Dental practices and the limits on what an AI can answer

Everything else in a dental phone buildout is scheduling logistics. This one boundary is where practices get into trouble, and it gets crossed for the friendliest possible reason: a patient asks what a crown will cost them, and the system tries to be helpful.

I am not a lawyer and this is not legal advice. It is the operating shape of the problem, and what to put in front of your counsel and your provider before anything answers a live line.

How many dental calls are actually being lost?

Enough that the phone is usually the largest untreated problem in the practice.

Patient Prism scored 11,552,668 patient calls across 8,280 dental and DSO locations in calendar 2025. Of every 100 calls, 69 connected with the front desk and 31 hung up before reaching an agent. CallRail's benchmark across 1.1 million conversations put the health care missed call rate at 32%, against 28% in legal, 14% in home services and 9% in real estate.

That spread is the point. Health care loses roughly three times the share of calls that real estate does. Anyone quoting you a single cross industry number is selling something.

The 31% is a specific kind of failure, and the distinction changes what you should buy. It is abandonment before reaching an agent, meaning hold time and routing, not "nobody picked up." That is a queueing problem, and queueing is what automation is genuinely good at.

Two caveats belong with it. This is a Patient Prism customer cohort, so the true market figure is plausibly worse than 31%. And of every 100 calls, only 34 are real booking opportunities while 35 are existing patients, vendors and wrong numbers. A third of your phone traffic was never going to become anything, which is why "calls answered" is a bad primary metric.

Where those calls come from also matters. Across 5,022,887 attributed calls, 90% came from the Google ecosystem, the Google Business Profile alone at 54%.

What can a dental AI actually say?

More than practices assume, as long as everything it says is verbatim and provider approved.

The legal basis matters, because it is what makes an AI viable at all. The Doctors Company, the largest physician-owned malpractice insurer in the US, states in its telephone triage guidance that unlicensed personnel should never be allowed to triage by telephone or provide telephone advice, and that failing to follow written advice protocols may be treated as practising dentistry beyond scope, with the clinician held vicariously liable. Don Balasa, CEO and legal counsel of the American Association of Medical Assistants, draws the line in the same article: triage requires independent clinical judgment and cannot be delegated, while non-triage communication, meaning following provider-approved protocols or decision trees in the verbatim receiving and conveying of information, can be.

That is the design specification. An AI may operate a provider-approved verbatim tree. The moment it generates novel clinical language, which is what a language model does by default, it has exercised independent judgment.

Inside that boundary it can confirm hours, location, and which providers are accepting new patients. It can state, as a fact about your practice, whether you are in network with a named carrier. It can collect the plan name, group number and member ID for the treatment coordinator to verify, and it can schedule, reschedule and cancel inside rules you set.

The word doing the work is verbatim. A decision tree with approved language at every node is a tool the practice controls. "Be helpful about insurance" is an improvising machine wearing your practice's name.

Get the provider to sign the tree. Not the office manager, and not the vendor. The Doctors Company is explicit about where liability lands: the practitioner is responsible for the instructions given to their patients by the service.

There is one case where a scripted tree is clinically valuable rather than merely safe. The International Association of Dental Traumatology's 2020 guidelines on avulsion of permanent teeth explicitly contemplate instruction by telephone to people at the emergency site: handle the tooth by the crown, replant immediately if possible, otherwise store it in milk, saline or the patient's saliva. Most periodontal ligament cells are non-viable after 30 minutes of dry time and likely non-viable beyond 60. Delivered as a provider-approved verbatim protocol while paging the on-call dentist, that is a legitimate use. Generated on the fly by a model, it is not.

What must it never say?

Two things, and they are close enough together that people conflate them.

It must never say a procedure is covered. In network is a statement about a contract between your practice and a carrier. Covered is a statement about a patient's benefits, remaining maximum, waiting periods, frequency limits, and downgrades. Only a verification produces that, and unlicensed personnel cannot produce it on the phone.

It must never quote an expected out of pocket amount. Not a number, not a range presented as what this patient will pay, not "usually around." The moment a patient hears a figure, that figure is what they believe your practice told them.

Beyond those two, the ordinary clinical line applies. No triage, no advice about pain, no opinion on whether a symptom can wait. A caller describing facial swelling gets a human, immediately. The Doctors Company's own closed claims analysis found that miscommunication contributes to over 30 percent of adverse patient events in the office setting, which is why the rule is written that strictly.

Two rules from the same guidance belong in the tree. End every call by telling the patient when to call back or seek emergency care. And handle repeat callers explicitly: a second unresolved call goes to the practitioner, a third requires an in-person examination.

