New Orleans is putting AI on 911 calls to handle surge backlogs

The sheer volume of emergency calls during a crisis often creates a deadly bottleneck where people are left on hold while waiting for a human dispatcher. New Orleans is tackling this by integrating AI to triage 911 calls specifically when the system hits a backlog. Instead of a dial tone or a recording, the AI steps in to categorize the urgency of the call, ensuring that a cardiac arrest or a house fire gets prioritized over a noise complaint or a stolen bike.

This is a high-stakes application of an LLM agent where the cost of a "hallucination" isn't just a wrong answer—it's potentially a lost life. For this to work in a real-world deployment, the AI workflow has to be incredibly lean. It isn't replacing the human dispatcher; it's acting as a sophisticated filter. The system listens to the caller, extracts the core emergency, and flags the priority level for the human operator who is currently overwhelmed.

If we look at this from a prompt engineering perspective, the challenge is likely around "intent classification" under extreme stress. People calling 911 aren't speaking in clear, structured sentences; they are screaming, crying, or speaking in fragments. The AI needs to be tuned to recognize urgency through sentiment and keywords without getting tripped up by the chaos of the audio.

Implementing this requires a very specific technical stack to ensure low latency. You can't have a 3-second lag while a cloud API processes a request when seconds matter. They are likely using a combination of speech-to-text (STT) and a highly optimized small language model (SLM) to keep the processing local or on a dedicated edge server.

For anyone looking to build a similar triage system, a basic logic flow for the AI agent would look something like this:

{
  "triage_logic": {
    "priority_1": ["unconscious", "not breathing", "active shooter", "chest pain"],
    "priority_2": ["car accident no injuries", "water leak", "broken window"],
    "priority_3": ["parking dispute", "noise complaint", "lost pet"],
    "action": "Route Priority 1 to immediate human dispatcher; queue Priority 3 for delayed response."
  }
}

The biggest hurdle here isn't the code, but the trust. Moving from a human-only system to an AI-assisted one in emergency services requires a massive amount of validation. However, the alternative—leaving people on hold during a disaster—is far worse. This is a great example of how AI can move beyond chatbots and productivity tools into critical infrastructure.