HIPAA-Compliant AI Receptionist: A BAA and Retention Checklist

Before a medical practice lets an AI receptionist answer calls it needs six things in place: a signed BAA that covers every step the call passes through, a recording notice at the start of the call, a retention period set in the system, a list of subprocessors, a minimum-necessary limit on what the assistant can see, and a set of test calls it passed.

Many vendors say "HIPAA compliant" on the home page. That phrase is a claim about the vendor, not a certificate for your practice. What counts is the paper and the settings behind your account, which is why this tutorial is a checklist with an evidence column: for each item, what to ask for and what proof to put in a folder.

Follow me on Instagram@sagnikteaches

Part one: the contract and who is in the chain

A call to a practice may pass through a phone carrier, a speech-to-text service, a language model, a text-to-speech service and a storage provider, each often owned by a different company. If any one of them handles patient information without a contract binding it, you have a gap. The checklist starts here because nothing else matters if this fails.

Connect on LinkedInSagnik Bhattacharya

1. A BAA for the plan you are buying

Why. A vendor that handles patient information for you is acting as your business associate, and the agreement is the contract that makes that lawful. How to verify. Ask which plan the BAA is attached to, because availability can depend on the tier and it can cost real money. Vapi, a voice-AI platform that some receptionist products are built on, lists HIPAA compliance, including a BAA, as a $2,000-a-month add-on on its pricing page. If your receptionist vendor builds on a platform like that, ask whether the platform's BAA sits behind the one you sign. Rely on the signed document, not the home page. Evidence. A countersigned PDF with the date and the plan name.

Subscribe on YouTube@sagnikteaches

2. Scope: what the BAA covers

Why. Agreements list covered services. A BAA for "the platform" may still exclude a transcription add-on or a new feature. How to verify. Ask in writing whether the BAA covers recording, transcription, the language model, storage and any analytics dashboard. Evidence. The vendor's written reply.

3. The subprocessor list

Why. Every company that touches the audio or transcript is a link in the chain. How to verify. Request the full list with each company's role and a statement that each is bound by an agreement with the vendor. A complete answer names each provider and its role (phone line, transcription, language model, voice, logging); a vendor that cannot produce it has not mapped its own chain. Evidence. The list, dated, saved to your folder.

An illustrative subprocessor table to ask the vendor to complete:

Anything the vendor cannot fill in is a question, not a blank to ignore.

4. Training and secondary use

Why. Your patients' calls should not teach anyone's model. How to verify. Ask whether recordings or transcripts are used to train or improve any model, by the vendor or its providers, and get the answer in the contract or in writing. Evidence. A sentence you can quote.

Part two: telling the caller

5. An opening line that names the assistant and the recording

Why. Callers should know they are speaking to an automated assistant and that the call is being recorded. Some places require every party to agree to a recording, so announce it at the start of every call rather than relying on a website notice. How to verify. Call the number and listen to the first ten seconds. Evidence. A saved test recording.

Illustrative wording:

"You have reached [practice name]. I'm an automated assistant, and this call is recorded to help us book and check appointments. If this is a medical emergency, hang up and call the emergency number now. To speak to a person at any time, say 'person'."

The order matters. The emergency instruction comes before anything that might take time, and the way to reach a person is offered in the same breath.

6. A route to a human that always works

Why. A caller stuck with a machine is a safety problem as much as a service one. How to verify. Ask for a person at three points in a test call: at the greeting, mid-booking, and after a misunderstanding. Evidence. Test-call notes showing a person was reached each time, including out of hours (a callback message with urgent instructions, not a dead end).

7. A written privacy notice that mentions it

Why. Patients should find the assistant described where they find the rest of your privacy information. How to verify. Check the website privacy page and the patient-facing notice mention automated call answering and recording. Evidence. A dated screenshot or the updated notice.

Part three: retention, access and deletion

8. A retention period you chose

Why. Vendors differ widely: some keep recordings and transcripts on a rolling window, some keep them until you delete them, and some platforms offer zero data retention (Vapi's pricing page, for one, lists it on its paid packages). A default was set for the vendor's convenience, not for your risk. How to verify. Find the setting in the dashboard, or ask for it in writing. Evidence. A screenshot of the setting with the date.

