Dental Emergency Call Triage: What AI Should and Should Not Decide

A patient with facial swelling at 10pm is the most consequential caller a dental practice receives, clinically and commercially. It is also, in the majority of US practices, the caller most likely to hear a recorded message asking them to phone back in the morning.

Automating this call correctly is therefore valuable, and automating it badly is genuinely risky. The line between the two is scope: the AI handles logistics and hand-off, and takes no clinical position whatsoever.

What it should do

• Recognise urgency signals — pain, swelling, bleeding, trauma, knocked-out tooth — and switch to the emergency flow immediately.
• Collect structured facts: what happened, when it started, whether there is swelling or fever, whether the patient is an existing patient, and a callback number confirmed back to them.
• State clearly and immediately what will happen next and in what timeframe.
• Escalate to the on-call clinician by call, SMS and email simultaneously, with the collected detail attached.
• Offer the recognised safety fallback: if the situation involves difficulty breathing or swallowing, uncontrolled bleeding, or facial trauma, direct the caller to emergency services without delay.

What it must never do

• Grade severity or tell the caller whether the problem is serious.
• Recommend medication, dosage, or any home remedy.
• Tell a caller it is safe to wait until morning.
• Decide that a case does not warrant escalation.

The design principle: the AI is a fast, accurate messenger with a defined safety fallback. Every judgement belongs to a clinician.

Escalation fails at the second step far more often than the first, because most practices define who is called and not what happens when that person does not pick up. Write the full ladder down.

1. Primary on-call clinician: simultaneous call, SMS and email with the structured summary.
2. If no acknowledgement within a defined window — five minutes is typical — move to the secondary on-call contact.
3. If neither responds within ten minutes, the AI calls the patient back with the agreed fallback message and directs them to the nearest urgent dental or emergency facility.
4. Every step is logged with timestamps, so the following morning shows exactly what happened and when.

Rotate the on-call list in the configuration rather than in someone's memory, and review the log weekly for the first month. Unacknowledged escalations are the failure mode to hunt for.

Emergency calls contain protected health information by definition — symptoms, history, identity. Recording, transcription, storage and the notification channel all fall inside HIPAA, and a business associate agreement must be in place with the vendor before a single live call is handled.

Pay particular attention to the notification channel. Sending clinical detail to a personal SMS number is a common and avoidable gap; agree with your vendor how much detail travels by which route. The HIPAA and BAA guide covers the wider requirements.