Last updated: August 7, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
New Orleans AI call triage is now handling part of the city's 911 traffic: the Orleans Parish Communication District (OPCD) is running Carbyne's Call Triage software to filter out duplicate calls about the same car crash before they reach a human dispatcher. Violent crime, medical emergencies, and anything life-threatening still go straight to a person, every time.
Short answer: New Orleans' 911 center is using Carbyne's AI Call Triage software to catch repeat callers reporting the same car accident, confirm the incident is already known, and free up dispatchers for calls that actually need a human. It's been running since 2023, limited to motor vehicle crashes, and OPCD reports it now recovers roughly two hours of call-taker time a day.
I spend my time testing the AI tools that answer phones, transcribe calls, and triage support tickets for businesses. A 911 center doing the same thing with life-or-death stakes got my attention fast. In my testing of consumer and small-business voice AI, the failure mode is almost always the same: the system is fine on the easy case. It falls apart the moment a caller says something it didn't expect. That's exactly the risk a 911 dispatch center can't afford. It's what makes this New Orleans AI call triage rollout worth picking apart in detail, not just skimming the headline.
What's Actually Happening in New Orleans
OPCD, which runs 911 for the city, was the first public safety agency in the country to deploy Carbyne's Call Triage feature, according to Carbyne’s own launch announcement from June 16, 2023. Carbyne CTO Alex Dizengof said at the time that the goal was to let "even a small or understaffed PSAP handle a major crisis" without dispatchers drowning in redundant calls.
The problem is a real one. A single pile-up on I-10 can generate 30 near-simultaneous 911 calls about the same wreck, and OPCD fields more than 1,000 emergency calls a day overall. Executive Director Karl Fasold put the human cost of that in blunt terms to StateScoop in 2024: "If you're calling about a heart attack, you go to the back of the queue" behind a stack of calls all reporting the same crash. He said the deciding factor for bringing in AI wasn't a budget line, it was burnout: "I made the decision that I was not going to burn out any more of my people," he told StateScoop on July 5, 2024.
By April 2026, the system had been running for roughly three years. Government Technology reported OPCD was recovering about two hours of call-taker time a day just from the AI handling car-accident-related calls. The district still operated at close to 40% below full staffing capacity, per Government Technology’s April 8, 2026 report. Tulane computer science professors Aron Culotta and Nick Mattei have been studying the deployment. OPCD also has a separate AI project in the works for its 311 non-emergency line, trained on three years of call recordings.
How Carbyne's Call Triage AI Actually Works
The system isn't a chatbot answering random calls. It only activates around a specific, known incident, and it hands off immediately the moment a call doesn't fit the pattern it's watching for.
1. A crash gets flagged in the dispatch system
When a car accident is logged in OPCD's computer-aided dispatch (CAD) system and calls about it start spiking, Carbyne's AI starts watching for new 911 calls that match the incident's location.
2. New callers near the scene get an automated check first
Callers within roughly 100 to 200 meters of the flagged crash, reports vary slightly by source, are met by an automated voice that asks if they're calling about that specific wreck.
3. Confirmed duplicates get a fast, honest answer
If the caller confirms it's the same crash and has no new details, the system tells them the incident is already known and help is on the way, then lets them disconnect without ever reaching a human queue.
4. Anything new, uncertain, or involved routes to a person
A caller who says no, who has new information, or who was actually involved in the crash gets sent straight to a live dispatcher, no different from calling 911 today.
5. Violent crime and medical emergencies never touch the AI
An OPCD spokesperson told Gadget Review the line is absolute: "Those will always be a human being, the first voice you hear," according to Gadget Review’s August 7, 2026 report. The AI only runs against the motor-vehicle-crash use case; the district has talked about expanding to other nonviolent incident types, but hasn't rolled that out at scale yet.
The Pilot's Numbers, By Report
Three different write-ups of this deployment, spread across almost three years, report different metrics because they're describing different phases of the same rollout. Here's what each one actually said, side by side.
| Source & Date | What It Measured | Key Numbers |
|---|---|---|
| Carbyne's own case study (covers an early 90-day window) | Events triaged, redundant-call reduction, time saved per incident | 3,500+ events triaged, 30% fewer redundant calls, up to 16 minutes saved per multi-caller incident, ~40 seconds faster average response |
| StateScoop, July 5, 2024 | Staffing and response-time context during a 4-month test | ~140 staff with 18 vacant call-taker roles, 70–80% of calls answered within 15 seconds against a 90% benchmark, AI output equal to roughly 2 full-time staffers |
| Government Technology, April 8, 2026 | Time saved after the system had been live for years | ~2 hours of call-taker time recovered daily; district running ~40% below full staffing capacity |
The direction is consistent even where the exact figures aren't: fewer redundant calls reaching a person, and measurable time given back to a chronically understaffed call center. A 2024 Carbyne/NENA survey found 82% of U.S. emergency communication centers report chronic staffing shortages, which is the backdrop that makes "2 hours a day" or "equal to 2 FTEs" worth taking seriously rather than dismissing as a rounding error.
