Get started

Urgent-call routing to humans

Urgent calls rarely arrive labeled correctly. A caller may choose a routine department, describe the impact before the cause, or use a phrase that sounds severe but lacks the facts a responder needs. If every caller receives the emergency route, the on-call team stops trusting it. If the first-line flow waits too long, a real operational problem remains in the queue.

Urgent-call routing uses business-written triggers and questions to recognize a priority path, capture the minimum useful context, and attempt a human handoff. The AI phone agent does not decide what is objectively dangerous or legally urgent. It applies the definitions the organization has approved and routes uncertainty conservatively.

A concrete urgent call

Suppose a tenant calls a property service line after finding water entering a locked equipment room.

Agent: I can help route an urgent maintenance report. What is the property address and where is the water visible?

Caller: It is coming through the ceiling outside the equipment room, and the door is locked.

The prompt marks active water near an equipment area as an urgent trigger. The agent asks whether anyone is in immediate danger, confirms a callback number, records the address and access constraint, and tells the caller it will attempt the on-call route. It does not tell the caller to enter the room or diagnose the source.

With attended warm transfer, the caller waits on hold while the agent calls the designated responder privately.

Agent to responder: A caller reports active water through the ceiling outside the locked equipment room at the stated property. No one is reported inside. I have the caller and callback number. Will you accept the call?

If the responder accepts, ThunderPhone connects the parties. If the responder declines, does not answer, or reaches voicemail, the agent returns to the caller and follows the written fallback. That might mean trying the next authorized contact, creating a priority record through REST, and sending a Slack alert. The agent should explain that a message or alert is not proof that a person has received it.

How to set it up on ThunderPhone

Define urgency by observable caller statements, not vague instructions to “use judgment.” For each trigger, document the questions to ask, the destination, the order of alternates, and the language used when no one answers. A practical prompt outline includes:

  • the urgent conditions the business recognizes and examples of ambiguous wording;
  • the minimum location, identity, callback, impact, and timing fields;
  • instructions to avoid diagnosis or risky advice;
  • the conditions that bypass routine intake;
  • the primary and alternate transfer targets; and
  • the final fallback, including what the caller should do if the situation involves immediate danger.

Urgent routing is often one branch within call screening and routing, but give it separate rules and tests. Attach only caller-safe escalation policies and location details to the knowledge base. ThunderPhone supports text, Markdown, CSV, PDF, and DOCX documents with per-agent scoping and mid-call search.

Use Slack for an on-call notification and HubSpot or Salesforce when the urgent record belongs in an existing customer workflow. Google Sheets can provide a simple reviewed escalation log. Custom dispatch, incident, or case systems can receive data through REST or webhooks. Remote Zapier MCP tools can be attached when the exposed action has been approved. Keep the transfer result, trigger, caller statements, timestamp from the receiving system, and fallback outcome together.

Use an eligible production number for attended transfer. In warm mode, ThunderPhone places the caller on hold, privately briefs the recipient, connects after acceptance, and returns to the caller when the attempt fails. Demo numbers and browser calls are limited to cold transfer because warm transfer requires a number able to place an outbound SIP leg.

Test both misses and false alarms: incomplete addresses, indirect urgency phrases, routine calls using dramatic language, disconnected transfer targets, voicemail, integration failures, repeat callers, threats, and reports of immediate danger. Reusable scenarios, graded call logs, and regression pass-rate gates can verify the routing rules after a prompt change.

What to measure

Measure whether the defined path reaches a responsible person:

  • urgent triggers recognized and missed in reviewed calls;
  • routine calls incorrectly sent to the urgent route;
  • required context captured before the transfer attempt;
  • transfer acceptance, failure, and alternate-route completion;
  • alerts or records delivered to the intended system;
  • repeat calls before a human response; and
  • incidents where the agent gave advice outside its prompt.

For operational dispatch after triage, connect this flow to field service dispatch. Property teams can keep industry-specific policies on the property management page rather than duplicating them here.

Honest limits

An AI phone agent is not an emergency service, clinician, safety officer, or legal decision-maker. It cannot see the scene, confirm that an alert was read, or guarantee that a recipient will answer. Callers reporting immediate danger need clear instructions to contact the appropriate emergency service, based on language approved by the organization.

Humans must own threats, medical symptoms, suspected crimes, safety hazards, and any ambiguous case where delay could cause harm. The agent's role is to recognize the written trigger, preserve the caller's facts, and move the call to that human path quickly.

Create your ThunderPhone account to configure and test urgent-call routing.