We sell an AI receptionist for dental practices, so treat this article with appropriate suspicion. But the fastest way to evaluate any vendor in this category, us included, is to ignore the capability list and interrogate the escalation rules. If you are still deciding whether you want an answering service or an AI receptionist at all, that is the earlier question.
A vendor who cannot tell you precisely what their system refuses to do has not thought carefully about deploying it in a dental practice. And in dentistry the cost of handling the wrong call badly is unusually high, because the calls that go wrong are attached to five-figure treatment plans and to clinical liability.
Here are the categories that should route to a human, and why each one is different.
1. Major treatment consults
Implants, full-arch and all-on-X cases, surgical extractions, endodontic treatment, full-mouth rehabilitation.
These calls should reach a person, and not because the AI cannot pronounce the words.
The reason is that a patient calling about an implant is at the start of a considered, expensive, often emotional decision. They have usually been thinking about it for months. They have questions about pain, about timeline, about whether it is worth it, and underneath those they are deciding whether they trust your practice with something significant.
That conversation is a relationship, not a transaction, and it belongs to your treatment coordinator or your doctor. Automating it saves a few minutes and risks a case worth many multiples of your annual software spend.
A practice we work with routes exactly this set to a human: all-on-X, implant consultations, extractions, endodontics, and other procedure scheduling all transfer to a named person on a direct line. The AI handles their hygiene, exams, and emergencies. It does not touch the big cases.
That is the correct division, and it is worth noting the practical detail that goes with it. When a transfer destination is configured, someone has to verify the number actually works. We have seen a configured transfer number turn out to be out of service during deployment testing, which would have silently dropped exactly these highest-value calls. Test the transfer path with a real call before go-live, every time, and re-test it when anything changes.
2. Anything clinical
Not "is this an emergency," which is triage against a fixed question set and is appropriate to automate. Anything past that.
Whether a symptom is serious. Whether a medication is safe with their history. What to do about post-operative pain, swelling, or bleeding. Whether a treatment plan is right for them. Interpreting an X-ray finding they half remember from their last visit.
These are clinical judgments. They belong to a licensed clinician, they carry liability, and no amount of model quality changes that. A system that will confidently answer a clinical question is not more capable, it is more dangerous.
The correct behavior is a clean handoff that does not make the patient feel dismissed:
"That's a question for the doctor rather than something I should answer. Let me get you to the right person."
Post-operative calls deserve specific mention. A patient calling three days after an extraction with concerns needs a clinician, promptly. That call should route immediately and it should be flagged, not queued.
3. Money disputes and treatment plan disagreements
A patient who is upset about a bill, disputing what they were quoted, or unhappy about what insurance paid is not looking for information. They are looking to be heard by someone who can actually resolve it.
An AI can capture the details and route them. It should not attempt to explain, defend, or negotiate. And it should never adjust a balance or make a financial commitment on the practice's behalf.
The same applies to a patient who has decided your treatment plan is wrong, or is asking for a discount, or is comparing your fee to a quote from down the street. These conversations decide whether the patient stays, and they need your team.
4. The genuinely angry patient
Distinct from the money dispute, and worth its own rule.
When someone is angry, the automated response that would be fine in any other context reads as institutional indifference. Everyone has been trapped in a phone tree while furious, and it converts a recoverable problem into a review.
The system should detect frustration and escalate quickly, with a short acknowledgment and no attempt to solve it first:
"I'm sorry, that sounds frustrating. Let me get you to someone on our team right now."
Ask any vendor how their system detects this and what it does. It is a revealing question, because it is genuinely hard and most have not addressed it.
5. Anything involving a minor's care where consent is unclear
A caller trying to schedule or ask about treatment for a child they may or may not have authority for is a judgment call about consent and custody. That is not a decision to automate.
6. Medical history that changes the plan
If a caller volunteers something like a recent cardiac event, a new anticoagulant, active cancer treatment, a recent joint replacement, or pregnancy, the appointment may need different handling entirely.
The AI should capture it accurately and flag it prominently. It should not decide what it means for the visit.
7. Anything the system is not confident about
The most important category, and the one vendors are least willing to discuss.
Every one of these systems has an uncertainty boundary. The question is what happens at it. A well-built system recognizes that it does not understand and routes to a human. A poorly built one guesses, confidently, and you find out from the patient.
Ask directly: what is the confidence threshold, what happens below it, and can you show me a call where it triggered? A vendor with no answer is telling you their system has no boundary, which is not a feature.
The dental-specific reason this matters more here
Escalation design matters in every industry. It matters more in dentistry for two reasons that compound.
Case concentration. A single full-arch case can be worth more than a year of everything else that came through the phone that week. In a business where value is that concentrated in a small number of conversations, mishandling the rare high-value call is not a rounding error. It is the whole quarter.
Clinical liability. Most industries deploying voice AI are not fielding calls where a wrong answer can hurt someone. You are. That raises the floor on what "acceptable failure" means, and it means the safety-relevant paths need to be tested rather than assumed.
The four questions to ask any vendor
- Show me your escalation list. Not the concept, the actual list of what routes to a human.
- What happens when the system is uncertain? Show me a real call where that fired.
- How do transfers work when the destination does not answer, and how do I find out it failed?
- What will your system never say, and how is that enforced?
If the answers are vague, the system probably does not have firm boundaries, which means you will discover them on a patient.
What this leaves for automation
Plenty, and it is most of the volume.
Answering on the first ring at any hour. Identifying the caller. Routine hygiene and exam scheduling. New patient intake and screening against your rules. Reschedules and cancellations. Emergency triage against your fixed questions and your eligibility policy. Insurance detail capture. Routing everything above to the right human with the context attached.
That is the large majority of your call volume, handled consistently, at 2am and during the lunch hour and on the fourth simultaneous call.
The goal was never a system that handles everything. It is a system that handles the routine reliably and knows exactly where to stop.
Telvana's escalation rules are configured per practice and we will walk you through ours line by line. If a vendor will not do that, it is worth asking why.
This article describes front office call routing policy. It is not clinical or legal guidance.
