Delivery journal · The context behind the implementationDiscuss a similar challenge ↗
Case study, United States

AI voice agent for dental practices that books into the practice's own schedule

We built an AI front desk for a US startup: a voice agent that answers, finds real openings and books them in Open Dental and other scheduling systems, plus an AI scribe, voice intake and owner reports. It serves dental practices and is designed for behavioral health practices too.

November 22, 2026 · 18 min read · Jitendra Choudhary, CTO & Co-Founder

AI voice agent for dental practices on a laptop and phone: the call overview with a caller detail panel, and the callback queue

Shared with the client's permission. The client, its practices and its people are not named, and every screen uses sample data. We publish a client's work only with their written approval. If you work with us, your system stays private unless you say otherwise.

This was an AI agent for healthcare build from day one: voice, scheduling, documentation and the screens around them.

Where a dental or therapy practice loses patients: the phone

Most new patients still call. A small practice has one or two people at the front desk, and they are also checking patients in, taking payments and chasing insurance. Calls that ring out after 5 pm, at lunch or during a busy morning often go to a competitor.

Missed calls after hours

Evening and weekend callers reach voicemail and rarely leave one.

Costs the owner: new patients who book somewhere else.

Front desk overload

The phone rings while someone is checking in a patient.

Costs the owner: hold times, and staff who burn out.

Scheduling across providers

Cleanings go to hygiene columns, exams to doctors, new patients to blocked times.

Costs the owner: wrong-length visits and empty chairs.

Clipboard intake

Paper forms filled in the waiting room, then retyped.

Costs the owner: late starts and typing errors.

Notes typed after hours

Dentists and therapists finish charting at night.

Costs the owner: fewer visits per day, and tired clinicians.

Recalls that slip

Patients overdue for a cleaning or a follow-up session are not called back.

Costs the owner: steady revenue that quietly stops.

No dollar figures from us: we did not measure the client's revenue. The client's team reports more than 1,100 after-hours calls handled and about $50K in revenue opportunity in 6 weeks at one deployment. That is their estimate, and we have not verified it.

Behavioral health practices have the same front desk problem, with longer sessions and recurring weekly visits. The platform was designed so the same agent and adapters can serve them; those connections are planned.

Why the product started

The client, a US startup building with practicing dentists, wanted one AI layer that sits on top of the scheduling system a practice already uses. No practice wants to switch systems to get a phone agent. So the agent had to read and write the practice's own schedule, the same one the front desk sees.

  1. 1

    Oct 2025

    First commit: portals, sign-in, intake

  2. 2

    Winter

    AI scribe, patient summaries, Open Dental adapter

  3. 3

    Spring

    Voice agent and per-practice scheduling rules

  4. 4

    Summer

    Voice agent dashboard, sync partner adapter

  5. 5

    Live

    Practices answering real calls

Dates from the commit history, read Nov 22, 2026. Season labels group months of work.

What the platform connects to

Integration diagram: Retell, Telnyx and Deepgram for calls; OpenAI and AWS Bedrock for models; Stripe, Sentry and AWS for operations; Open Dental, NexHealth, Dentrix, Curve and Denticon for scheduling; SimplePractice and TherapyNotes plannedOpen full size

Five people, one front desk: what each one does now

Screens below show the delivered product, restyled. We kept its layout, fields and words. Concepts appear only in "What comes next". Sample data only; no patient data.

The patient who calls

Calls at 7:40 pm about an overdue cleaning. Nobody is in the office. The agent answers on the second ring.

  1. Is recognized

    The caller's number is looked up before the greeting ends.

  2. Spells their name

    First, then last, read back letter by letter. Date of birth read back.

  3. Hears two real times

    From the live schedule, for the right visit type and length.

  4. Is booked

    The agent confirms only after the schedule says yes.

  5. Fills the form by voice

    Speaks medications and allergies; checks what was filled in.

Patient intakes list on a phone with review statuses

Intakes. Every form, with what still needs review.

Callback queue on a phone, oldest first, with the reason for each call

Callbacks. Who to call, why, and how long they waited.

