How AI Hands Difficult Calls to Your Team

The Demo Never Shows You the Transfer

Every AI receptionist demo you will sit through is a booking. A patient calls, the agent is warm and quick, an appointment appears in the schedule, and everyone in the room nods. It is a fair demonstration of the ordinary case, and the ordinary case is most of your call volume.

The calls that decide whether you keep the product are the other ones. A patient is in pain and frightened. A parent is angry about a bill. Someone asks a question about a treatment plan that no software should answer. On those calls the agent has exactly one job, which is to get the caller to a human quickly and without making them start over.

That moment is called a handoff, and it is the least documented part of this category. Vendors list it as a bullet. Almost nobody publishes how theirs decides, what it carries across, or what it does when the transfer fails.

This page sets out the four things that have to be true for a handoff to work, what each vendor actually publishes about its own, and the failure mode that is missing from every feature list we read.

Four Things That Have to Be True

A handoff is not one feature. It is four, and a product can do three of them well and still lose you a patient.

The trigger. Something has to decide that this call needs a person, and it has to decide early. A trigger that fires after four minutes of the agent trying to help is technically a handoff and practically a complaint.

The context. Whoever picks up needs to know who is calling, what they have already said, and why the call was escalated. Without that, the patient repeats themselves, which is the single thing they were promised would not happen.

The destination. There has to be a specific person or phone that receives the call, and it has to be correct at that hour. The destination at two in the afternoon and the destination at two in the morning are not the same, and most published descriptions quietly assume the first.

The fallback. The destination will sometimes not answer. What happens in the next ten seconds is the difference between a recovered patient and a caller who hangs up and dials the practice down the road.

Hold each vendor’s description against those four. Most cover the first two.

What the Vendors Publish

We read each of these pages on 16 September 2026. Where a vendor does not publish something, this page says so rather than filling the gap from a third-party blog, because the third-party blogs in this category are almost entirely written by competitors.

Adit is the most specific of the platform vendors. Its AI Front Desk page describes “warm transfers with context,” and states that when a caller needs a human, the AI transfers with a full summary so the team picks up where the conversation left off, with no need for the caller to start from the top. The same page says the agent handles routing the ones that need a human alongside booking and rescheduling. Adit publishes scale figures next to this: 5,000 or more practices, 2.4 million calls handled every month, 200,000 appointments booked monthly, and a 4.8 rating across G2, Capterra and Software Advice.

Viva AI publishes a section headed “Knows when to get a human.” It describes two distinct paths, which is more than most vendors separate out: a complex case or an explicit request for a person triggers an instant transfer to the right staff member, while emergencies trigger what it calls HIPAA-compliant alerts. Viva also states SOC 2 Type II certification and a signed BAA for every practice.

Weave lists “seamless handoff to staff” among its AI Receptionist bullets, and describes the receptionist as handling incoming calls directly and intelligently routing complex inquiries to staff, completing tasks over voice or text within a unified conversation thread. Weave states that over 35,000 locations use it. The mechanism behind the routing is not published.

Arini publishes no description of its escalation or transfer behaviour on its homepage as read on that date. It publishes production and call-volume figures, including $39,000 in average monthly production and 1,600 or more calls handled per month on average, and case-study figures for named groups, but the handoff is not among the capabilities it describes there.

Dentina publishes a detailed list of the scheduling rules its agent reads before booking, covering provider restrictions, operatory constraints, appointment-type rules, procedure-specific blocks, provider hours and lunches, family booking, same-provider rescheduling, age-based appointment types, existing-patient matching and duplicate-record prevention. Its homepage carries “30-day free trial, cancel anytime, no setup fee,” a live counter reading 2,800,000 or more calls handled, 2,000 or more dentists and DSOs, and a 100 percent call answer rate described as a median across more than 1,000 practices over a trailing 90 days. Escalation is not described on that page.

Smith.ai has published the most technically detailed account of handoffs we found anywhere in this category, in a blog post by Maddy Martin dated 16 December 2025. It sets out confidence thresholds that activate a handoff when AI certainty falls below 60 to 70 percent with a hard floor at 40 percent, negative sentiment detection, explicit customer requests such as asking to talk to a person, complex issue identification, conversation loop detection, regulatory and security requirements, and urgency indicators. It also specifies a minimum viable context payload: complete conversation history with timestamps, collected customer data, synchronised profile information, and metadata such as sentiment scores and intent classifications.

Smith.ai is a horizontal business answering service rather than a dental product, and that framework is written for contact centres. But it is the clearest public statement of what a handoff rule set contains, and it is a fair standard to hold a dental vendor to.

The Distinction That Changes the Question

There is a difference between vendors that is easy to miss and that changes what “hand off to a human” even means.

Smith.ai and Ruby employ human agents. When their AI escalates, it escalates to their staff, who pick up on your behalf. Ruby is a human virtual receptionist service rather than an AI one, staffed entirely in the United States.

We do not do that, and neither do the dental-specific AI vendors above. When our agent escalates, it escalates to your team. There is no room of people behind us waiting to take the call.

