Overview
Heron is built to stay on the administrative/reception side of a call — booking, rescheduling, answering practice questions — and to hand off immediately, not attempt to help, the moment a call sounds like it might involve an emergency or a mental health crisis. It doesn't try to assess, counsel, or triage the caller. It recognises certain words and phrases, plays a fixed safety message, ends the call so the line is free for the caller to dial emergency services, and flags the interaction for your team to follow up.
How detection works
Heron is an LLM so will look to match associated phrases with medical emergencies and mental health crises (see categories below). It isn't making a clinical judgement about what's happening — it's pattern-matching against known crisis language, the same way you'd train a receptionist to recognise "I think I'm having a heart attack" as a hang-up-and-call-111 situation rather than something to triage themselves.
Because this is word/phrase-based, it will occasionally trigger on things that aren't genuine emergencies (e.g. a caller referencing a past incident, or using one of the phrases in an unrelated context) — that's expected, and treated as an acceptable trade-off given what's at stake if a real one were missed.
What happens when it triggers
Heron stops the normal conversation flow. It doesn't continue trying to book an appointment or answer the original question.
It plays a fixed safety message. The message tells the caller this isn't something Heron can help with directly and directs them to emergency services (111 in NZ / 000 in Australia / 911 in USA / 192 in Brazil currently configured).
The call ends. Heron disconnects deliberately, so the line is immediately free for the caller to dial emergency services themselves rather than staying on hold or in a queue.
The interaction is tagged and routed to "Needs Staff." It lands in your Heron inbox flagged for follow-up, with the transcript, timestamp, and caller number (where available), so your team can review and action it appropriately — including reaching out to the caller if that's the right call.
Trigger word/phrase categories (illustrative — confirm against the live list)
The categories below are the kind of thing a system like this is typically built to catch. Treat this as a starting point for the conversation with product/engineering about what's actually configured, not a final list to publish:
Suicide / self-harm: e.g. - I'm feeling suicidal", 'I want to hurt myself', 'No body cares or loves me', "I'm very depressed", 'I hate everything', 'I want to end it all', "I don't think I can keep going", "I need help, I'm really not okay anymore", 'Can I speak to someone about mental health", "suicide," "kill myself," "end my life," "want to die," "hurt myself"
Life-threatening medical symptoms: e.g. - 'Cant breathe', 'struggling to breathe', 'very short of breath', 'Chest pain, crushing chest pain, dull chest pain/ache', 'Pain radiating down left arm, up neck, jaw pain', 'Sudden rash all over body', 'Feels like throat is closing up', 'Numb feeling one side of body', "Any bleeding that is very heavy or won't stop", 'Feels like they have had a TIA/mini stroke', 'Face is drooping', 'Struggling to speak', 'Child is choking', 'Child has blue lips/not waking easily', 'Child is fitting',
Immediate danger from another person: e.g. "he's hurting me," "I'm in danger," references to weapons in a threatening context.
Explicit emergency framing: e.g. "this is an emergency," "I need an ambulance"
What lands in "Needs Staff" and how to action it
Each of these calls will show up in the "Needs Staff" queue in your Heron workflow. Because the caller has already been told to contact emergency services and the call has ended, this isn't a queue to action "live" the way you might jump on a ringing phone — but it should be checked promptly (recommend building it into your regular inbox-check cadence, not left for end of day), since:
A caller may not follow through on calling emergency services themselves.
Your team may recognise the caller and have context that changes the right response.
It's useful practice-level visibility even when the situation has already been escalated elsewhere.
What this feature is not
Not a clinical or diagnostic triage tool. Heron doesn't assess severity or make a judgement call — it recognises trigger language and hands off, full stop. This keeps it on the administrative side of the line, consistent with how Heron is positioned generally (see the existing "keep-human" framing in your webinar materials — clinical triage, diagnosis, and distressed/complex patients are explicitly out of scope for the AI).
Not a replacement for your practice's own emergency procedures. Staff should keep following your existing protocol for handling emergency calls — this is a safety net on the Heron side, not a substitute for it.
Not guaranteed to catch every phrasing. Word/phrase-based detection will miss some genuine emergencies expressed in unexpected language, and will occasionally trigger on non-emergencies. Practice managers should set expectations with staff accordingly rather than treating it as infallible.