What can an AI front desk safely answer for a medical practice?

Logistics only: hours, locations, parking and accessibility, which insurance plans the practice participates in, whether it is accepting new patients, what a first visit involves, what to bring, how to request records, and how to book. Every one of those is a fact about the practice rather than about a patient, which is the line that makes them safe, and together they are the large majority of what a medical phone line actually carries.

What must it never answer?

Symptoms, diagnoses, medications and interactions, test results, whether something is serious, whether somebody should come in, and anything about another patient. Those are refused before an answer is attempted rather than answered with a disclaimer, because a careful answer to a clinical question is still an answer to one. The reply says who will follow up and, for anything urgent, directs the patient to emergency services in wording a clinician has approved.

Can it ask patients about their symptoms as intake?

It can reasonably ask whether somebody is a new or existing patient, which insurance they hold, and whether the visit is routine or about a new concern. It should not collect symptoms by text. Clinical intake belongs in your clinical systems, where it is stored and handled appropriately, and a text thread is read on a bus and kept in a backup indefinitely.

What should an appointment reminder for a medical practice say?

The time and the place, and not the reason. Tuesday at 2pm is fine; your diabetes review on Tuesday is a disclosure about a patient's health sitting on a lock screen where anybody can read it. The same applies to the clinician's name where the specialty implies the condition. Reminders go only to patients who agreed to receive them, with a record of what they agreed to.

What should we ask a vendor about privacy?

Whether they will sign a BAA, where the data is held, who inside their organisation can see patient messages, how long messages are retained, and what happens to them if you leave. Ask all five in writing rather than accepting a general assurance of compliance, and have your compliance lead rather than your front desk read the answers.

An AI front desk for medical practices

Logistics answered at nine in the evening, every clinical question refused before an answer is attempted, and the two questions that account for most of a medical phone line. Plus the reminder wording that is itself a disclosure.

Sohail Safdar · November 4, 2026 · 6 min read

A medical practice's phone line carries two kinds of traffic, and they need opposite handling.

Most of it is logistics: hours, insurance, parking, are you taking new patients, what should I bring. None of it needs a clinician and all of it occupies the front desk.

The rest is clinical, and the correct behaviour for software there is to get out of the way completely, as quickly as possible, with the patient's own words preserved.

Getting the second part right is what makes the first part usable.

The two questions that dominate

"Do you take my insurance?" and "are you accepting new patients?"

Both are high volume, both have simple answers, and both have a trap.

The insurance question is really three questions. Are you in-network with my plan; will you bill my plan if you are not; and what will my plan actually cover for what I need. A practice that is out-of-network but files claims as a courtesy can answer "yes, we work with that plan" truthfully, and the patient will hear the answer to the first question. They arrive, are treated, and then have a conversation with billing that nobody wanted.

So the rule is narrow and holds: state which plans you participate in, from a list you maintain, and nothing about what a patient's plan will cover. Participation is a fact about the practice. Coverage is a fact about the patient, and it depends on their deductible, their remaining benefit and how a procedure codes.

The new-patient question's trap is staleness. The answer changes, sometimes monthly, and it is exactly the kind of fact nobody updates. An assistant confidently telling somebody you are accepting new patients when a panel closed six weeks ago produces a wasted registration and an annoyed phone call. Put it on the same weekly list as everything else that changes.

What it answers from your facts

Participating insurance plans, as the list, per the section above.
New patient status, and what a first visit involves.
Hours, locations, parking, accessibility.
Forms, what to bring, how to request records.
How to book, with the link.

Every item on that list is a fact about the practice rather than about a patient. That is the line, and it is a more useful test than any list: if answering requires knowing something about the person asking, it is not a logistics question.

What it never answers

Symptoms. Diagnoses. Medications and interactions. Test results. Whether something is serious. Whether somebody should come in. Anything about another patient, which in practice means anything from a shared household number about somebody other than the sender.

Two things about how the refusal has to work.

It must happen before an answer is attempted. A system that produces an answer and then appends "but please consult your physician" has answered a clinical question and added a sentence. A system that detects the category before the answering path runs has refused. Those look similar in a transcript and are completely different in the case that goes wrong.

The urgent branch needs clinician-approved wording. For anything that reads as an emergency the reply has to direct the patient to emergency services, and that sentence must be written by somebody qualified to write it rather than by a vendor or generated by a model. Have your clinical lead sign off the exact words.

Intake without clinical detail

It is reasonable to ask: new or existing patient, which insurance, routine or a new concern, preferred days.

It is not reasonable to collect symptoms by text. Clinical intake belongs in your clinical systems, and the four questions above are enough to route the enquiry and book the appointment. Those four live as qualifying questions in the practice's own knowledge, which is the only place the assistant reads an answer from.

The practical argument alongside the regulatory one: a text thread is read on a bus by whoever is beside the screen, and it sits in a backup indefinitely. The four questions produce no record that matters if a phone is lost.

Reminders: the wording is itself a disclosure

An appointment reminder should carry the time and the place, and not the reason.

"Reminder of your appointment on Tuesday at 2pm at Riverside Medical, 14 Mill Street."

Not "your diabetes review", not "your follow-up with the oncology team", and be careful with the clinician's name where the specialty implies the condition. A reminder is read on a lock screen, which is a notification board anybody nearby can see.

The same applies to a confirmation and to a reschedule message. And reminders go only to patients who agreed to receive them, with a record of the wording they agreed to.

Shared phone numbers, which are common

A household phone number is ordinary, and it is the mechanism behind most accidental disclosures in a practice.

Two habits worth having. Confirm the number belongs to the patient rather than to the person who gave it. And honour a request to use a different channel or a different number without making it difficult, because a patient asking for that usually has a reason they would rather not explain.

What to ask a vendor, in writing

Will you sign a BAA?
Where is the data held?
Who inside your organisation can see patient messages?
How long are messages retained, and can we change that?
What happens to them if we leave?

Ask all five, get the answers in writing, and have your compliance lead read them rather than your front desk. A general assurance of compliance is not an answer to any of the five.

Not legal or compliance advice. HIPAA and texting patients covers the practical rules in more depth, and your own counsel covers your own obligations.

What to measure

Calls and messages answered without a person, which is the volume justification.
Clinical escalations, and how long each waited for somebody. This is the number that matters most and it is about your staffing rather than the software.
New patient enquiries that became registered patients.
Questions it could not answer, read weekly. In a medical practice this list is unusually useful, because it is written in patients' own words and tells you what your website does not say.

In Frontiva

Answers come only from structured facts and FAQs your practice approved, with no match logged as a gap rather than guessed at. The guardrails for prompt injection and sensitive clinical topics run before any answer is looked up, as a deterministic pattern check with no model and no network call in it, so they cannot be argued out of by the message they are checking. Your own escalation keywords sit alongside them, in your patients' own phrasing.

A scored readiness suite runs before going live, and its two emergency checks are required rather than averaged: an agent that cannot get an urgent message to a person cannot be switched to unsupervised answering however well it scored on everything else. What an AI readiness test should cover explains why that distinction matters.

Approval mode is the default and is where a medical practice should expect to stay longer than most trades: the assistant drafts and a person sends. How much patient history the assistant may read is a setting with a narrow default, and the widest option is one you choose deliberately.

Reminders carry the time and place from the appointment rather than its reason, go only to patients with recorded consent, and defer out of quiet hours in the recipient's own timezone.