Voice agent overview on a phone with the callbacks banner

Staff on the go. The overview, on a phone.

The front desk

Arrives in the morning to a list, not a voicemail box. Every call has a summary, an outcome and, when needed, a callback with the reason.

  • Overview: calls, callbacks and booking rate against the week before.
  • All calls: recording, transcript, summary and staff notes in a side panel.
  • Callback queue: who to call, why, and how long they have waited.
Voice agent overview with weekly numbers, the call list and the caller detail panel

Voice agent overview. Sample data.

All calls list with call detail panel: recording, summary, action needed and transcript

All calls, with the call detail panel.

Callback queue with the caller detail panel, transcript and held slot

Callback queue. Each callback shows the held slot and its expiry.

The dentist and hygienist

Start Scribe at the chair. By the time the patient is at checkout, the SOAP note and a plain-language summary are waiting for review.

Dentist dashboard with the patient queue and a note ready for review

Patient queue: Not started, Ready for review, Pending Summary, Signed.

Dental AI scribe recording with a live transcript and a draft note

Scribe recorder: live transcript beside the draft note.

The therapist (designed, planned)

The same pieces fit a therapy practice: the agent books the next weekly session, the scribe drafts a DAP note, and measures such as the GAD-7 come in through the portal. If a caller or a message mentions a safety concern, the agent does not counsel; it gives the crisis line and flags the clinician.

Planned behavioral health intake with PHQ-9 and GAD-7 scores and a therapist match

Planned module, shown as a design: intake with PHQ-9 and GAD-7.

Voice agent stats: last 24 hours, booking funnel, agent performance and call heat map

Stats. Sample data, not client results.

The practice owner

Wants three answers: how many calls did we catch, how many became visits, and what is still waiting on my team. The owner also decides what the agent may do.

  • Stats: calls, booking funnel, first-call resolution, when callers reach out.
  • Notifications: every booking, reschedule and cancellation the agent made.
  • Incident reports: anything the agent got wrong, tracked to a fix.
  • Staff & Access: who sees what, and the BAA, roles, two-step sign-in and the audit trail.
Notifications list with a cancelled appointment and what the agent did next

Notifications.

Incident reports with severity, status and an update timeline

Incident reports.

Staff and access with roles, status and access checkboxes

Staff & Access.

How the voice agent works, step by step

The agent runs on Retell. Its prompt, tools and analysis fields live in the code, one set per practice, pushed from a single source. Every fact the agent says about the schedule comes from one of our tools; the prompt forbids it to invent a time.

Retell call workflow: call start hook, greeting, red-flag check with 911 or warm transfer, then book, reschedule, cancel, question or other branches with each tool call, and the post-call analysisOpen full size

Purple boxes are tool calls. Each tool waits up to 15 seconds (5 for logging a callback) and the agent says "One moment while I pull that up" meanwhile.

The tools the agent can call

ToolWhat the agent sendsWhat comes backIf it fails
Find or create the patientName (spelled), date of birth, phone, new or existing, reasonThe patient's record in the practice system, and whether they are newAgent offers a callback; never guesses a record
Check availabilityDay as spoken ("next Monday"), visit type, reason, preferred providerOpen slots and a sentence the agent reads word for word"Trouble with the schedule right now"; callback offered
Book the appointmentThe chosen slot, visit type, reason, providerSuccess, written to the practice scheduleAgent never says "booked"; logs a callback
Find the patient's appointmentsName and date of birth; never phone aloneUpcoming appointmentsAsks for date of birth if a phone is shared
Move an appointmentThe appointment id and the new slotSuccessCallback logged
Cancel an appointmentThe appointment id, after offering to move itSuccessCallback logged
Log a callbackWho, why, and who should call backSaved to the callback queueShort timeout so the call is never stuck
Transfer to the on-call doctorWarm transfer, caller shown as the callerA live person on the lineCalm apology and a callback; no intake questions

From the agent configuration and backend handlers in the repository. A ninth handler warms the cache when a call starts; it is not a spoken step.

