officer or counsel, and check what the agent records when callers volunteer health details.
Can the vendor use our calls to train its AI?
Only if your agreement allows it. Ask directly and put the answer in the contract, including whether you can opt out.
How long should call transcripts be kept?
Keep them only as long as you need them for operations and quality review, set that period in the vendor's settings and contract, and document the decision.
References
Legal Information Institute, Cornell Law School (current). 45 CFR § 164.502 - Uses and disclosures of protected health information: General rules (opens in a new tab) ↑
KFF (2026). KFF Tracking Poll on Health Information and Trust: Use of AI For Health Information and Advice (opens in a new tab) ↑
American Medical Association (2026). More than 80% of physicians use AI professionally: AMA survey (opens in a new tab) ↑
Medical Group Management Association (MGMA) (2026). AI governance in medical group practices: Rules for the humans in the loop (opens in a new tab) ↑
Legal Information Institute, Cornell Law School (current). 45 CFR § 164.404 - Notification to individuals (opens in a new tab) ↑
Legal Information Institute, Cornell Law School (current). 45 CFR § 102.3 - Penalty adjustment and table (opens in a new tab) ↑
Federal Communications Commission (2024). FCC Makes AI-Generated Voices in Robocalls Illegal (news release) (opens in a new tab) ↑
Legal Information Institute, Cornell Law School (current). 45 CFR § 164.530 - Administrative requirements (opens in a new tab) ↑