In this article
- Why HIPAA applies to an AI receptionist
- There is no official HIPAA certification
- What a business associate agreement covers
- A BAA review checklist
- Where PHI shows up in an AI call flow
- Minimum necessary and role-based access
- Call recording, notice, and consent
- State recording laws
- Recordings are PHI
- Outbound AI calls and the TCPA
- What to ask an AI receptionist vendor
- Privacy and contract
- Security
- Operations
- Clinical safety limits and escalation
- Practice responsibilities that a vendor cannot take over
- A testing plan before patients call
- Why a billing company cares about the phone line
- How Mednex Group approaches AI receptionists and PHI
- Common mistakes
- Worked scenarios: what the rules mean on a real call
- Hypothetical example: a new patient books a visit
- Hypothetical example: a family member calls for results
- Hypothetical example: after-hours voicemail transcripts by email
- Hypothetical example: the vendor changes its language model provider
- Adding the system to your risk analysis
- Contract terms to check beyond the BAA
- Next steps
Why HIPAA applies to an AI receptionist
A phone call to a medical office often includes PHI. The caller gives a name, a date of birth, a reason for the visit, a medication, or a symptom. When the practice is a covered entity and a vendor handles that content for it, the vendor stands in the shoes of the practice for that task.
HHS defines a business associate as a person or organization that performs functions or activities for a covered entity, or provides certain services, where the work involves creating, receiving, maintaining, or transmitting PHI. A call-answering service, a transcription tool, and a cloud platform that stores call audio can all meet that definition.
Two points are easy to miss:
- Using an AI voice agent does not change the analysis. If the system hears, transcribes, or stores PHI, the vendor is in scope.
- The vendor’s subcontractors are in scope too. Speech recognition, language model hosting, telephony carriers, and storage providers may all touch the data.
For what these systems do in a practice, see what an AI receptionist can handle. For the broader service, see AI automation for healthcare.
There is no official HIPAA certification
Vendors often advertise “HIPAA compliant” or “HIPAA certified.” HHS states that its Office for Civil Rights does not endorse, certify, or recommend specific technology or products. Compliance depends on how the covered entity and the business associate meet the Privacy, Security, and Breach Notification Rules, and on the agreement between them.
Treat a badge on a website as a starting question, not an answer. Ask what the vendor does, in writing: safeguards, access controls, audit logs, encryption, incident response, and subcontractor terms. A third-party security report or attestation can help, but it does not replace a BAA or your own risk analysis.
What a business associate agreement covers
A BAA is a contract that sets the rules for how a business associate may use and protect PHI. According to HHS, it must, among other things:
- Describe the permitted and required uses and disclosures of PHI.
- Bar the business associate from using or further disclosing PHI except as the contract or the law allows.
- Require the business associate and its subcontractors to follow the applicable Security Rule requirements.
- Require reporting of security incidents, including breaches of unsecured PHI.
HHS also states that a business associate must have a BAA with a subcontractor before disclosing PHI to it, and that the covered entity does not have to contract directly with those subcontractors. This is why you ask who the subcontractors are. Your practice relies on the vendor to pass the obligations down.
A BAA review checklist
| Item | What to look for | Red flag |
|---|---|---|
| Permitted uses | Limited to providing the service to you | Broad rights to use data for the vendor’s own purposes |
| Subcontractors | Named or categorized, with flow-down terms | No mention, or “third parties as needed” |
| Breach and incident reporting | Defined reporting duty and contact path | Vague or no timeline |
| Data return and deletion | Return or destroy PHI at termination, with confirmation | Vendor keeps data indefinitely |
| Model training | Clear statement on whether your data trains models | Silence on the question |
| Location and storage | Where audio and transcripts are stored and for how long | Unknown retention |
| Audit and access | Access logs and the ability to request them | No logging |
| Liability and insurance | Reviewed by your counsel | Not reviewed |
Have your attorney review the BAA. A template from the vendor is a draft, not a final answer.
Where PHI shows up in an AI call flow
Map the data path before you sign. Each point below is a place where PHI may exist or move.