One booking, message by message

Sequence diagram: caller, Telnyx number, Retell agent, our API, scheduling adapter, Open Dental and AWS text alerts, with 21 numbered messagesOpen full size

How the Telnyx number hands calls to Retell is our reading of the setup; the code confirms the numbers and the agent, not the trunk details.

The call, as five lanes

Call-to-booking pipeline in five lanes: call answered, caller identified, intent and slot search, booking written to the schedule, confirmation and follow-up, with configuration above and supporting services belowOpen full size

Every call ends in one visible outcome

State machine of a call from ringing to confirmed, with final outcomes booked, rescheduled, cancelled, question answered, callback needed, sent to on-call and caller hung upOpen full size

166

calls across the workspace, Aug 26 to Sep 22, 2026

about 1 min

average call length (1 min 3 s)

2

most calls at the same moment

2.3 s

average response latency in the same window

Measured in the voice platform's analytics for the whole workspace, which includes demo lines. These describe use, not business outcomes. No before-and-after measure of bookings or staff time exists yet.

Guardrails written into the agent

No clinical advice. The agent schedules. It does not diagnose, and it does not tell a caller whether a symptom is serious.

Emergency routing. Life-threatening events: "Please hang up and dial 911." Red-flag dental symptoms: warm transfer to the on-call doctor, before any intake questions.

Minimum patient data. Name, phone, date of birth and a short reason. Never medical history, insurance numbers or ID numbers on the call.

A person when it matters. Billing, insurance and anything unsupported become a callback with the reason. Urgent calls also text the practice's emergency phone.

True state only. "Booked" is said only after the practice system confirms it. Times come only from the tool's answer.

Honest promises. Callbacks are promised "during our regular business hours", never "as soon as possible", so callers do not ring back all night.

The scheduling adapter layer

Each practice runs its own scheduling system, and a wrong booking is heard by a real patient. So everything in the platform speaks to one interface, and one adapter per system translates. Adding a system means writing an adapter, not changing the agent.

Scheduling adapter layer: voice tools, staff screens, intake, reminders and scribe call one interface; adapters for Open Dental, a sync partner for Dentrix, Curve and Denticon, and planned behavioral health EHRs; sync and conflict rules belowOpen full size

What the agent hears, and where it lands

Data mapping table from spoken words to platform fields to Open Dental and sync partner fieldsOpen full size

Field names checked against the adapter code on Nov 22, 2026. Rules per practice decide operatory, length and whether a visit goes to the hygiene or doctor column; some practices only allow new patients inside reserved blockouts.

The AI scribe and patient summaries

The scribe records the visit, labels who spoke, and drafts two things in parallel: the clinical note and a summary the patient can understand. The dentist approves each one. Only then does it go to the chart, the PDF and the portal.

AI scribe flow: recording, speech to text, speaker labels, draft note and patient summary, saved as a version, provider approves or edits, signed and locked, then written to the chart, PDF and portalOpen full size

The architecture

One Next.js codebase serves the staff, dentist and patient portals and the API the agent calls. Work after a call runs on a queue. Everything runs on AWS in a US region.

119

API routes

33

screens

85

database migrations

118

test files, unit and end-to-end

2,200

commits, Oct 2025 to Sep 2026

Counted in our copy of the main repository on Nov 22, 2026. The client's latest code is newer than our copy.

The whole system on one page

System architecture: callers and staff on top; Telnyx, Retell, Deepgram, OpenAI, AWS Bedrock, Open Dental, NexHealth, Stripe and AWS messaging on the left; bands for edge, app and API, voice agent orchestration, AI layer with human review, scheduling adapters, workers and queues, data and platform; technology stack and failure handlingOpen full size

Follow the black circles 1 to 9 to see one call move from the phone to the staff screen.

Deployment on AWS

AWS deployment: load balancer, production app server with web app, worker, scheduled jobs and reminders containers, PostgreSQL on RDS, object storage, SQS, Cognito, Secrets Manager, messaging, Bedrock and Transcribe, CloudWatch, hosted scheduling servers and admin consoleOpen full size

Buy, build or integrate

The client bought proven services where they exist, and we built the parts that make a front desk agent safe to put on a practice's phone line.

