Multi-Location Dental Group: A Voice AI Agent Inside the Scheduling System

Overview

A multi-location dental group serving a major Midwest metro area operates with **26 providers** across **4 clinical locations**. With a high daily patient volume, the organization faced significant scheduling and operational challenges that directly impacted patient experience and revenue.

Client identity is withheld under our clinical data-privacy agreement to protect the operator's competitive scheduling advantage.

The Challenge

The practice encountered serious scheduling inefficiencies that were costing them patients, revenue, and staff morale.

The Solution

Techland deployed a **Voice AI Agent** directly integrated into the practice's EMR system. The integration was seamless, like *"flipping a switch"*, requiring minimal effort while following the organization's complex scheduling rules.

Key Features

How It Was Built

The interesting constraint on this project was not the AI. Voice agents are a solved problem in the abstract. The constraint was that this one had to run in a clinical environment, against protected health information, inside a scheduling system that predates the entire category, and it had to answer fast enough that a patient does not notice they are talking to software.

Why the models run locally

Every alternative we looked at meant streaming patient audio to a third-party API. That is workable, but it means PHI leaving the practice's boundary, a Business Associate Agreement with every vendor in the chain, and a dependency on someone else's uptime during clinic hours.

So we did not do that. Speech recognition and the language model both run on-premise, on a single GPU host inside the practice's own network. Audio is transcribed, reasoned over, and answered without any of it crossing the network boundary. The only thing that leaves is what already left before we arrived: nothing.

This is the part that transfers to other clients, and it is why we lead with it. Any business with a data-residency rule, a regulator, or a customer contract that restricts where information can be processed can have a capable AI system. It just has to be built to run where the data already is, and that is an architecture decision made at the start, not a setting toggled at the end.

The latency budget

A voice agent lives or dies on turn-taking. Past roughly a second of silence after a caller stops speaking, the conversation stops feeling like a conversation. That single number drove most of the engineering.

We budgeted the whole round trip, from the caller finishing a sentence to the first syllable of the reply, at under 800 milliseconds, and then spent that budget deliberately: streaming transcription that begins before the caller stops talking rather than after, speculative lookups against the schedule while the sentence is still being parsed, and speech synthesis that starts emitting audio on the first clause instead of waiting for the full response.

Running the models locally helps here too. There is no round trip to a public API in the middle of the loop.

Rules the model is not allowed to improvise

A dental group's scheduling logic is genuinely intricate: which providers can perform which procedures, how long each takes for a new patient versus a returning one, which operatory and which of the four locations, hygiene recall intervals, and the blocks that exist for reasons only the practice knows.

None of that is left to the language model. The model's job is to understand what the caller wants and to hold a natural conversation. The booking decision itself goes through a deterministic rule engine that reads the practice's actual constraints from the EMR. If the rules say a slot is not valid, the slot is not offered, regardless of what the model thinks. A language model that is confidently wrong about a medical appointment is worse than no automation at all, and the architecture makes that class of failure structurally impossible rather than unlikely.

When the agent cannot resolve a call within its competence, it does not guess. It warm-transfers to a human with the transcript and the caller's record already on screen, so the patient does not start again.

The stack

Layer	Approach	Why
Telephony	SIP trunk into the existing practice number	No number change, no porting, no patient-facing disruption
Speech recognition	Streaming ASR, self-hosted on-premise	Transcription begins mid-sentence; audio never leaves the network
Language model	Open-weight model served locally on a single GPU host	No PHI to third parties, no per-token cost, no vendor uptime dependency
Booking logic	Deterministic rule engine over the EMR's own constraints	The model never decides a slot is valid; the rules do
EMR integration	Direct read/write against the practice's scheduling system	One source of truth; the front desk sees bookings appear in the system they already use
Escalation	Warm transfer with transcript and patient record attached	A call the agent cannot finish is not a call the patient has to repeat
Observability	Every call logged with transcript, decision path, and latency trace	When something goes wrong the practice can see exactly why, and so can we

None of the individual pieces here are exotic. What made the project work was refusing to let the interesting part, the language model, make decisions it should not be trusted with, and putting the whole thing where the data already lived. That pattern is not specific to dentistry. It is how we would approach any regulated workflow where the cost of a confident mistake is higher than the cost of the automation.

Impact & Results

Call Handling

• Sub-800ms response latency from the end of a caller's sentence to the first spoken syllable, maintaining natural turn-taking.
• Hold times reduced from over 30 minutes to under 5 minutes across all 4 clinical locations and 60-patient daily clinic schedules.
• Zero-repetition warm transfers to human staff with live transcripts and EMR records pre-loaded on screen for edge-case calls.

After-Hours Performance

• 24/7 automated appointment booking via the existing practice phone number with zero third-party PHI exposure.
• 100% deterministic rule enforcement outside clinic hours, matching new and returning patients to eligible providers, operatories, and recall intervals directly inside the EMR.

Business Outcomes

Conclusion

By integrating a Voice AI agent directly into its EMR system, the dental group transformed its scheduling operations. The result: faster patient access, reduced operational strain, significant incremental revenue, and a scalable solution aligned with complex clinical workflows.

This case demonstrates how AI-driven automation can directly improve healthcare operations, patient experience, and financial performance, without disrupting existing systems.

Due to clinical confidentiality agreements, the practice's name is withheld from public distribution. Qualified healthcare operators can request a live architecture walkthrough and anonymized call recordings on a discovery call.