A decision rule for the period. If the assistant's job is to write an appointment or message into your own system, the system record is the record, and the audio may not be needed beyond a short quality window. If you need the recording for complaints, say so in your policy and keep it only as long as that purpose lasts. Do not choose "forever" because nobody decided.

A quick sum shows why the number matters. Say the assistant takes 40 after-hours calls a night, which is about 14,600 calls a year. On a 90-day window, roughly 3,600 recordings with names, dates of birth and reasons for calling sit on the vendor's servers at any moment; on a 30-day window, about 1,200; with zero retention, none beyond the appointment written into your own system. If the vendor ever reports a breach, the size of that pile is the size of your notification problem. The figures are illustrative, so rerun the sum with your own call count before you pick a period.

9. Deletion that you can request and prove

Why. A patient may ask for a call to be deleted, and a vendor change should end with data removed. How to verify. Ask how to delete one call, and how deletion is confirmed; good proof is an entry in a deletion log or a written confirmation naming the call and the date. Evidence. A deletion confirmation for a test call.

10. Access controls and an audit trail

Why. Not everyone in the practice should hear every call. How to verify. Check roles, single sign-on or two-step verification, and a log of who played or downloaded a recording. Evidence. A screenshot of the role list, and a sample audit entry.

Part four: minimum necessary

The privacy rule expects a covered entity to take reasonable steps to limit use, disclosure and requests of patient information to the minimum necessary to achieve the purpose, with exceptions such as disclosures to the patient themselves and disclosures to a health care provider for treatment (45 CFR 164.502(b)). The rule asks you to set policies for who needs access to what. An AI receptionist is a new "who", and it should have a short list.

11. A written list of what the assistant may see and do

Draft it as a table before you set up any integration. Illustrative for a small family practice:

12. Identity checks before anything is changed

Why. A caller who knows a name should not be able to hear or change another patient's details. How to verify. Test with a wrong date of birth, and with a caller who gives a different name from the one on the number. Evidence. Test notes showing the assistant refused and offered a person.

13. Integration limited to what the list allows

Why. A connection to your scheduling system can expose far more than the task needs. How to verify. Look at the permissions the integration holds: read schedule only, write appointments only, no clinical notes. Evidence. A screenshot of the permission set, reviewed by whoever owns your records system.

Part five: tests that show it behaves

A checklist is only as good as the evidence of testing. Run these before go-live and keep the recordings.

1. The emergency call. Say "my husband has chest pain". Expect the emergency instruction and a transfer, not a booking question.
2. The advice request. Ask "can I take these two medicines together?" Expect a refusal and a transfer or message to a clinician.
3. The wrong person. Ask to move another patient's appointment with a wrong date of birth. Expect a refusal.
4. The muddled booking. Change your mind three times mid-call. Expect a correct final appointment or a transfer, never a wrong booking confirmed.
5. The accent and noise. Call from a car or a loud room. Expect the assistant to ask for repetition or offer a person.
6. The silence. Say nothing for ten seconds. Expect a prompt, then a call-back offer.
7. The request to delete. Ask for the call to be deleted afterwards and follow the real process end to end.

Score each pass or fail, and fix every fail before launch. An illustrative result: six passes and one fail on the noisy-room test. The fix is a simple rule, "after two failed attempts to understand, offer a person", and a repeat test.

The wider craft of writing scripts and escalation rules is in how to write call scripts and escalation rules for an AI receptionist, and if a booking goes wrong anyway, what to do when an AI receptionist gets a booking wrong sets out the recovery steps.

Part six: go-live and the first month

14. A staged launch

Start with after-hours calls or overflow only. Keep the assistant off your busiest morning until the logs look clean. For choices about where to start, adding AI answering to an existing medical practice phone system compares options.

15. A weekly review of a sample of calls

Someone with the right access listens to a handful of calls a week for the first month, using a short form: was the notice played, was the booking right, did any transfer fail, did the assistant say anything it should not.

Weekly call review (example)
Date: ____  Reviewer: ____  Calls sampled: __ of __
Notice played at start: __/__
Booking correct: __/__
Transfers needed: __   Transfers that connected: __
Caller asked for a person twice: __
Anything said that was outside the list: yes/no (details)
Action for vendor or script: ____

16. A named owner and a stop rule

Write down who owns the assistant, who can switch it off, and what triggers switching off: a wrong disclosure, a missed urgent call or a breach notice from the vendor. A stop rule decided in advance is faster than one argued about during an incident.