CapabilityOptionsWhat we choseWhy
Phone numbersCarrier contracts, or a programmable carrierBought: TelnyxNumbers by API for each practice line.
Voice agent runtimeBuild turn-taking and speech, or a voice agent platformBought: RetellTurn-taking, interruptions, voices and transfers were ready. Two other voice providers were tried along the way; the code keeps one interface across all three.
Speech to textOne engine, or twoBought: Deepgram, AWS TranscribeStreaming for the scribe, with a second engine available.
ModelsTrain, or hosted modelsBought: OpenAI, AWS BedrockHosted models behind our prompts and schemas.
Payments and errorsBuild, or buyBought: Stripe, SentrySubscriptions and error tracking are not where a practice wins.
Three scheduling systemsOne adapter each, or a sync partnerIntegrated: NexHealthOne connection reaches Dentrix, Curve and Denticon practices.
The product's edgeBuy a generic answering service, or buildBuilt by usThe agent's tools and rules, the adapters, the scribe review workflow, intake, portal, messaging, staff and owner screens, and the audit trail.

Where AI acts and where a person decides

The agent can book on its own because every booking goes through rules and the practice system's own answer. Everything clinical waits for a person. More in our healthcare AI solutions.

AI doesA person decides
Books a routine visit by phone
Staff review every call next morning
Spots red-flag symptoms
On-call doctor takes the call
Drafts the SOAP note and summary
Dentist edits and approves
Fills intake fields from speech
Patient checks, staff review

When things go wrong

What the platform does in the bad cases.

The scheduling system is slow or down+
Each tool call has a time limit. If the schedule does not answer, the agent says it is having trouble with the schedule and offers a callback during business hours. It never invents a time.
The booking is not confirmed+
The agent does not say "booked". It apologizes, tells the caller the team will finalize it, and logs a callback with the details. Staff see it at the top of the callback queue.
The slot was taken at the desk while the caller decided+
The schedule is read again before writing. If the slot is gone, the agent offers the next opening.
Two family members share one phone+
The agent never finds a record by phone alone. Name and date of birth decide who the patient is, and family members are booked one at a time.
A caller describes an emergency+
Life-threatening events get "please hang up and dial 911". Red-flag dental symptoms go straight to a warm transfer to the on-call doctor. If nobody answers, the caller hears that the team will call back, and the post-call analysis texts the practice's emergency phone.
The post-call report arrives without a call record+
The post-call hook rebuilds the record from what the report carries, so no call goes missing from the log.
The AI draft of a note is wrong+
The dentist edits it or rewrites a section in a concise, standard or detailed style. Each change is a new version, and an earlier version can be restored.

Security as standard practice

These are the defaults we bring to every healthcare build, and the ones this platform uses.

BAA on request

The platform offers a business associate agreement and a data processing addendum to practices on request.

Audit trail

Who viewed or changed a record, and when, including support staff who open a practice's view.

Less patient data on calls

The agent collects only what booking needs. Staff text alerts go through an AWS service covered by the AWS BAA.

Keys out of code

Secrets in a managed store, two-step sign-in for staff, and roles that limit each screen.

We do not claim a certification for this platform. Compliance is a practice's and vendor's shared responsibility, and we document our part.

What we found, what we decided, what improved

A phone agent is judged by callers in real time. These are the lessons we now bring to every voice project.

What we foundWhat we decidedWhat it improves
Models are unreliable at working out "next Tuesday" from today's date.The agent passes the words as spoken; the backend resolves the date against the practice's clock.Fewer wrong-day offers.
Names heard over the phone are often misspelled.First and last name asked separately, spelled, and read back letter by letter.Records found instead of duplicated.
"As soon as possible" made callers phone back all evening.Callbacks are promised "during our regular business hours".Calmer callers, fewer repeat calls.
Each practice books differently: hygiene columns, child lengths, new-patient blocks.Scheduling rules stored as data per practice, not in the prompt.A new practice is configuration, not code.
Three voice providers were tried over the year.One provider interface and one metrics format for all of them.Latency and cost compared fairly; switching is cheap.
Booking bugs are hard to test with real calls.A scenario catalog run against a fake Open Dental server and a test agent.Changes tested before a patient hears them.

