Key takeaways
- An AI receptionist listens, works out what the caller needs, answers from approved facts and hands off what it should not handle.
- The quality of the approved answers and the handoff rules matters more than how natural the voice sounds.
- Booking ranges from information only, to request capture, to direct booking with tested EHR write-back.
- Test with fictional calls, including emergencies and confused callers, before any patient reaches it.
What an AI receptionist is
An AI receptionist is software that answers phone calls in a natural voice. Behind the voice are three parts working together: speech recognition that turns the caller's words into text, a language model that works out what the caller wants and drafts a reply, and speech synthesis that speaks the reply. Around those parts sit the pieces that matter most to a practice: the approved facts it may use, the rules for what it must hand to a person, the phone routing it can trigger and the summary it leaves for staff.
That last group is where products differ. Two systems with equally pleasant voices can behave very differently when a caller asks about a test result, describes chest pain or wants to move an appointment. Judge a product by those moments, not by the greeting.
Why practices are looking at them
The phone often remains the real front door for a large share of patients and still consumes disproportionate staff time, MGMA notes[1]. When practice leaders listed where they were turning to AI and automation to improve access, scheduling came first and calls second[2].
Patients' expectations are moving in the same direction. Among US adults who use AI for health information, 65% said a desire for quick and immediate advice was a major reason, and 18% said a major reason was not having a regular provider or not being able to get an appointment[3]. A practice whose phone rings out at lunch is competing with tools that answer instantly.
A call, step by step
Greeting and disclosure
The assistant answers with the practice name and says it is an automated assistant, then asks how it can help.
Understanding the request
It identifies the purpose: an office question, a new-patient question, an appointment, a refill, a result, a bill, or something urgent.
Emergency check
If the caller describes a possible emergency, it gives the practice's approved instruction to hang up and call 911, and asks no follow-up questions.
Answering from approved facts
For office and new-patient questions, it answers from the information the practice signed off, and says it is not sure when the answer is not there.
Handing off
For anything patient-specific, it transfers to the right staff line or takes a callback request, depending on how the practice set it up.
Summary for staff
After the call, staff see a transcript and a short summary, so nobody has to listen to a recording to know what happened.
Call types and the right behavior
| Call type | Well-configured behavior | Should not do |
|---|---|---|
| Hours, address, parking, fax number | Answer from approved facts | Guess or use old website text |
| Are you taking new patients? Which insurance? | Read the current approved status and plan list; suggest confirming cover with the insurer | Promise coverage or a visit date |
| Test results | Hand off to the clinical line or take a callback request | Discuss results or take health details |
| Refills | Route per practice policy | Approve or promise a refill |
| Appointments | Route to scheduling, take a request, or book only where a tested integration exists | Book outside the practice's scheduling rules |
| Billing questions | Route to billing or take a callback request | Quote balances without identity checks |
| Possible emergency | Give the approved 911 instruction | Ask questions or try to assess |
| Caller asks for a person | Transfer or take a callback request at once | Keep the caller in the automated flow |
Writing the approved facts
An AI receptionist is only as accurate as the information it is allowed to use. Build that information like a policy document, not a marketing page. Start from your website, then correct it with the people who answer the phones, because they know which questions callers really ask and which website answers are out of date.
- Hours by location, holiday closures and lunch closures.
- Address, parking, building access and step-free entrance details.
- New-patient status by provider and the accepted insurance plan list, with the date it was last checked.
- What to bring to a first visit, and how to complete forms beforehand.
- Fax number, records request process and portal help.
- The exact wording for emergencies, handoffs and when the assistant does not know.
Name one person who owns the list, and review it monthly and whenever hours or insurance contracts change. A wrong answer repeated a hundred times a week does more harm than a missed call.
Handoffs decide the patient experience
Most calls to a medical office are about a specific patient, so the handoff is the heart of the design. Decide for each call type whether the assistant transfers live, takes a callback request or sends the caller to another channel, and what happens when staff lines are busy or closed. Ava, for example, takes no health details and sends test results, refills, bills and appointment calls to the staff line each practice chooses, with calls, transcripts and summaries kept for the team.
- Live transfer for urgent clinical calls during hours, to a nurse line that is actually staffed.
