Product · AI voicemail triage

AI voicemail triage for dental offices

Your voicemail box fills up overnight, every night. Oriva reads what is in it and ranks it by how soon someone needs to act, so your team opens a sorted queue instead of a stack.

What the first hour actually looks like

One person is checking in a patient, taking a call, and working last night's messages, all before nine.

Call-tracking analyses of dental practices put roughly a third of inbound calls as unanswered. That is not a staffing failure. It is arithmetic. A front desk that is seating a patient, chasing an insurance question, and handing back a treatment plan cannot also be on the phone, and the calls that land in those minutes go to voicemail.

So does everything that arrives after six. By morning the box holds a confirmation, a reschedule, a billing question, a daughter calling about her mother, and one message about swelling. In a plain list they all look identical. The only way to find the urgent one is to play every message in the order it was left, while the phone is already ringing again.

About one in seven appointments is a no-show, according to published no-show research. A cancellation left at 9pm is the kind of gap you can still fill before the morning, and in an unsorted box it sits behind messages that could safely have waited.

Triage, not just answering the phone

Most of this category sells a way to stop voicemail from existing. Oriva starts from the box you already have.

Answering is about capture. Triage is about order. Even a practice that answers the phone well ends the night with messages waiting, and the question at 7:52am is not how many there are. It is which one to open first, and what to do once it is open.

Answering the callTriaging the message
What it solvesFewer calls go unheardThe right message gets worked first
What comes backA recording, or a booked slotA transcript, an intent, a one-line summary, the callback number, and any flags
An emergency at 9pmDepends who is on the lineFlagged, lifted out of the queue, and shown on its own until someone acts
When the read is unclearThe message still has to be sorted by handParked for review with no suggested action
What your team opensEvery message, in the order leftA queue already ranked by urgency

How the queue is ordered

Four intents, an emergency flag, and an honest bucket for everything that does not fit. Here is the whole rubric, in public.

Emergency, first

Swelling, bleeding, a knocked-out or broken tooth, severe pain. It leaves the queue, gets its own card, and stays there until a person acts on it.

Cancel or reschedule, next

These are the time-sensitive ones. A cancellation frees a slot you can still refill today, and a reschedule carries the time the caller asked for, in their own words, onto your waitlist.

Everything else

Running late, billing, insurance, a prescription, a post-op question, a wrong number. It routes to a person with a summary, and running late is never filed as a cancellation.

Confirmations, lower

The patient is keeping the appointment, so nothing is at stake if it waits an hour. One click clears the message and the confirmation is recorded.

Needs review, never blocking

Below the confidence line, more than one request in a single message, or a poor recording. Oriva shows no suggested action at all, and it never sits in front of work your team can act on.

Urgency, not diagnosis. Oriva ranks messages by how soon a person should look at them and surfaces a likely cause for clinician review. It is information to support the clinician, not a diagnosis. Emergency detection runs twice, once through the model and once through a fixed list of clinical terms, so an urgent message is still flagged when the model is unavailable. A bare "as soon as possible" with no pain or injury does not count as an emergency.

What your front desk actually sees

One message, start to finish.

A message comes in at 9:14pm. The next morning it is already above the confirmations, and the card reads: Reschedule, 92%. Under that, one line. Patient wants to move Thursday's cleaning to next week.

The detail sits underneath. Caller name and callback number, extracted as heard, with missing digits left missing rather than invented. The time the caller asked for, quoted the way they said it: next week, morning. When the caller is not the patient, the card says so in a sentence, such as Ruth calling for her mother Eleanor, so your team can confirm authorization before anything changes.

The original recording sits one tap away, under a standing note that intents are AI-suggested and can be wrong, so verify with the recording before acting.

Then four buttons. Confirm, cancel, reschedule, dismiss. You press one, and that is the whole interaction. Nothing changes until you do, and the board records who acted and which classifier produced the read. Press reschedule and the patient joins your waitlist with the window they asked for attached, so the next opening can go to someone who already told you when they wanted it.

About that 92%. The percentage is the model's own confidence in its read, not a measured accuracy rate. Below the 0.75 line, or when a message carries more than one request, or when the recording is poor, Oriva shows no percentage and no suggestion at all. It parks the message for a person.