California sharpened the picture further. AB 489, in force since 1 January 2026, makes every healing arts title protection provision enforceable against an AI system, and states that each use of a prohibited term is a separate violation. No "Dr. Ava," no "I'm one of the hygienists," no DDS, DMD, RDH or RN in the persona, greeting, SMS signature or marketing page. AB 3030 separately requires an AI disclaimer and human contact instructions on generative patient communications, including verbal ones. Recording adds a third layer, since California requires consent before recording begins. All three belong in one conversation with counsel.

Restaurants have exactly one boundary of the same kind, and the hard line on allergen questions is worth reading because the reasoning transfers cleanly.

Why is insurance the hardest problem in the category?

Because it is simultaneously the most asked question and the least answerable one.

Patient Prism categorized 1,163,398 opportunities where a patient did not book. Consideration and timing accounted for 45.2%. Financial barrier accounted for 34.4%, and inside that cluster insurance outweighs price by more than 11 to 1. Service fit was 10.3% and connection failures 9.3%.

Read that carefully. The single largest actionable barrier is insurance, and it beats price by an order of magnitude. The most-asked hard question is the one carrying the highest liability.

What you can do is change what the answer feels like. The difference between a compliant script and a good one is entirely in the wording.

Bad: "I am not able to discuss coverage."

Good: "I can see we are in network with your plan. Our treatment coordinator verifies your specific benefits before anything is scheduled, so you will have exact numbers in writing before you commit. Can I take your plan details now so she has them ready?"

The second one is compliant, it collects the data you needed anyway, and the patient feels handled rather than blocked. That is the difference between a script written by a lawyer and one written by an operator with a lawyer's list.

Then the escalation has to work. A transfer that rings a desk nobody is at is worse than no transfer, and how good implementations handle the handoff to a human is what practices most often leave until after launch.

What is the biggest recoverable pool, and why is it also the riskiest?

Followup, and it is almost entirely un-worked.

Here is the arithmetic on the Patient Prism funnel, and you can rerun it with your own call volume. Of every 100 calls, 21 book and 13 walk away without booking. Only 4.8 of those walkaways receive a followup call, and only 0.7 come back and book. Across the 2025 cohort, 890,621 patients walked away and never got a callback at all. The cohort-average followup rate is 38%. The best operators exceed 80%.

Apply it to a practice taking 1,200 calls a year. That is roughly 156 walkaways, of which about 59 currently get a callback and about 8 return. Lift the followup rate from 38% to 80% and you are calling 125 people instead of 59. Even at the cohort's poor conversion on callbacks, that is a meaningful number of chairs.

Now the part vendors do not mention. An AI answering an inbound call the patient placed is not a robocall. An AI placing outbound callbacks is making artificial-voice calls, and the FCC declared in February 2024 that AI-generated voices are artificial under the TCPA and require prior express consent of the called party.

There is a genuine advantage for a dental practice here, and it has a hard edge. Appointment reminders, confirmations, recall notices and post-op instructions from a covered entity are health care messages and get relaxed consent treatment. The moment the same call pitches a whitening special, it is telemarketing and needs prior express written consent. Keep transactional and promotional campaigns on separate consent records. Damages run at $500 per message, trebled to $1,500 for willful violations.

Do not build the outbound program off this article. Build it off your counsel's reading of your consent records, and build the inbound side first.

What about calls that come in after you close?

This is where a dental AI earns most of its money, and it is also where the boundary is easiest to hold.

Nobody expects a coverage answer at nine at night. They want to book, or they are in pain and need to know what to do. Booking is scriptable. Pain is a routing decision your provider defines in advance, usually to an emergency line.

Worth sizing before you buy: what after hours calls are actually worth depends on your new patient value and how many of those calls exist. There is no published national figure for the value of a new dental patient, so that number has to come off your own practice management system.

What to do this week

Pull your call log and count calls that rang out or hit voicemail during business hours. That number, not the vendor's number, is your case.

Write the coverage script line and the pain routing line exactly as you want them spoken, plus the end-of-call instruction about when to call back or seek emergency care. Take all of it to the provider for signature, and to your attorney for the AB 489, AB 3030 and recording questions.

Then call your own office three times as a new patient and try to get a price out of whoever answers. If a human gives you one, an AI would have too, and you have found the real problem.

While you are looking at new patient flow, the five things a first time visitor needs above the fold usually determines whether they call at all, and service area pages determine whether they find you.

Questions about the boundary in your practice?

Email me at eric@seod.com with the five questions your front desk hears most often. I will sort them into three buckets, safe to script, safe only with exact approved wording, and never say, plus the handoff line I would use for each in the third bucket.

I will flag which items need your attorney or your provider rather than my opinion. That flagging is the useful part and it costs you nothing.

More context sits in the AI phone and lead response library.