Can an AI receptionist answer a practice’s phones? What it can and cannot do

A dentist I work with counted her missed calls for one week. Most of them came in at lunch and in the half hour after closing, and most of the callers wanted one of four things. That count is the right place to start with an AI receptionist, because it tells you which calls a machine could take and which ones it never should.

What can an AI receptionist actually handle on a practice phone line?

It handles calls where the answer is already written down and the next step is a system action: office hours, directions, parking, which insurance plans you take, booking or moving an appointment, and taking a message with a callback number. Those calls follow a script your front desk could hand to a new hire on day one.

The test I use is simple. If a new front-desk hire could handle the call on their first morning with a one-page sheet, an AI receptionist can probably handle it too. If the call needs judgment your best receptionist built over five years, keep it with a person.

Start with the calls that go to voicemail today. Nobody is answering them now, so any correct answer is an improvement, and there is no good human conversation being replaced by a worse machine one. Your team keeps the calls it already handles well, and you learn how the system behaves on the calls that matter least.

What should an AI receptionist never do?

It should never assess symptoms, suggest what a caller should take, or decide whether something is urgent. Those are clinical calls. The script should send any call that mentions pain, bleeding, a reaction or an emergency straight to a person, or tell the caller to hang up and dial 911 when the practice’s own protocol says so.

Write that rule down before you pick a vendor, in words your clinical lead has signed off. Then test it. Call the line yourself and describe a problem in plain, rambling language, the way a scared patient would. If the system tries to be helpful about the symptom instead of handing you off, it is not ready for your patients.

Does an AI receptionist need a business associate agreement?

Almost always. HHS describes a business associate as a person or company that performs services for a practice involving the use or disclosure of protected health information. A phone system that hears a patient’s name, date of birth and reason for visit, then records or stores it, fits. HHS requires a written agreement before you share that information.

Storage counts even when the vendor cannot read what it stores. HHS guidance on cloud services says a provider that maintains electronic patient information is a business associate even if it holds only encrypted data and lacks the key. Call recordings and transcripts sitting on a vendor’s servers are exactly that.

Can an AI voice call patients back or make reminder calls?

Inbound answering and outbound calling are treated differently. In February 2024 the FCC confirmed that AI-generated voices count as an “artificial or prerecorded voice” under the Telephone Consumer Protection Act. Calls using one need the called party’s prior express consent unless an exception applies. Check your consent records and any exception with counsel before switching outbound calls on.

Voicemail matters too. HHS says a practice may leave appointment messages on an answering machine, and should limit what it says. Its own suggestion is to leave the practice name, a number and what is needed to confirm the appointment, or simply ask the patient to call back. Put that limit into the script word for word, because an AI system will happily read out whatever it has.

How do you test an AI receptionist before patients hear it?

Use your own calls. Pull twenty real call types from last month, including the awkward ones, and run each through the system with a staff member playing the patient. Score every call on three questions: was the answer right, did it hand off when it should have, and would you be comfortable if the patient heard the recording.

1. Write the one-page sheet first: hours, services, insurance list, booking rules, handoff triggers. The system is only as current as that sheet.
2. Get the BAA signed before any real patient call reaches the system.
3. Run the twenty-call test with staff, including a caller who mentions a symptom and a caller who is angry.
4. Go live on overflow only, such as lunch and after hours, with every call logged for review.
5. Review a sample of recordings weekly for the first month. Widen the hours only when the handoffs are reliable.

Be skeptical of claims that sound too clean. In its 2024 sweep of deceptive AI marketing, the FTC’s chair said there is no AI exemption from the laws on the books. A vendor promising that its system “handles every call” is making a claim you can check with your own recordings in an afternoon, so check it.

Who should not start with the phones?

If your scheduling rules live in one person’s head, the phone system will expose that on day one. Write the rules down first, then automate them. And if most of your calls are from existing patients with clinical questions, an AI receptionist will spend its day handing off, and your team will feel it as an extra step rather than a relief. In that case the better first project is usually somewhere else in the practice.