How Triagents handles emergencies

The single most important safety behaviour on Triagents is what happens when a patient describes an emergency. There is one answer, and it is the same on every call, at every site, in staffed hours and out of them.

The agent tells the caller to hang up and call 000. It repeats it once. Then it stops talking. The number is the emergency number for the country the clinic operates in: 000 for Australian clinics, 911 in the United States, 999 in the United Kingdom. The behaviour around it does not change with the country, and this post says 000 because that is where the platform runs today.

No hold music. No “please stay on the line”. No transfer to reception, no conference to an on-call clinician, no attempt to work out how serious it is first. The practice is notified with the transcript and a recording reference, and the event is written to a permanent audit record. Recognition and response both land within 1–2 seconds of the utterance.

Why this is not configurable

Almost everything else on the platform is set per site. Opening hours, escalation paths, what the agent asks, how it responds, which referrer gets routed where. This one is not, and the reason is worth stating plainly.

Every alternative puts something between the patient and an ambulance. A transfer to reception is fast when someone picks up and is a silence when they do not. A conference to an on-call clinician depends on a phone being answered. Each of those paths is defensible on the day it is designed and fails on the day the person it assumed is in a room with a patient. The 000 instruction has no such dependency, which is the entire argument for it.

It also removes a decision from a configuration screen that should never have been on one. A practice choosing its emergency escalation path is a clinical-governance decision being made months in advance, by someone reading a settings page, about a call they cannot picture. Making it uniform means it cannot be set wrong, cannot drift after a staffing change, and cannot differ between two sites in the same group because two managers filled in the same form differently.

The agent is not qualified to triage an emergency and the platform is not built to. So it does the one thing that is unambiguously right and gets out of the way.

What “emergency” means

The recognition is configurable. The response is not.

Each practice configures its own emergency vocabulary in TriageWorkflow. The defaults cover a clinically broad set: chest pain, severe bleeding, stroke symptoms, severe allergic reaction, mental-health crisis vocabulary, paediatric emergency words, and the practice can add specialty-specific terms.

For radiology, the default list also includes contrast-reaction vocabulary, post-procedure complication terms, and a few imaging-specific clinical phrases that don’t show up in generic medical NLP corpora.

Widening that list widens what gets the 000 instruction. It never changes what the instruction is.

What happens in those 1–2 seconds

1. The intent classifier flags the utterance as an emergency at high confidence.
2. The agent stops working whatever it was working on, mid-booking if necessary.
3. It delivers the instruction: hang up and call 000, or the emergency number for the clinic’s country. It says it once more, and ends its own involvement.
4. The practice is notified with the full transcript and a recording reference.
5. The event is logged with a permanent audit record and reviewed within 24 hours.

Step 2 matters more than it looks. The failure mode in a voice agent is not usually missing the emergency; it is hearing it and finishing the sentence it was already saying. The booking flow is abandoned, not completed and then escalated.

How we validate it

Three layers:

• Per-release regression suite. Several hundred red-team utterances, including clinically obscure phrasings, sarcasm, indirect statements, and adversarial inputs designed to evade simple keyword matchers.
• Per-practice configuration validation. Before any practice goes live, we run their configured vocabulary against a sample of real call transcripts to confirm the recognition fires where it should. The response needs no validation per practice, because it is the same one every time.
• Live monitoring. Every emergency recognition is reviewed by a clinical engineer within 24 hours of occurrence.

What we don’t do

We don’t triage the emergency. We don’t give clinical advice. We don’t grade severity, ask follow-up questions to decide how urgent something is, or conclude that it is probably nothing. We don’t hold the caller while we find someone, because the seconds that costs are the only ones that matter.

Recognising that a caller has described an emergency is a routing decision. Deciding what it means is a clinical one, and the only safe route is off the platform and onto 000.

That’s the floor. Nothing about Triagents exists if that floor isn’t held.