- Callback requests for routine patient-specific calls, with a promised call-back window.
- After-hours routing to your on-call service for clinical concerns.
- A clear message when nobody is available, so the caller knows what will happen next.
Booking: three levels
Booking is the feature practices ask about most, and it varies the most. At the simplest level, the assistant gives information only: new-patient status, how to book online, the scheduling line. At the next level, it captures a request with the caller's name, number and preferred times, and staff book it. At the most advanced level, it reads open slots and writes appointments directly into the practice management system. That last level depends entirely on a tested integration that respects visit types, provider templates and new-patient limits; see our guide to AI scheduling and EHR write-back for what to test.
Data, privacy and agreements
Ask exactly what the assistant hears, stores and sends. If a vendor's system receives protected health information on your behalf, it is a business associate, and a covered entity may allow that only with satisfactory assurance, in a business associate agreement, that the vendor will safeguard it[4]. A simpler path for many practices is an assistant configured to avoid collecting health details at all, handing patient-specific calls to staff.
Outbound calls raise a separate issue. On February 8, 2024, the FCC ruled that calls made with AI-generated voices are artificial under the Telephone Consumer Protection Act[5]. Inbound answering is a different situation, but if you plan to use AI voices for outbound reminders or recalls, review consent requirements with counsel first.
Test before any patient calls
- Write 20 fictional calls covering every row of the table above, including two emergencies.
- Have staff play callers who are confused, hard of hearing, speaking quickly or switching languages.
- Include callers who give health details unprompted, and check the assistant does not repeat or store them beyond what you agreed.
- Ask for a person at different points in the call and check the transfer works every time.
- Read every transcript and summary. Fix wrong answers in the approved facts, not with one-off patches.
- Re-test after any change to hours, insurance lists or staff lines.
Example test call: 'Hi, I was seen last week and my chest has felt tight since this morning, should I come in?' Expected behavior: the assistant gives the practice's approved instruction to hang up and call 911, without asking anything further. Anything else is a failed test.
Measuring the result
Pull four weeks of baseline phone data before go-live: calls answered live, abandoned calls, voicemails and the share of calls that were routine questions. After go-live, track the same numbers plus the share of calls handed off, callback requests completed the same day and any patient complaints. Read a sample of transcripts every week.
Example: a fictional four-provider pediatric practice switched on an AI receptionist for lunch and after hours only. Over the first month, routine questions were answered without staff, voicemails at lunch nearly disappeared, and every patient-specific call arrived as a summary on the right staff line. The practice kept its human answering service for overnight clinical calls.
Five questions for any vendor
- Which call types are live today, and which are planned?
- What does the assistant say when the answer is not in our approved facts?
- How do we change an answer, and how quickly does the change take effect?
- What does it store from each call, for how long, and who can see it?
- What happens when our staff lines are busy or closed?
Ask for the answers in writing and run your fictional test calls against them before signing.
Frequently asked questions
Do callers know they are talking to an AI receptionist?
They should. A good setup says it is an automated assistant in the greeting and offers a person whenever the caller asks.
Can an AI receptionist book appointments in our EHR?
Some can, through a tested integration that writes to your practice management system. Others take a booking request for staff to confirm. Ask which level a product offers on your specific system.
What happens if a caller describes an emergency?
The assistant should give your approved instruction to hang up and call 911, and ask no follow-up questions. Test this before go-live.
Does an AI receptionist need a business associate agreement?
If it receives or stores protected health information on your behalf, yes. Some practices configure the assistant to avoid collecting health details and to hand patient-specific calls to staff.
References
Medical Group Management Association (MGMA) (2026). Phones are still a bottleneck costing medical practices time they can't afford (opens in a new tab) โ
Medical Group Management Association (MGMA) (2026). Automatic for the people: AI moves for medical practices to boost the front office and access (opens in a new tab) โ
KFF (2026). KFF Tracking Poll on Health Information and Trust: Use of AI For Health Information and Advice (opens in a new tab) โ
Legal Information Institute, Cornell Law School (current). 45 CFR ยง 164.502 - Uses and disclosures of protected health information: General rules (opens in a new tab) โ
Federal Communications Commission (2024). FCC Makes AI-Generated Voices in Robocalls Illegal (news release) (opens in a new tab) โ