- Telephony. The carrier routes the call and may hold call metadata and audio.
- Speech to text. The system converts speech to text. Ask whether this happens on the vendor’s systems or a subcontractor’s.
- Language processing. A model interprets the request. Ask where it runs and whether prompts and outputs are stored or used for training.
- Integrations. The system may read or write to your scheduling or practice management software. Confirm what data it can see and change.
- Storage. Recordings, transcripts, and summaries sit somewhere. Know where, how long, and who can access them.
- Notifications. Texts, emails, and voicemail transcriptions to staff or patients can carry PHI. Check how they are secured.
- Analytics and dashboards. Reports may show call content. Control who can open them.
If any step is unclear, the vendor should be able to draw it. A vendor that cannot describe its own data flow is not ready for a medical office.
Minimum necessary and role-based access
HIPAA’s minimum necessary standard asks covered entities to take reasonable steps to limit PHI uses and disclosures to what is needed for the purpose. HHS says policies should identify which workforce roles need access, what PHI each role needs, and the conditions for access.
Apply that to an AI receptionist:
- Give the system access to appointment slots and basic scheduling data, not the full chart.
- Limit what the system may read back to a caller. Do not let it recite clinical details to a caller who has not been verified.
- Restrict who on the staff can listen to recordings and read transcripts.
- Document why each integration needs the data it uses.
A voice agent should confirm identity before sharing appointment details or balances. Decide what counts as enough verification for each type of disclosure, and have the system hand off to staff when a caller cannot be verified.
Call recording, notice, and consent
Many AI receptionists record and transcribe calls. Two sets of rules matter.
State recording laws
States differ on whether one party or every party must consent to a recording. If a patient calls from another state, more than one state’s rule may apply. A common approach is a short disclosure at the start of the call, such as stating that the call may be recorded and that an automated assistant is answering. Whether that is enough in your states is a question for counsel.
Recordings are PHI
When a recording contains identifiable health information held by your practice or its business associate, it is PHI and falls under HIPAA protections. Decide who may access recordings, where they are stored, how long you keep them, and how patients’ requests are handled. A vendor providing recording needs to be under a BAA.
Outbound AI calls and the TCPA
The Telephone Consumer Protection Act (TCPA) is a separate law from HIPAA. In February 2024 the FCC ruled that calls using AI-generated voices fall under the TCPA’s “artificial or prerecorded voice” category, which generally requires the recipient’s prior express consent, with limited exceptions. If your AI system places outbound calls, such as reminders or recall outreach, have counsel confirm consent, opt-out handling, and the applicable exceptions. This applies to outbound calling, not to a system that only answers inbound calls.
What to ask an AI receptionist vendor
Use this list in your first call. Ask for answers in writing.
Privacy and contract
- Will you sign a BAA before any PHI is processed, and will you flow it down to every subcontractor that touches PHI?
- Which subcontractors handle telephony, speech recognition, language models, and storage?
- Does any of our data train your models or anyone else’s?
- How long are recordings, transcripts, and logs kept, and can we require deletion?
- What happens to our data when the contract ends?
Security
- How is data encrypted in transit and at rest?
- Who at your company can access our recordings, and are those accesses logged?
- Do you require multi-factor authentication for staff accounts?
- Can you share a recent independent security assessment, and what does it cover?
- What is your incident response process, and how will you notify us?
Operations
- What can the AI say and do, and what does it hand to a human?
- How does it handle emergencies and symptom descriptions?
- How does it verify callers?
- What does it write into our scheduling system, and can we review changes?
- Can we review call summaries and listen to samples during the trial?
The test is whether the vendor answers directly. Evasive answers about data flow or subcontractors are a reason to pause.
Clinical safety limits and escalation
HIPAA is one risk. Patient safety is another. An AI receptionist should not give clinical advice, interpret symptoms, or reassure a patient who describes an urgent problem.
- Script a clear emergency message that tells callers to contact emergency services when appropriate.
- Define keywords and situations that trigger immediate transfer to staff or an on-call line.
