[Developers]

AI Triage

At 2:47 AM, a 911 call centre receives 34 simultaneous calls. Four of them contain gunshot sounds in the background. Two involve callers who sound like children. One is a prank. The remaining 27 are noise complaints and

Category: Api DomainsLast Updated: Apr 2, 2026
api-domainsaireal-timecompliance

Overview#

At 2:47 AM, a 911 call centre receives 34 simultaneous calls. Four of them contain gunshot sounds in the background. Two involve callers who sound like children. One is a prank. The remaining 27 are noise complaints and fender-benders. A dispatcher handling calls manually cannot sort that queue fast enough. Every second that a life-threatening call sits unrouted is a second that matters.

The AI Triage module processes every call through a real-time intelligence pipeline. Speech is transcribed as it happens. Sentiment, audio events, and keywords are scored in parallel. Each call receives a P1-through-P5 priority before the dispatcher even picks up. Life-threatening situations reach the top of the queue automatically. The dispatcher works the list; the AI keeps the list sorted.

Key Features#

  • Real-Time Call Transcription: Speech-to-text transcription of emergency calls with multi-language support, keyword extraction, and streaming capabilities for live call processing.
  • Sentiment Analysis: Analyse caller emotional state to identify panic, distress, aggression, and confusion, with automatic escalation recommendations when urgency thresholds are exceeded.
  • Background Audio Classification: Classify audio events such as gunshots, explosions, screams, vehicle crashes, and other emergency indicators with automatic escalation for critical events.
  • Dynamic Priority Scoring (P1-P5): Multi-component risk scoring that combines sentiment, audio, and keyword signals to assign priority levels from P1 (critical, immediate life threat) to P5 (informational), with configurable target response times.
  • Emergency Keyword Detection: Automatic extraction and categorisation of emergency-relevant keywords (weapons, medical, fire, crime, traffic, danger) from call transcripts with weighted scoring.
  • Surge Redistribution: Monitor PSAP capacity in real time and automatically redistribute calls during surge conditions to partner PSAPs, overflow centres, or AI deflection queues.
  • AI Callback System: Schedule and manage automated callbacks for non-urgent (P4/P5) calls, reducing queue pressure while maintaining service quality with multi-attempt retry logic and human escalation.
  • Predictive Staffing: Forecast call volumes and staffing needs by hour, generate proactive alerts for understaffing or surge conditions, and provide shift coverage analysis.
  • Smart Triage Gate: A structured triage evaluation runs on every significant transcript update, producing one of five outcomes: escalate immediately (P1/P2 life threats), dispatch needed (P3), follow-up queue (P4), prank logged (P5), or pending. The triage gate enforces a one-way escalation rule where a call that reaches immediate escalation can never be downgraded, ensuring life-threatening situations are never missed. Triage decisions control when dispatcher broadcasts are sent, preventing information overload from non-emergency calls.
  • Vulnerable Caller Detection: Over 30 phrases associated with children, elderly, and distressed callers trigger immediate escalation regardless of other signals. Phrases such as "help my mom", "my dad won't wake up", and "I'm a kid" are treated as critical indicators requiring human dispatcher involvement without delay.
  • Prank and Non-Emergency Filtering: The triage system identifies prank indicators (8 phrases including "just kidding", "it's a joke") and non-emergency keywords (20 terms including "noise complaint", "pothole") to route calls appropriately. Emergency keywords always override prank or non-emergency indicators to prevent false filtering of genuine emergencies.
  • PSAP Priority Floor: Emergency line calls enforce a P3 minimum priority floor. P4 (follow-up) and P5 (prank) outcomes on PSAP lines are automatically clamped to P3, ensuring every emergency line caller receives at least a standard response.
  • Manual Priority Override: Supervisors can override AI-assigned priorities with documented reasoning for full audit trail accountability.
  • Programmable API Access: Full API support for all triage operations including transcription, analysis, priority evaluation, surge management, and staffing predictions.