What It Takes to Run Something Like This
You can't bolt this onto a 911 center over a weekend. OPCD and Carbyne spent three months training the model on real local 911 recordings before it went live, specifically to handle New Orleans accents and speech patterns rather than a generic voice model. It also needs a CAD system the AI can read incident data from in real time, a narrowly scoped use case (here, motor vehicle crashes only) rather than "handle everything," and a hard rule that keeps violent crime and medical emergencies routed to a human by default, not by exception.
That last part is the one I'd push back on if any vendor tried to skip it. In my testing of AI phone and voice systems for ordinary businesses, the products that hold up under real callers are the narrow ones with a clear human escalation path, not the ones promising to "handle anything." A 911 center just happens to be the highest-stakes version of that same design principle.
Questions Worth Asking Before Your Agency Copies This
If you run a call center, public or private, and this story makes you want to look at AI triage for yourself, these are the prompts I'd run past a vendor or through an AI assistant before signing anything:
- "What percentage of calls does your AI handle end-to-end versus escalate to a human, and how is that measured?"
- "Walk me through exactly what happens if a caller says something the model wasn't trained to expect."
- "How much local call data do you need before this performs reliably on my region's accents and slang?"
- "What's your hard-coded list of call types that always route to a human, and can I audit that list?"
Copy those into ChatGPT, Claude, or a vendor demo and use the answers as a checklist. A vendor that gets vague on the second and fourth questions is the one to slow down on.
Common Mistakes People Make Reading This Kind of Story
The biggest one is assuming "AI is answering 911 calls" means an algorithm is making dispatch decisions. It isn't. Carbyne's system filters duplicate reports of a single confirmed incident; a human dispatcher is still the one who decided a crash happened and sent units to it in the first place.
The second is assuming this replaces dispatchers. OPCD's own framing is the opposite: Fasold adopted it specifically to stop burning out the staff he already has, not to cut headcount, at a district still running well below full staffing.
The third is treating the 100-to-200-meter radius and "motor vehicle crashes only" scope as a permanent ceiling. Reports from 2024 and 2026 both describe planned expansion to other nonviolent call types, so the scope you read about today is a snapshot, not the finished product.
Tools That Make It Easier to Judge a Story Like This
You don't need to be a public safety expert to sanity-check a claim like "AI is triaging 911 calls." It helps to have spent real time with the AI voice and call-handling tools the same way you'd want a dispatcher to. My AI tool ratings and AI tool reviews hubs score chatbots and voice assistants against the same reliability and escalation questions I listed above. If you're curious how far consumer voice AI has come, my guides to using ChatGPT’s voice mode and using ElevenLabs cover the same speech-recognition and synthesis building blocks Carbyne's system relies on, just in a lower-stakes setting. Small businesses weighing an AI phone or support system of their own should start with my best AI tool for small business guide before picking a vendor. And if this story leaves you wanting the bigger picture on how city and state governments are betting on AI right now, my piece on governments making a dangerous bet on the AI boom covers that pattern beyond just 911 centers. My free AI tools roundup is a decent starting point if you want to test similar voice or triage logic yourself before spending anything.
My Take
What sold me on this isn't the AI, it's the scope. Carbyne's system doesn't try to be smart about ambiguous situations; it asks one narrow yes-or-no question about a crash the dispatch system already knows is happening, and it kicks anything uncertain straight to a person. That's the same design choice I look for when I review AI tools for paying businesses: narrow scope, hard-coded escalation, no attempt to sound human when the stakes are high. The numbers OPCD and Carbyne report, roughly two hours of staff time back a day inside a center running 40% below capacity, are believable precisely because the system isn't claiming to do more than that.
Frequently Asked Questions
Is Carbyne's AI making 911 dispatch decisions?
No. It only filters duplicate calls confirming a crash the dispatch system already logged. Human dispatchers still decide how to respond to every incident; the AI's job is to stop redundant callers from clogging the queue ahead of unrelated emergencies.
Does this replace 911 dispatchers in New Orleans?
No. OPCD Executive Director Karl Fasold adopted it to reduce burnout on an existing, understaffed team, not to cut staff. The district was still running roughly 40% below full capacity as of Government Technology's April 2026 report.
What happens if I call 911 about a medical emergency or a crime?
You reach a human dispatcher immediately. OPCD has confirmed violent crime and medical emergencies are always routed to a live person; the AI only operates on the motor-vehicle-crash use case.
How long has this system been running?
Since 2023. Carbyne's Call Triage feature launched with OPCD as the first deployment site in June 2023, and as of 2026 the district was preparing a separate AI system for its 311 non-emergency line.
Could my city's 911 center use something similar?
Possibly, but it took three months of training on local call recordings before this system went live in New Orleans. Ask any vendor exactly what percentage of calls get escalated to a human and what triggers that escalation before you sign anything.