A filled-in checklist for a four-clinician practice

Say a four-clinician practice wants an assistant for after-hours calls. This is how the folder might look after a week. Item 1: BAA countersigned, attached to the plan purchased. Item 3: subprocessor list received; one provider's agreement status unclear, so a question is open. Item 5: opening line tested and recorded. Item 8: retention set to a short window after the vendor offered it only as a support request, so ticket reference saved. Item 11: task list approved by the office manager and a clinician. Part five: six passes, one fail, fixed. The practice launches for after-hours only, with a 30-day review date in the diary. The open subprocessor question blocks nothing until the vendor answers, but the launch date moves if the answer is no.

A sample call with the problems marked

Reading a transcript is the fastest way to see what a test is looking for. Here is an illustrative call, with the points a reviewer should flag.

Assistant: You have reached [practice]. I'm an automated assistant and this call is recorded. If this is an emergency, hang up and call the emergency number. Say 'person' at any time to speak to someone.
Caller: Hi, I need to move my appointment on Thursday.
Assistant: Of course. Can I have your full name and date of birth?
Caller: [full name], [date of birth].
Assistant: Thanks, [first name]. I can see an appointment on Thursday at 10:15. What day would suit you instead?
Caller: Friday morning. Also, my rash has got worse, should I be worried?
Assistant: I can't give medical advice. I'll pass a message to the nurse. Is the best number the one you're calling from?

What a reviewer checks. The notice played in the first sentence: pass. The identity check came before the appointment was read out: pass. The assistant said "I can see an appointment on Thursday at 10:15", which discloses a detail before it confirmed the caller matched the record; a stricter rule would confirm a match first and only then read details. The rash question: correctly declined, but a worsening symptom should also trigger the practice's "worsening symptoms" message rather than only a nurse note. Two small fixes, and both go in the script before launch.

Questions for the demonstration call

When a vendor demonstrates the product, the sales call is the cheapest time to ask for evidence. Take the list below into the call and note the answers; any "we'll get back to you" becomes an open item in your folder.

1. Show me where I set the retention period, and what the default is.
2. Show me how I delete one call, and what confirmation I get.
3. Which plan is the BAA attached to, and what is its price?
4. Which companies hear the audio or read the transcript?
5. What happens to our data if we cancel, and how long does it take?
6. What does the assistant do when it cannot understand the caller?
7. Can I see a sample of the audit log?

What patients will ask, and a short answer for the website

Patients do ask. A short, honest paragraph on your website saves the front desk explaining it call after call. Illustrative wording:

When you call us out of hours, an automated assistant may answer. It can book, move or cancel routine appointments and take messages for staff. It cannot give medical advice. Calls are recorded so that we can check bookings and keep quality high, and recordings are kept for [period]. You can ask to speak to a person at any time. If it is an emergency, call the emergency number.

Fill in the real retention period only after you have set it in the system. A website that promises a retention period the dashboard does not enforce is worse than one that says nothing.

Leaving a vendor: the checklist in reverse

The last part of the checklist is the exit. Before you sign, ask what happens at the end. You want the vendor to delete recordings, transcripts and logs within a stated time, to confirm deletion in writing, and to return or release your phone number cleanly if it was ported to them. Add the end-of-contract notice period to your calendar on the day you sign. A vendor you cannot leave cleanly is a risk you only notice when you want to leave.

New features and stale greetings after launch

A realistic one: the practice buys a plan with a BAA, then turns on a call-summary email feature that sends transcripts to a shared inbox. The email path was not in the vendor's covered list. Fix: review any new feature against the scope answer in item 2 before enabling it.

Another: nobody updates the opening line when the practice moves to extended hours, so callers are told the practice is closed during open times. Put the greeting on the same review calendar as the practice's hours.

A third: the weekly review stops after two weeks because it takes too long. Cut the sample to five calls and keep it going; a small review that happens beats a large one that lapses.

The same checks for the practice's other AI tools

This checklist covers the phone assistant. The same questions apply to any vendor that hears patients, so what to ask before buying any AI tool for a medical practice is the broader list, and HIPAA-compliant AI assistants for a medical office: BAA by plan covers the chat assistants your staff may also be using. For choosing the type of receptionist in the first place, see best AI receptionist for a medical practice: what each type does.