Designed, not built yet

What comes next

These screens are concepts we designed with the client's team. None of them is in the product today. Each one reuses the same call data, scheduling adapter and review steps.

Concept: insurance check that reads the payer reply, cites the page and flags what needs a person

Concept. Insurance check with the payer page cited for each benefit.

Concept: recall and reactivation campaigns with progress and chairs filled

Concept. Recall campaigns that fill open hygiene chairs.

Concept: tomorrow's confirmations with no-show risk and reasons

Concept. No-show risk, with the reasons shown.

Concept: multi-location owner dashboard with a stale data warning

Concept. One view across locations, with stale data marked.

How we work with a team like yours

How this engagement ran

Time and material, from the first commit. Our engineers worked as an embedded consultant team inside the client's tools, using accounts the client issued, so the code history carries the client's name. Partner engineers joined under our engagement when the scope grew. The client set priorities week by week with their dentists.

You own the code

Repositories, cloud and vendor accounts sit in your name.

Clinicians in the loop

Your dentists or therapists review the agent's wording and the note formats.

US hours overlap

Engineers in India with daily overlap for stand-ups and demos.

Planned next steps

Behavioral health adapters

SimplePractice and TherapyNotes on the same interface

Measure before and after

bookings after hours, staff phone time

Revenue next to cost

per practice, in the owner report

Recall campaigns

the agent calls overdue patients

These are planned steps, not shipped features.

Want a voice agent that books into your practice's schedule?

Tell us which scheduling system you run and what your front desk misses today. We will show you what an AI agent can safely take on, built as custom healthcare software with real interoperability.

Book a 30-min call

Sources: commit history and source code of the main repository, read Nov 22, 2026; agent configuration in the same repository; voice platform analytics for Aug 26 to Sep 22, 2026. Screens show our redesign of the delivered product with sample data. No patient data is shown. No outcome percentages are claimed.

Questions buyers ask

Can an AI voice agent book appointments directly into Open Dental?
Yes. In this build the agent calls our backend tools, and an Open Dental adapter reads open slots and writes the appointment through Open Dental's API. The agent says a visit is booked only after Open Dental confirms it.
Which dental scheduling systems does the voice agent work with?
Open Dental directly, and Dentrix, Curve and Denticon through a certified sync partner. Each system sits behind its own adapter on one shared interface.
Does it work for behavioral health practices?
The platform is designed for behavioral health too: the same agent, scribe and adapter interface apply. Connections to SimplePractice and TherapyNotes are planned, not live.
What happens if a caller describes a dental emergency?
Life-threatening events are told to dial 911. Red-flag dental symptoms go straight to a warm transfer to the practice's on-call doctor, before any intake questions. If nobody answers, a callback is logged and the practice gets a text alert.
How does the agent avoid booking the wrong patient?
It never finds a record by phone number alone. It asks for first and last name separately, has the caller spell each one, reads them back, and confirms the date of birth.
Does the AI give medical advice on calls?
No. The agent schedules and answers practice questions such as hours and address. Clinical questions go to the on-call doctor or to a staff callback.
Is the platform HIPAA compliant?
We do not claim a certification. The platform offers a business associate agreement and a data processing addendum on request, keeps an audit trail, collects only booking details on calls, and sends staff alerts through an AWS service covered by the AWS BAA.
How does the AI scribe handle dentist review?
The scribe drafts a SOAP note and a plain-language patient summary from the recorded visit. The dentist edits or approves each one; only approved notes go to the chart, the PDF and the patient portal. Every version is kept.
How many calls has the voice agent handled?
The voice platform recorded 166 calls across the workspace from Aug 26 to Sep 22, 2026, about one minute each on average, with at most two at the same time. The client's team separately reports more than 1,100 after-hours calls at one deployment over 6 weeks; that is their figure.
How long does it take to build an AI front desk like this?
This platform grew over about a year of steady work, from the first commit in October 2025, with an embedded team on a time-and-material basis. A first voice agent on one scheduling system is a smaller slice; we scope it with you on a call.