- Ban the system from speculating about test results, medications, or diagnoses.
- Test the system with hypothetical scenarios, including angry callers, unclear speech, and callers in distress, before it goes live.
Keep a record of what the system is allowed to do. If the practice’s policies change, update the configuration and note the date.
Practice responsibilities that a vendor cannot take over
Signing a BAA does not transfer your HIPAA duties. The practice still needs to handle the following.
- Risk analysis. Include the AI system and its data flows in your Security Rule risk analysis, and update it when the configuration changes.
- Policies and training. Staff need to know what the system does, how to review its work, and how to respond when it hands off a call.
- Notice of privacy practices. Check whether your notice and patient communications need updates.
- Access management. Add and remove staff accounts for the vendor’s dashboard when roles change.
- Incident handling. Know how you will receive a breach notice and who decides on patient notification.
- Documentation. Keep the BAA, vendor answers, test results, and configuration decisions on file.
Your privacy officer or compliance advisor should own this list. If you do not have one, assign the role and write it down.
A testing plan before patients call
- Dry run with staff. Call with scripted scenarios: new patient, reschedule, billing question, refill request, urgent symptom, wrong number.
- Verify the data path. Confirm where recordings and transcripts appear and who can open them.
- Test the handoff. Confirm transfers reach a live person, and that voicemail and after-hours routing work.
- Check the scheduling writes. Make sure appointments created by the system match the right provider, visit type, and time.
- Review transcripts. Read samples for accuracy and for statements the system should not make.
- Pilot with a limited call type. Start with after-hours or a narrow category, then widen.
- Review weekly at first. Track errors, handoffs, and patient complaints.
Why a billing company cares about the phone line
Front-desk calls drive revenue cycle outcomes. Wrong insurance details, missing authorization, and unclear scheduling notes at the call become denials weeks later. If an AI receptionist captures insurance information, confirm what it collects, how it is validated, and where it is stored. Eligibility and benefits verification still need a check against the payer, not the caller’s recollection. See prior authorization services and medical billing services for how the front end connects to claims.
For a view of where front-end errors are showing up in your claims, request a Free Medical Billing Audit. It reviews recent claims, denial reasons, and A/R aging by payer. It does not include legal or HIPAA compliance review.
How Mednex Group approaches AI receptionists and PHI
Mednex Group offers AI receptionists and AI provider-office assistants for call handling, routing, scheduling, and approved administrative workflows, along with virtual medical assistants and receptionists and billing services. The approach is to define the tasks in writing, limit the data the system can see, and send anything outside the approved list to a person.
Mednex handles PHI only under a signed BAA and a secure workflow. Practices should not send patient information through the website contact form. Mednex does not claim to make a practice HIPAA compliant, and it does not provide legal advice. Your counsel and compliance advisor should review the agreement and configuration.
Common mistakes
- Accepting a “HIPAA compliant” claim without a BAA.
- Ignoring subcontractors, especially the speech and language model providers.
- Letting the system read back account or appointment details without caller verification.
- Forgetting state recording consent rules and outbound TCPA consent.
- Giving the AI access to the full chart when it needs only the schedule.
- Launching without an emergency script and a tested handoff.
- Leaving the AI system out of the risk analysis.
Worked scenarios: what the rules mean on a real call
Rules are easier to apply when you walk through calls. These are hypothetical examples with no real patient data.
Hypothetical example: a new patient books a visit
A caller gives a name, birth date, phone number, insurance plan, and a reason for the visit. The system hears PHI, may transcribe it, and writes an appointment into your scheduler. Each step needs coverage: the vendor under a BAA, the speech and language subcontractors under their own agreements, the scheduling integration limited to the fields it needs, and the recording stored under your retention rule. If the system then texts a confirmation, check that the text carries only the minimum, such as the time and location, and not the reason for the visit.
Hypothetical example: a family member calls for results
A caller says they are the patient’s spouse and asks whether lab results came back. The system cannot verify the relationship or whether the patient has authorized the disclosure. The correct behavior is to decline to share anything, take a callback message, and route it to staff who can check the patient’s recorded preferences. The practice, not the AI vendor, decides who is allowed to receive information, so the preference data and policy have to exist before the system goes live.