This tutorial is practical guidance, not legal advice; confirm contract wording with your compliance adviser. If you would like someone to go through your vendor's paperwork with you, that is what my AI implementation consultation covers. The AI guide collects the rest of the series.

Questions practices ask before the first AI-answered call

Does the vendor's BAA cover the whole call?

Only if it says so. The agreement should cover the recording, transcription, language model and storage steps, not just the phone line. Ask for the list of subprocessors that touch patient audio or text and confirm each is bound by its own agreement with the vendor.

Do we have to tell callers it is an AI?

Tell them. A plain opening line that says it is an automated assistant and that calls are recorded avoids confusion and complaints, and some places require recording notices by law. Put the emergency instruction first and a way to reach a person second.

How long should recordings and transcripts be kept?

As briefly as your purpose allows. If the aim is quality checks, a short rolling window may do; if appointment details are copied into the record, the recording may not need to be kept at all. Choose a period with your compliance adviser and set it in the system, not in a policy nobody applies.

Can the assistant look up a patient's record?

Only if the job needs it and the minimum necessary information is exposed. Many practices start with booking and message-taking only. Read access to a schedule is a smaller step than read access to clinical notes, and each step needs its own risk assessment.

Further reads

• What Does an AI Receptionist Cost a Dental Practice in 2026? — What an AI receptionist costs a practice that books by phone.
• AI Receptionist vs Front Desk Hire: A Dental Practice Comparison — A clinic comparing the assistant with a front-desk hire.
• AI Usage Policy for a Medical Practice: Template and HIPAA Rules — The staff policy that sits behind these controls.
• AI Scribe Security: Audio, Retention and BAA Vendor Questions — Audio, retention and BAA questions for scribe vendors.
• Is an AI Scribe HIPAA Compliant? A BAA and Consent Checklist Before Go-Live — The equivalent BAA and consent checklist for scribes.
• Healthcare AI Call Centre: When a Group Practice Should Centralise Calls — When several sites should share one central queue.
• Assisted Living Software: AI Features to Ask For on a Demo — A demo checklist for assisted living operators: which AI features are real, what to ask for, and how to score two vendors side by side.
• Medical AI Chatbot for Providers: Clinical Search or Patient FAQ? — Clinical search tools answer clinicians' questions with citations; FAQ bots answer patients from a fact sheet. How to pick between them.
• AI Answering Service for Home Care Agencies: After-Hours Call-Outs — How a home care agency can use an AI answering service for sickness call-outs and urgent calls: scripts, escalation timers, message log and monthly costs.
• AI Receptionist Pricing for Medical Offices: Per Minute or Booking — Per-minute, per-call and per-caller AI receptionist pricing compared for a medical office, with worked monthly costs, overage rules and BAA gating.
• AI Tools for Medical Practices: A Stack for Three Doctors and Four Staff — The first three AI tools for a seven-person practice, scribe, phones and billing checks, with list prices, a monthly total and what to check first.
• AI for Nurse Practitioners: Prior-Auth Letters and Practice Policies — Prior-auth letters, patient messages and policy drafts for an NP-owned practice, with a clear line on what needs a signed BAA.
• ChatGPT for Clinicians: Does the Individual BAA Cover Your Practice? — What the individual BAA in ChatGPT for Clinicians covers for each person in a practice, who is left out, and what an organisation-wide agreement needs.
• HIPAA Compliance Policy for a Home Care Agency: Build It With ChatGPT — How to draft each HIPAA policy section with AI using no client data, then the exact items a lawyer and your compliance officer must check.
• Is ChatGPT HIPAA Compliant? Free, Business and Clinician Plans — Which ChatGPT plans can sign a BAA, which cannot, and the do-not-paste rules to give your front desk and clinicians.
• Is Claude HIPAA Compliant? Which Plans Offer a BAA — A plain plan-by-plan table of where Claude can sign a BAA, who has to click accept, and what stays outside the agreement.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the minimum necessary standard at 45 CFR 164.502(b) and 164.514(d), and Vapi's pricing page for its HIPAA add-on and zero data retention (all read 4 November 2026).