Priority Levels#

  • P1 (Critical): Immediate life threat, target response within 60 seconds. Risk score 80-100.
  • P2 (High): Urgent response needed, target response within 3 minutes. Risk score 60-79.
  • P3 (Medium): Standard response, target response within 10 minutes. Risk score 40-59.
  • P4 (Low): Non-urgent, target response within 30 minutes. Risk score 20-39. Eligible for AI callback.
  • P5 (Informational): Information only, target response within 1 hour. Risk score 0-19. Eligible for AI callback.

Use Cases#

Emergency call centres process incoming 911/emergency calls through a complete AI pipeline: transcribe, analyse sentiment, classify audio events, extract keywords, and assign dynamic priority for dispatcher routing.

PSAP operators managing capacity across multiple sites use surge redistribution to automatically balance call load when a major incident overwhelms a single centre, maintaining acceptable wait times without manual intervention.

Public safety agencies managing high call volumes during major events deflect P4/P5 calls to an automated callback queue, freeing live operators for genuine emergencies while ensuring every caller receives a response.

Emergency service planners use predictive staffing models to identify shift coverage gaps before they become service failures, acting on volume forecasts rather than reacting to overloaded queues.

Integration#

The AI Triage module integrates with other Argus modules and external systems:

  • Alert Management: Triage results and priority evaluations feed into the broader alert management system for unified incident tracking.
  • Case Management: High-priority calls can automatically generate cases for investigation follow-up.
  • Audit Trail: All priority evaluations, manual overrides, and callback outcomes are logged for compliance and quality assurance.
  • Command Center: Real-time surge status and staffing metrics display in the operational command centre dashboard.

Open Standards#

  • NENA i3 / NENA-STA-021 (NG9-1-1): The triage pipeline ingests calls via the NENA i3 Emergency Incident Data Object (EIDO) JSON schema, applying the standardised P1, P5 priority vocabulary and i3 service URN taxonomy (RFC 5031 / RFC 6443) to classify and route emergency calls to the correct PSAP queue.
  • SIP (RFC 3261) with PIDF-LO: Inbound NG9-1-1 calls arrive as SIP INVITE messages whose PIDF-LO bodies are parsed to extract geodetic and civic caller location, feeding the triage gate's risk scoring before a dispatcher answers.
  • RFC 7852, Additional Data Related to an Emergency Call: Subscriber, device, and communications information blocks defined in RFC 7852 and NENA-STA-012 are consumed during triage to enrich call context, supporting more accurate priority assignment for P1/P2 escalations.
  • SIPREC (RFC 7865 / RFC 7866): Call recordings ingested by the SIPREC recorder module supply the audio stream fed into real-time transcription and background audio classification; RFC 6873 metadata extensions carry participant and stream identity into the triage audit trail.
  • Real-Time Text / RFC 4103 + ITU-T T.140: The NG9-1-1 RTT bridge surfaces text-based emergency calls (deaf and hard-of-hearing callers) through the same triage gate as voice calls, with ITU-T T.140 character encoding over SIP/RTP and WebRTC data channels (RFC 8865).
  • ISO 639-1: Language codes are carried on every transcription session, enabling the multi-language speech-to-text pipeline to route non-English calls to the correct keyword dictionaries and AI classification prompt.
  • GraphQL (June 2018 specification): All triage operations, including priority scoring, rule management, bulk triage, manual override, and feedback, are exposed as authenticated Strawberry GraphQL mutations and queries, consumed by the dispatcher UI and third-party integrations alike.
  • W3C Verifiable Credentials v2: NENA i3 additional-data blocks attached to high-priority calls are wrapped in an Ed25519-signed Verifiable Credential envelope with deterministic SHA-256 canonicalisation, providing tamper-evident provenance for P1/P2 triage decisions in the audit trail.

Last Reviewed: 2026-04-02 Last Updated: 2026-04-14

Ready to Build?

Get started with our APIs or contact our integration team for support.