Hypothetical example: after-hours voicemail transcripts by email
An after-hours system sends a transcript of each message to a shared mailbox. The transcript contains a name and a symptom. If the mailbox is a general email account without encryption or access controls, PHI may sit where too many people can read it. Fix this by sending a notification that says a message is waiting and requiring staff to open it in the secured dashboard, or by using a secured email arrangement your compliance advisor approves.
Hypothetical example: the vendor changes its language model provider
A vendor switches the underlying model provider mid-contract. If the vendor has not told you and has no BAA with the new provider, PHI may flow to a party with no contract at all. Your BAA should require the vendor to notify you of subcontractor changes, and your annual vendor review should ask for the current list.
Adding the system to your risk analysis
The HIPAA Security Rule expects covered entities to analyze risks to electronic PHI and to act on what they find. A new AI vendor changes the picture, so update your analysis rather than treating the purchase as a separate project. A short working sequence:
- List the assets. Add the AI service, the phone number it answers, the dashboard, the recordings, and the scheduling integration.
- Map the data. Use the data path from the earlier section and name an owner for each point.
- Identify threats. Examples: misrouted calls, misheard identity, unauthorized dashboard access, over-retention of recordings, a vendor breach, and a prompt or integration that exposes more data than intended.
- Rate likelihood and impact. Use the scale your practice already uses.
- Decide controls. Examples: MFA on dashboard accounts, shorter retention, narrower integration scope, caller verification rules, and a monthly access review.
- Document and schedule review. Record the decisions and review them when the vendor, scope, or integrations change.
This is a general outline. Your compliance advisor can tailor it to your size and systems.
Contract terms to check beyond the BAA
The BAA covers PHI. The service agreement covers everything else, and gaps there create problems later.
- Service levels. What happens when the system is down, and does calling fail over to a human line or to voicemail?
- Change notice. Will the vendor tell you before changing models, subcontractors, or data locations?
- Ownership. Who owns recordings, transcripts, and configuration, and can you export them?
- Indemnity and limits. How is liability allocated if the vendor causes a breach? Counsel should read this clause closely.
- Termination. What are the notice terms, and how quickly is PHI returned or destroyed after you leave?
- Insurance. Ask whether the vendor carries cyber liability coverage and whether your own policy has conditions about vendors.
Next steps
Write down the data path, ask vendors the questions above in writing, and have counsel review the BAA and recording language. For practices that also want a review of front-end errors in claims, the Free Medical Billing Audit is the entry point. More guides are on the Mednex blog, and general questions can go through the contact page without patient details.
This guide is educational. It is not legal, coding, compliance or clinical advice. Rules differ by payer, plan, state and contract, so confirm against the payer's current policy before acting.
Frequently asked questions
Does an AI receptionist need a BAA?
Usually yes. If the vendor creates, receives, maintains, or transmits PHI for your practice, such as hearing callers give names and reasons for visits, or storing recordings and transcripts, HIPAA generally treats the vendor as a business associate, and a covered entity needs a signed BAA before PHI is handled. The vendor also needs BAAs with its own subcontractors that touch the data, such as speech recognition or hosting providers. Ask for the BAA before a trial with real patient calls, and have your attorney review it. Do not let real patient calls reach the system until the BAA is signed.
Is there a HIPAA certification for AI receptionist software?
No. HHS states that its Office for Civil Rights does not endorse, certify, or recommend specific technology or products. A vendor claiming to be HIPAA certified is not referring to an HHS program. What matters is whether the vendor will sign a BAA, what safeguards it has, how it handles subcontractors, and whether the practice completes its own risk analysis and policies. Independent security assessments can be useful evidence, but they do not replace the agreement or your own compliance work. Keep a copy of every vendor answer with the BAA in your compliance file.
Can an AI receptionist record patient calls?
