On this page
  1. What is an AI receptionist?
  2. What can an AI receptionist safely handle?
  3. What must it hand to a human?
  4. What should you ask an AI receptionist vendor?
  5. How do you test one before patients meet it?
  6. Where does it go wrong?
  7. Where we fit, and where we don't
  8. Common questions
Short answer

An AI receptionist for a medical practice is software that answers patients for you, either by voice on the phone or in writing by text, email and web chat. It can safely handle hours, booking links, reminders and rescheduling. It should hand anything about symptoms, doses, refills or an urgent problem to a person, and it needs a BAA before it touches a patient's details.

This guide is for solo NPs and physicians running cash-pay hormone, weight-loss and functional medicine practices, often with one VA or none. We don't name or rank vendors here. The products change every few months, and the questions below will still be useful after they do.

What is an AI receptionist?

The term covers two different products. Know which one you are being sold before you compare anything else.

The two kinds of AI front desk
KindWhat it doesWhere it fitsWhat to watch
Phone voice agentPicks up calls, speaks with the caller, books or takes a messagePractices where most new patients callMishearing names and dates, callers who are upset or unwell, call recording rules
Written assistantReplies to texts, emails, web chat and form inquiries, and sends links and remindersPractices where patients come from ads, Instagram and web formsHealth details landing in a text thread, replies that sound sure and are wrong

Some products do both. Many cash-pay practices that advertise online get most inquiries in writing, at night, from a form. If that is you, a written assistant covers more of your front desk than a voice agent does.

What can an AI receptionist safely handle?

Admin work that has one right answer, which you have written down. If your VA could answer it from a one-page FAQ without asking you, an assistant can too.

  • hours, location, time zone and whether you see patients in a given state
  • how a first visit works, in the order you set: intake form, labs, consult
  • sending the booking link, the intake form link or a payment link
  • appointment confirmations and reminders
  • intake and payment reminders when a step is still open
  • rescheduling and cancelling inside your policy
  • "do you take insurance?" and other questions your website already answers

Fees are a judgment call. An assistant can repeat what your website says. It should not quote a fee for a case it has to interpret.

What must it hand to a human?

Anything clinical, anything urgent and anything where the patient is upset. The assistant's job in those cases is to stop, say a person will reply, and tell you.

  • Symptoms and side effects. "Is this normal?" is a clinical question, however casually it is asked.
  • Doses and medication changes. An assistant should never tell a patient to take more, less or none.
  • Refills. It can note the request. The clinician approves it.
  • Lab results. It can say results are in the portal. It should not read or explain them.
  • Anything urgent. The reply should give emergency instructions, such as calling 911, and alert you straight away.
  • Complaints, refunds and billing disputes. A person should own these.

A workable middle ground is a draft. The assistant writes a reply to a clinical question, and nothing sends until the clinician reads and approves it. That is how we set up the practices we work with: admin messages go out on their own, and anything about doses, symptoms or refills waits for the clinician.

What should you ask an AI receptionist vendor?

Ask these in writing and keep the answers. A vendor that can't answer the first two is not ready for a medical practice.

  1. Do you sign a BAA, and on which plan? A phone call or text thread with a patient is patient information. Ask whether the BAA covers every company the vendor uses behind the scenes, including the AI model and the phone or texting carrier. See which practice tools sign a BAA.
  2. Where is the data stored, for how long, and is it used to train models? You want recordings, transcripts and messages kept only as long as you need them, and not used for training.
  3. How does handoff work? Ask what triggers it, who gets told, how fast, and what the patient hears or reads while they wait. Ask to see the list of topics it will not answer.
  4. Are patients told it is an assistant? They should be, at the start. Patients forgive an assistant that says what it is. They don't forgive finding out later. Some states also have rules about automated calls and call recording, so check yours.
  5. What happens after hours? Find out whether it answers at night, what it does with an urgent message at 2 am, and who is alerted.
  6. Does it connect to your EHR or scheduler? Ask for the name of your EHR on a list, not "we integrate with most systems". Ask whether it reads your real calendar or only sends a booking link.
  7. How do you review what it said? You need a log of every call or message, searchable by patient, and a way to correct an answer so the mistake doesn't repeat.
  8. Who writes and updates its answers? If the answer is "you, in a settings page", count that time as part of the cost.

How do you test one before patients meet it?

Play the patient yourself for a week before you turn it on. Use made-up names and no real patient details.

  1. Write down the 20 questions new patients ask most. Your sent folder has them.
  2. Ask each one the way a patient would, with typos, half sentences and two questions at once.
  3. Add five it should refuse: a symptom, a dose question, a refill request, a lab result question and something urgent. Check that each one is handed off and that you get the alert.
  4. Try to book, reschedule and cancel. Then look at your calendar to see what happened.
  5. Message it at 10 pm and on a Sunday.
  6. Read the full log. Mark every answer that was wrong, vague or said something you would not say.

Then start small. Turn it on for new inquiries only, or after hours only, and read every conversation for the first two weeks. Widen it once a week passes with nothing you needed to correct.

Where does it go wrong?

  • It guesses. An assistant with no answer in its FAQ may make one up. Ask the vendor how it is stopped from doing that, then test it with a question you never wrote an answer for.
  • It over-shares in a text. A reminder that names the treatment puts health details on a lock screen. Our guide to HIPAA-safe texting covers what to leave out.
  • Nobody reads the handoffs. A handoff that lands in an inbox nobody checks is a missed message with extra steps.
  • It replies fast and the next step isn't ready. A quick answer only helps if the intake link goes out with it. See how fast to reply to a new patient inquiry.

Where we fit, and where we don't

We don't answer the phone. If most of your new patients call, you need a voice agent, an answering service or a person, and the questions above apply to all three.

What we do is the written follow-up around the EHR you already have. Assistants reply to inquiries by text, email and web chat in your practice's voice, send the intake and payment links, and send reminders. Patients are told an assistant helps the practice reply. Anything about doses, symptoms or refills is drafted and waits for your approval. For how that compares with hiring help, see what to hand a VA in a medical practice.

This is general information, not legal advice. Confirm BAA terms with each vendor, and check your state's rules on automated calls, call recording and telling people they are talking to an assistant.

Related: check the tools you already use with the HIPAA and BAA checker, and read is ChatGPT HIPAA compliant? before a general AI assistant goes anywhere near patient messages.

Common questions

What is the best AI receptionist for a medical office?

The one that signs a BAA, hands clinical and urgent questions to a person, tells patients it is an assistant and connects to the scheduler you already use. We don't rank vendors because products change fast. Use the eight questions above and test with made-up patients before you choose.

Is an AI receptionist HIPAA compliant?

It can be, if the vendor signs a BAA that covers the AI model and the phone or texting service behind it, and you set it up to keep health details out of texts. Without a BAA, it should not handle patient calls or messages.

Can an AI front desk answer medical questions?

It shouldn't. Questions about symptoms, doses, refills and results belong to the clinician. A good assistant says a person will reply, alerts the practice and, for anything urgent, gives emergency instructions.

Do patients need to be told they are talking to an AI?

Tell them. It is the honest thing to do and patients take it better than finding out later. Some states also have rules on automated calls and recording, so check yours. This is not legal advice.

Can an AI receptionist book appointments in my EHR?

Some can, for some EHRs. Ask the vendor whether it reads and writes your real calendar or only sends a booking link. For a solo practice, a booking link from your own scheduler is often enough.

Do I need a phone agent or a text and email assistant?

Look at where your last 30 new patients came from. If most called, start with the phone. If most filled in a form or sent a message, a written assistant covers more of them.