That is not a weakness, but it is a real difference and it changes what you have to configure. If you buy a service with human agents, the destination problem is theirs. If you buy a dental AI receptionist, including ours, the destination is a phone number you nominate, and if nobody at your practice is holding that phone at eleven at night, the handoff has nowhere to go.

Decide which of those two products you are buying before you compare escalation features, because the feature means something different in each.

A related gap catches practices with bilingual patients: an escalation that hands a Spanish-speaking caller to an English-only colleague. See can AI answer your Spanish-speaking patients.

The Failure Nobody Lists

Here is the one that is missing from every feature list we read, including, until recently, our own.

The transfer is attempted. The on-call phone rings. Nobody picks up.

This is not an edge case. It is a Saturday, the dentist on call is driving, and the phone is in a bag on the back seat. The agent has done everything right. The caller is now listening to an unanswered ring, and the product has produced a worse outcome than a plain answering machine, because at least the answering machine takes a message.

Ask every vendor you are evaluating the following, in these words, and write the answer down:

When the escalation destination does not answer, what happens next, and how do I find out that it happened?

The answers that satisfy are: the call falls back to a second destination, or the caller is offered a message that reaches someone who is actually awake, and either way the event appears somewhere you will see it the next morning without going looking. The answer that should worry you is any version of “it rings through.”

A related question, which separates the careful vendors from the rest: does an attempted but failed escalation get recorded differently from a completed one? If both are logged as escalated, your reporting will tell you the system is working on exactly the nights it is not.

What Ours Does

Our agent triages urgent situations and then follows the escalation protocol your practice configures. You choose one of three behaviours: forward the call to an on-call dentist, send an urgent SMS to a named phone, or read the caller your own emergency instructions.

Every escalation carries a transcript and summary, so whoever picks up has the conversation rather than a name and a callback number. That is the context requirement, and it is the part we would hold ourselves to hardest, because it is the part patients notice.

On the destination question we are plain: it is a number you nominate, and choosing it is your configuration work, not ours. On the trigger, the protocol is applied the same way on every call, which is the actual argument for a system over a tired person at five to six on a Friday.

We answer in under a second with no hold queue and no phone tree, and handle unlimited simultaneous callers, so an escalation is never competing with a queue behind it. Calls are unlimited on our $399 a month flat plan, with the first month free and a one-time $500 setup fee payable up front, so a month with a lot of difficult calls costs the same as a quiet one. That matters more than it sounds: on per-minute or per-credit pricing, a long call that ends in a transfer is a call you paid for twice.

Five Test Calls Before You Sign

Configuration is the work that determines whether any of this holds. Run these during your trial, from a mobile the agent does not recognise, and time them.

1. Ask for a person immediately. Say that you would like to speak to someone in the first sentence. Count the seconds until a human voice arrives. Anything past thirty is a patient you would have lost.
2. Describe an emergency. Use the words a frightened patient would use rather than clinical ones. Check that the escalation fires on the description and not on a keyword.
3. Be difficult without being urgent. Complain about a bill. This should escalate on sentiment, not on content, and it is the trigger most likely to be missing.
4. Call the after-hours destination when you know nobody will answer. This is the fallback test, and it is the one nobody runs. Then check what the log says happened.
5. Escalate, then ask the person who picked up what they could see. If they had to ask the patient to repeat anything, the context payload is not good enough.

Then read the transcripts the next morning. The gap between what you think you configured and what the agent actually does lives in those transcripts.

Where This Sits

If you are still deciding between vendor shapes rather than testing one, the five facts that actually decide it is the better starting point, and our buyer’s guide covers contract terms and the scorecard. For the specific case of urgent calls outside opening hours, how AI handles dental emergencies after hours goes further into the clinical boundary than this page does.

FAQs

Does the patient know they are being transferred? They should. A transfer that happens silently, where the voice simply changes, reads as a trick. Ours tells the caller what is about to happen before it happens.

Can we set different destinations for different times of day? That is the question to ask any vendor, and it is the one that most often exposes a shallow implementation. A single destination applied around the clock is not a configuration, it is a default.

What if we do not have anyone on call? Then say so during configuration and choose the instruction path instead, where the agent reads the caller your own guidance. That is a legitimate choice for many practices. What is not legitimate is nominating a phone that nobody is holding, which produces the silent failure described above.

Is a warm transfer different from a forward? Yes. A forward moves the call. A warm transfer moves the call and the context together. Adit uses the term warm transfer explicitly on its AI Front Desk page as read on 16 September 2026. When a vendor says only transfer, ask which of the two they mean.

Sources

Vendor claims on this page were read on 16 September 2026 from each vendor’s own pages: the Adit AI Front Desk landing page, the Viva AI homepage, the Weave AI page, the Arini homepage and the Dentina homepage. The Smith.ai handoff framework is from a Smith.ai blog post by Maddy Martin published 16 December 2025. Vendor pages in this category change frequently, so treat every figure above as accurate on the date read and re-check before relying on it.

---

Related reading: once calls are being handed over, decide who’s accountable when the AI books it wrong.