Dentrix AI Receptionist Integration: What the PMS Allows

How Do You Stop an AI Receptionist From Double-Booking?

Treat the booking as unconfirmed until the PMS says otherwise. While your AI talks, a front-desk person can book the same chair, another call can take it, and a polled schedule can already be out of date. Open Dental’s create call, for one, doesn’t say it checks for overlap, so the check is yours.

The sequence we use:

1. Read open time and offer one or two slots.
2. When the caller agrees, write the appointment.
3. Wait for the PMS to accept it and return the appointment’s ID.
4. Read that appointment back, and the slot around it, to confirm nothing else landed there.
5. Only then say “you’re booked.”

If step 3 fails or step 4 shows a clash, the caller hears “that time just went, can I offer 10:40?” and never a false confirmation. Retries need care too: if a write times out, look for the appointment before writing it again, or a network blip becomes a double booking. On a PMS with no event feed, we plan for the polling delay up front and keep the read-back in every call.

What Should the AI Do When the PMS Write Fails?

Decide the wording for each failure before you decide the happy path, because a caller remembers a broken promise longer than a slow answer.

• Write rejected (slot gone, a required field missing, a provider not scheduled): offer the next open time and try once more.
• PMS unreachable (office server off, eConnector down, a vendor outage, an expired key): don’t pretend. Take the request, say the practice will confirm, and put a task in the front desk’s queue with the caller’s number and preferred times.
• Patient match unclear (two records share the phone number, the name doesn’t match): book nothing against a guess. Collect the details and hand it to staff.
• After hours, urgent: follow the practice’s own rules for emergencies, and keep clinical triage out of the AI’s hands unless the practice has written those rules down.

Log every attempt, the PMS response and what the caller was told. When a patient says “your system booked me,” that log is how the practice answers.

How Should the AI Match Callers and Capture Insurance?

Matching is where duplicate records and privacy problems both start. Phone number alone isn’t enough, because families share numbers and a parent often calls for a child. Match on phone plus name plus date of birth, and create a new patient only when nothing matches. Anything in between goes to staff, since merging duplicates later is a PMS clean-up job nobody enjoys.

Before the AI says anything from the record (the next appointment, a balance, what treatment is planned), HIPAA’s verification rule applies. 45 CFR 164.514(h) requires a covered entity to “verify the identity of a person requesting protected health information and the authority of any such person to have access” when it isn’t already known. For a voice agent, that’s a scripted identity check before any disclosure, and a rule for what it says when the check fails.

Insurance capture is simpler than verification. The AI collects the payer, member ID, subscriber name and date of birth, and the patient’s relationship to the subscriber, and saves them to the patient. The eligibility check itself runs as an electronic request and response (X12 270/271, the HIPAA eligibility standard under 45 CFR 162.1202) before the visit.

Our dental insurance verification automation page covers how that writes back to the PMS, and the step-by-step front-desk version is in how to verify dental insurance. In our dental eligibility automation webinar, our VP of Engineering, Arun Badole, runs the Mindbowser build live on Open Dental with Stedi’s product team, from the 271 response to the benefits landing in the patient’s record.

HIPAA, BAAs and Outbound Calls for an AI Receptionist

The practice is the covered entity. Your AI receptionist company creates, receives and transmits protected health information on the practice’s behalf, so it’s a business associate and signs a BAA with each practice. That part most vendors know.

The part that gets missed is downstream. The speech-to-text service, the model provider and the telephony platform all handle the call audio or the transcript. Under the HIPAA definition in 45 CFR 160.103, a subcontractor that creates, receives, maintains or transmits protected health information on behalf of a business associate is itself a business associate. 45 CFR 164.502(e) lets you pass PHI to that subcontractor only with satisfactory assurances, which means a BAA.

Before a single live call, list every company in the audio path and check you hold a BAA with each. Our guide to business associate agreements under HIPAA covers what goes in one.

Read only what the call needs, and treat recordings and transcripts as PHI with the same access controls as the chart.

Outbound calls are a separate question from answering the phone. In February 2024 the FCC confirmed (FCC-24-17) that AI-generated voices count as “artificial or prerecorded voice” under the Telephone Consumer Protection Act, so calls using them need the called party’s prior express consent unless an exemption applies.

If your AI also calls patients for recalls or confirmations, get consent handling and the exemptions reviewed by counsel. Recall outreach itself is covered in our guide to dental recall software. This section isn’t legal advice.

How to Test an AI Receptionist Integration Before the First Live Call

Test the failure paths first, in each vendor’s test environment (Figure 4) and against the PMS versions your customers actually run. The happy call works in every demo.

The cases to have passing before go-live:

• Slot race: a second booking lands in the same slot mid-call.
• Rejected write: a missing required field or an unscheduled provider.
• Duplicate patient: two records share the caller’s phone number.
• Family number: a parent books for a child.
• Server off: the office server or eConnector is down at 7:40 pm.
• Timeout and retry: the write times out and is retried.
• Version mismatch: a call the customer’s PMS version doesn’t support.
• Unreadable insurance: the caller doesn’t know the member ID.

Every one should end in a correct booking, a new offer or a staff task, and the log should show which.

Building the PMS Side of Your AI Receptionist

We build the PMS integration layer for AI-receptionist products: the local service each desktop Dentrix or Eaglesoft office needs, the booking and patient writes, confirm-after-write, polling where a PMS publishes no events, and the failure paths above. We help you through each vendor’s partner application (approval is the vendor’s call), and after launch we keep it running, adjusting the integration when a vendor changes its API.

For Dentrix, Eaglesoft, Denticon, CareStack and Curve, we build to each vendor’s route. The voice-agent method itself is covered in our guide to AI voice agents in healthcare, and building the dental AI product is on dental AI development.

Outside dental, we built the connector that lets an AI receptionist work inside a hospital EHR. For a venture-backed AI health company, our Cerner Millennium connector for AI scheduling and documentation agents gives the scheduling agent patient search and appointment write, plus shared insurance and practitioner-directory lookups.

Booking into medical EHRs as well? Some AI receptionists serve dental groups inside health systems, or move into medical practices. ConnectHealth, our integration platform, runs that side: you deploy it in your own cloud environment, it connects to Epic and athenahealth over FHIR (each subject to that EHR’s own app approval), and its workflow engine includes AI agent tools with MCP support.

For AI steps, use a model inside your environment or the Guardrails node so PHI stays there. See ConnectHealth for Epic and athenahealth. The dental PMS side stays a Mindbowser build.

We work in 2-4 week sprints under a signed BAA, with SOC 2 Type II in place, and you own the code and the data.