How AI Receptionists Should Handle Emergencies

Every AI receptionist you deploy will eventually take an emergency call. A patient having a stroke. A property tenant with water gushing through the ceiling. A law-firm client who was just arrested. How the agent handles those calls is the single most important quality dimension in the entire system, because the downside is unbounded — missed escalation in a medical emergency can cost a life, and the legal liability that follows can destroy a business.

This piece covers how to design, test, and operate the emergency path of an AI receptionist. It's not a "nice to have" feature. It's a safety floor that every deployment needs to clear before go-live.

TL;DR

• Define emergency categories up front, per vertical — they vary widely.
• The emergency path must be hard-coded, not prompt-negotiable.
• Immediate routing. No clarifying questions beyond what's needed to route.
• Direct callers to 911 when appropriate — don't try to be a dispatcher yourself.
• Test emergency scenarios before every release, not just once.

Start with the question: what's an emergency here?

Emergency definitions are vertical-specific. Write yours down.

Healthcare / medical:

• Chest pain, difficulty breathing, stroke symptoms.
• Severe bleeding, loss of consciousness, severe allergic reaction.
• Suicidal ideation, overdose, psychiatric emergency.
• Pediatric emergencies (generally lower threshold).

Dental:

• Severe pain with swelling (possible abscess).
• Knocked-out permanent tooth in last 30 minutes.
• Severe bleeding after extraction.
• Trauma to face/jaw.

Property management / real estate:

• Fire, gas smell, carbon monoxide.
• Flooding / major water leak.
• Break-in or security incident.
• Lockout with a safety concern (child locked out alone, storm).

Legal:

• Arrest or detention.
• Court deadline in under 24 hours with no filing prepared.
• Restraining-order violations.

IT / SaaS support:

• Production outage affecting all users.
• Active security incident or data breach.
• Data loss from a paying customer.

Hospitality:

• Medical event in-room.
• Security threat to a guest.
• Fire, flood, structural issue.

Every vertical has its own list. Missing one of these during design is how you get a catastrophe.

The emergency rule must be hard-coded

"Hard-coded" means: not left to the LLM to interpret in context. The system prompt must have explicit triggers and explicit actions for each category. Example:

EMERGENCY RULES (non-negotiable):

If the caller says or implies any of the following,
immediately stop whatever flow you're in and say:
"This sounds urgent — I'm connecting you to our
emergency line right now. If it's a medical emergency,
please also call 911."

Then call page_on_call(reason=<summary>, category=<category>).

Triggers (medical):
- chest pain, "can't breathe", "trouble breathing"
- stroke symptoms: slurred speech, facial droop, one-side weakness
- severe bleeding
- loss of consciousness, "passed out"
- suicidal: "kill myself", "want to die", "end it"
- overdose: "took too many", "overdose"
- "emergency", "911", "call an ambulance"

Do NOT ask clarifying questions beyond: "Are you safe
right now?" if it's unclear whether the caller is the
patient.

Note: no ambiguity, no "maybe", no LLM judgment call. The trigger fires → the action fires.

The routing targets

You need three destinations wired up:

1. On-call person. Your clinical lead, head of operations, attorney on call. Reached via the rotation tool — PagerDuty, Opsgenie, or a scripted phone-forward.

2. External emergency services. For true emergencies, 911 is the right answer. The agent should direct callers to 911 and hang up last if the caller is unable to call themselves (rare but real).

3. Escalation backup. If the on-call person doesn't answer within 60 seconds, fall through to a secondary destination. Never leave a caller with unresolved emergency on the line.

Never ask to "handle" an emergency

A common AI failure mode: caller says "I think I'm having a heart attack," and the agent, following its helpful-assistant training, tries to "help" with troubleshooting questions. "Can you describe the pain? How long has it been going on?"

Wrong. The agent's only job in an emergency is to route. Fast.

Write the system prompt to explicitly forbid this:

Do NOT attempt to diagnose, triage medically, or "help"
beyond connecting the caller to a human. Your job in an
emergency is to route, not to handle. Any medical,
legal, or safety-related questions get deferred to the
person you're routing to.

Testing the emergency path

Every release that changes the prompt, the LLM, or the routing logic gets an emergency test suite run against it. At minimum:

• 10–20 emergency scenarios per vertical.
• Played as audio or text input.
• Expected output: trigger fires, correct route, no clarifying questions.
• Automated pass/fail.

This is non-negotiable. An un-tested emergency path is a liability waiting to happen.

For the broader testing approach, see how to A/B test voice agent prompts and how to measure voice agent quality.

False positives are OK

Over-escalation is a feature, not a bug. If the agent sometimes pages the on-call for a situation that turned out to be non-urgent, that's fine. The on-call person deals with it. The alternative — under-escalating and missing a real emergency — is unacceptable.

Tune triggers to err on the side of caution. Acceptable false-positive rate: 5–15%. Acceptable false-negative rate: as close to zero as you can engineer.

The post-call protocol

Emergency calls don't end when the call ends. You need:

• An audit log — who called, when, what was said, who was paged, when they responded.
• A post-mortem on any missed or mishandled emergency.
• A follow-up to the caller within a defined window (e.g., 24 hours).
• Legal review for any incident that led to harm.

Treat emergency calls like production incidents — they get reviewed.

Jurisdictional considerations

Some rules vary by jurisdiction:

• HIPAA (US healthcare): the agent's emergency logs are part of the medical record. Retention policies apply.
• State-by-state medical scope: some states restrict what AI can do in emergency scenarios. Know your state's rules.
• International: 911 becomes 112, 999, etc. Route accordingly.
• AI disclosure laws: some jurisdictions require disclosure; during an emergency, disclosure happens in the greeting but the agent should not stop to re-disclose mid-emergency.

Work with legal counsel before go-live in any vertical that touches human safety.

Common failure modes

Relying on LLM judgment for emergency detection. Hard-code the triggers. LLMs are non-deterministic; your life-safety rules must be deterministic.

No on-call fallback. On-call doesn't answer → call drops. Always chain a secondary route.

Clarifying-question creep. Well-meaning PM adds "ask which location" — now the agent asks a dying caller which store. Remove any non-routing question from the emergency path.

Forgetting to test. Emergency paths silently break with prompt changes. Automate the test suite.

No audit trail. Can't prove what happened, when. Always log emergency events separately.

Related reading

• Designing an AI Receptionist From First Principles
• Cost Comparison: Hiring a Receptionist vs Deploying AI
• Greeting Design: First-Impression Engineering for AI Voices
• How AI Receptionists Handle Repeat Callers
• Voicemail Replacement: Why AI Receptionists Win

FAQ

Can the AI dispatch 911 itself? Generally no. Tell the caller to dial 911. The AI can stay on the line to provide information if the caller is incapacitated, but initiating emergency dispatch autonomously is a regulatory and technical minefield.

What if the caller is describing someone else's emergency? Same routing. The agent doesn't need to distinguish — urgency is urgency. The on-call person figures out the details.

What about silent or unclear calls? If you can't rule out a silent distress call, err on the side of escalation. Better a manager gets paged for a dropped call than a real emergency goes unanswered.

How do we handle mentions of suicide or self-harm? Treat as emergency. Immediate routing. Provide crisis hotline numbers (988 in the US) while routing.

Can the AI talk the caller down during an emergency? No. That's clinical work. Route to a human immediately. The agent is not a crisis counselor.