It can, but recording brings two checks. First, state recording consent laws differ: some require every party to agree and others require one party, and calls across state lines may involve more than one state. A short disclosure at the start of the call is common, and counsel should confirm what is enough for your states. Second, a recording that contains identifiable health information held by your practice or its business associate is PHI, so storage, access, retention, and vendor terms need to be controlled and covered by a BAA. Keep the retention period short and written down, and let patients know calls may be recorded.
What PHI does an AI receptionist handle?
Anything a caller says that identifies them and relates to care or payment can be PHI: name, date of birth, phone number, reason for visit, provider, insurance details, appointment times, and balances. It also creates PHI in recordings, transcripts, summaries, and text notifications. Map each place where this data is heard, processed, stored, or sent, including telephony, speech to text, language models, integrations with scheduling software, and dashboards. A vendor that cannot describe that data path clearly is not ready for a medical office. Ask the vendor to draw the data path on one page and keep that page on file.
Does the minimum necessary standard apply to an AI receptionist?
Yes, as a practical design principle. HIPAA's minimum necessary standard asks covered entities to take reasonable steps to limit PHI use and disclosure to what the purpose needs, and HHS says policies should define which roles need which PHI. For an AI receptionist, that means giving it access to scheduling data rather than the full chart, limiting what it may read back to callers, requiring identity verification before disclosure, and restricting which staff can open recordings and transcripts. Document why each integration needs the data it uses. Review the settings again whenever the vendor adds an integration.
Do outbound AI calls need patient consent?
Often yes, under the TCPA, which is separate from HIPAA. In February 2024 the FCC ruled that calls using AI-generated voices fall under the TCPA's category of artificial or prerecorded voice calls, which generally require the recipient's prior express consent, with limited exceptions. If an AI system places reminder or recall calls, have counsel confirm consent, opt-out handling, and which exceptions apply. This rule concerns outbound calling. A system that only answers inbound calls raises different questions, mainly around PHI handling and recording. Ask counsel to confirm the exceptions that might apply to your reminder calls.
What questions should I ask an AI receptionist vendor about HIPAA?
Ask whether the vendor signs a BAA before any PHI is processed and flows it to subcontractors, which subcontractors handle telephony, speech recognition, language models, and storage, whether your data trains any model, how long recordings and transcripts are kept and how they are deleted, how data is encrypted, who can access recordings and whether access is logged, how incidents are reported, and what happens to data when the contract ends. Request the answers in writing and keep them on file with the BAA. If a vendor will not answer in writing, treat that as a signal to keep looking.
Who is responsible for HIPAA compliance, the practice or the vendor?
Both, in different ways. The vendor, as a business associate, is directly responsible for meeting applicable Security Rule requirements and its BAA terms. The practice remains responsible for its own compliance: risk analysis that includes the AI system, policies and staff training, access management, incident handling, and deciding what the system may do. A BAA does not move the practice's duties to the vendor. Keep the BAA, vendor answers, test results, and configuration decisions documented, and have a privacy officer or compliance advisor own the process. Review this split in your annual risk analysis and when the vendor changes.
Can an AI receptionist give clinical advice or triage symptoms?
It should not. An AI receptionist is suited to administrative tasks such as scheduling, routing, and basic information. It should not interpret symptoms, discuss results or medications, or reassure a caller who describes an urgent problem. Configure a clear emergency message, define situations that trigger immediate transfer to staff or an on-call line, and test with scenarios such as distressed callers and unclear speech before launch. Review transcripts regularly after launch. These limits are a patient safety control as well as a compliance control. Update the escalation script whenever clinical leaders change the policy.
How should a practice test an AI receptionist before going live?
Run scripted calls covering new patients, rescheduling, billing questions, refill requests, urgent symptoms, and wrong numbers. Confirm where recordings and transcripts are stored and who can open them, that transfers reach a live person, that after-hours routing works, and that appointments the system creates match the right provider, visit type, and time. Read transcripts for accuracy and for statements the system should not make. Start with a narrow call type such as after-hours, review results weekly at first, and widen scope only when errors are low. Keep the test results as part of your compliance file.