What Oriva will not decide for you

The defaults are conservative on purpose, and the conservative ones are the ones that ship.

Manual is the default

Automation is a toggle you turn on, not one you turn off. Out of the box, every state change is a click by a person.

Emergencies stay with a human

Even with automation on, a message flagged as an emergency is excluded from it by rule rather than by setting.

It abstains instead of guessing

Low confidence, more than one request, or a poor recording, and Oriva offers nothing. Declining to guess quietly is the point.

Only the unambiguous is automated

With automation on, the agent acts on single-intent, non-emergency, high-confidence messages, one at a time. Everything else queues for you.

Every action is on the record

Each message carries which classifier read it and who acted on it, and each status change writes an audit entry.

Where it fits with what you already run

Triage starts at the recording. The ingest is designed to take the audio your phone system already produces, transcribe it, and post it to the board, so patients keep calling the number they have always called and nothing about how they reach you changes on day one.

The rest of the front desk is where Oriva already runs. Veronica answers live calls around the clock in 30 or more languages, which matters when more than one in five US residents speaks a language other than English at home, according to the US Census. The smart questionnaire asks about seven questions in under two minutes, roughly three times more efficient than a standard forms stack. The triage engine turns what the patient said into a structured, pre-visit picture for a clinician to review.

If you want the longer version of the problem, we wrote up a day spent shadowing a dental front desk and what the front-desk software stack actually costs in retyping.

What is not built yet. Writing decisions back into Dentrix, Open Dental, or Eaglesoft is in active development, so today a decision on the board is one your team then enters in your practice management system. The waitlist policy is designed to offer an opening to one patient at a time, in priority order, on a short exclusive window, and the messaging that carries those offers is planned rather than shipped. We would rather list that here than let a demo imply it.

Questions we get

What is dental voicemail triage?

Voicemail triage is the step between a message being left and a person acting on it. Each message is transcribed, read for what the caller is asking for, summarized in a line, and ranked by how soon someone should respond. Urgent messages come out of the queue and onto their own card, routine ones queue behind them, and anything unclear is parked with no suggested action. It is sorting rather than diagnosis: the read exists to help your team decide, not to decide for them.

What happens to the voicemails patients leave overnight?

Each message is transcribed and classified as it lands, so by the time your team opens the board the whole night is already sorted. A message that reads as an emergency is lifted out of the queue onto its own card and stays there until a person acts on it. Cancellations and reschedules come next, because those are the gaps you can still refill today. Confirmations, which cost nothing by waiting, sit further down.

How does it tell an urgent message from a routine one?

Two checks run on every message. A model reads the transcript, and a fixed list of clinical terms is checked separately, so swelling, bleeding, a knocked-out or broken tooth, or severe pain still raises the flag even when the model is unavailable. Urgency words on their own are not enough, which is why an ASAP about a bill reads differently from an ASAP with pain. When the read is uncertain, or one message carries more than one request, Oriva offers no suggestion at all and parks it for a person.

Who decides what happens to an urgent message?

Your team does. Nothing changes on its own: every state change on the board is a click by a person, and messages flagged as emergencies are excluded from automation by rule rather than by setting. When an urgent message comes in, someone on your team chooses how it is handled, including whether to hand the callback to Veronica. That handoff is always a choice a person makes.

How does this work with the phone system we already have?

Triage starts at the recording, so patients keep calling the number they have always called. The ingest is designed to take the audio your phone system already produces, transcribe it, and post it to the board. Writing decisions back into Dentrix, Open Dental, or Eaglesoft is in active development, so for now a decision made on the board is one your team then enters in your practice management system.

Is this HIPAA compliant, and what happens to the recording?

Oriva is built to be HIPAA compliant. Data is encrypted in transit and at rest, access follows least-privilege principles, and we sign a Business Associate Agreement with practices using Oriva for protected health information. The recording is kept alongside the transcript so a read can always be checked before anyone acts on it, and every status change writes an audit entry. Formal third-party attestations such as SOC 2 are on the roadmap rather than in place, and we would rather say so here than on a sales call.

Type a message and watch it get sorted

Write a message the way a patient would leave one, and see exactly where it lands. Invented messages only, please: the demo runs on synthetic data and is not a place for patient information.