The pitch is simple and it is true: let the AI answer the phone so a human stays free for the emergency. In Craighead County, Arkansas, the AI answered a crash call too, and it told the caller to go online and file the accident report. Every collision gets an officer. That is not a preference, it is the entire reason the number exists. Automation took a wreck, and turned it into a web form.
The published numbers are the good ones. Between September 29 and October 28 the system, a non-emergency call-taker the center calls Lezle, picked up close to 5,000 calls and closed out more than half of them without a dispatcher touching the line. Over those 30 days the county’s call load fell 56%. E-911 dispatcher Terra Summers put the labor saved at roughly 70 hours, then did the arithmetic out loud: at five minutes a call, 70 hours is a lot of calls.
She is right about that. What the rollout does not have is the number that would tell a resident whether the thing is safe, which is the count of calls it got wrong. Dispatchers log bad outcomes to upper administration, which either patches the workflow or forwards it to the programmer. That is a bug tracker. A bug tracker is a reasonable way to run a phone system. It is a strange way to run the door to 911.
Terra Summers — Seventy Hours Is Real Relief, and Relief Is Not the Same as Safe
Start with the staffing, because the staffing is what actually built this. When Lezle began its trial on September 2, Jonesboro’s dispatch center employed 22 full-time dispatchers on 12-hour shifts, and some shifts ran with three people on duty. Those 22 humans held the radio for more than 200 officers across nine police departments and 11 fire departments. Savannah Tinsley, two years into the job, described the job as running a radio while answering 911s while answering the non-emergency line, and said a system that takes one of those three jobs off her desk relieves real stress.
That is not spin. Three people covering nine departments is a staffing model that has already lost. The system is named for Lezle Wilson, a Jonesboro dispatcher who died in 2023, which tells you how the county feels about the people who used to sit in that chair. Nobody at that center is lying about the relief.
The relief is also exactly why the error matters. A center that has been running on three bodies has no slack to add a verification step, so the AI’s mistakes get caught the way they got caught here: after the caller has already been sent somewhere.
The Crash Report — The Dangerous Failure Is Not a Wrong Sentence, It Is a Wrong Door
A wrong answer can be argued with. A caller who is told the wrong fee for fingerprints will hear the correction and shrug. A wrong routing decision cannot be argued with, because the caller believes the transaction is complete. She hangs up, she files the report from a parking lot, and nobody is dispatched. The county’s own dispatchers describe this as an early glitch, and early is doing a lot of work in that sentence, because the failure was not exotic. It was a car crash, the single most common thing a police radio exists to answer.
The safeguard that is supposed to prevent it is a keyword list. Words like emergency, breathing and fire push a caller straight to a live dispatcher. That is the load-bearing wall of this entire architecture, and keyword lists fail precisely where frightened people are least articulate: a caller who says my car is in the ditch and never says the word emergency, a caller shouting over traffic, a caller with an accent the model was not trained to flatter. The handoff works when the caller says the right word. The caller does not know there is a right word.
Cristy Hundley — The Promise That It Never Touches 911 Lasts Exactly as Long as the Routing Does
The center’s training coordinator told K8 News at launch that the system will never answer a 911 call, and that anything which turns out to be an emergency goes to a person immediately. Both statements are almost certainly true today. Neither is a property of the software. They are properties of a phone configuration and an intent classifier, and those are two things a budget cycle can quietly change.
The public has been given a promise about scope and no data about accuracy. Jonesboro has published the 56% and the 70 hours. It has not published a false-negative rate, a count of misrouted calls, or an average delay on the calls Lezle handed back. The most generous reading is that nobody has been asked for it yet.
Arlington County — Seventy Percent of the Calls Were Never Emergencies, and That Is the Whole Case
The honest argument for this technology is not the flashy one. It is paperwork. Arlington County’s account of its own AI diversion program, published through the federal 911 program office, records 264,414 administrative calls against 78,552 emergency calls in 2022, which is 70.3% of the volume arriving at a line that was never built for it. Tow notifications and out-of-service fire alarm notices alone ate about three hours of every shift. The county’s AI diversion removed more than 22,000 calls in six months.
That is the case, and it is a good one. American non-emergency lines are drowning in clerical work, and the people best placed to say so are the dispatchers losing their voices to it. Brown County, Wisconsin, reports the same shape from the other end: 160,000 non-emergency calls a year, with its assistant handling around 60% of them. The problem is real. The question was never whether the volume exists.
Nicholas Sides — The Resident Is Not Asking Whether It Works, He Is Asking Who Answers When It Doesn’t
A Jonesboro resident interviewed for the October report said he has not called the system himself, is generally wary of AI, and thinks it can be a useful tool if it is used the right way. That is the whole consent problem compressed into one sentence. The public is being asked to accept a system whose errors are discovered by the people it fails, one at a time, with the fix applied after the fact and never published.
Summers’ position is the pragmatic one and she is probably correct about it: the system improves, it is still learning, and on balance it should stay. A center staffed at three people a shift is not going to hand back a tool that returns 70 hours a month, and it should not have to. But keeping it and measuring it are not the same decision, and only one of them is currently being done.
The fix that would actually matter is small and unglamorous. Publish the error log. Publish the count of crash calls that resolved to a web form. Write a hard rule that any call containing a vehicle, an injury or a collision can never terminate in a self-service workflow, because the cost of a false transfer is a human being waiting in a parking lot for a police car that was never sent. The counties racing to install these assistants are all making the same bet on the same keyword list, and every one of them is doing it without a published number to check it against.
The software did not make a mistake a tired dispatcher would never make. It made a mistake a tired dispatcher makes at the end of a twelve-hour shift, which is the honest comparison — and then it made it at nine in the morning, without a tremor in its voice, on the call that decides whether somebody’s bad day becomes a wreck on a police blotter or a form nobody reads. The AI-mekna is not the villain here. The missing number is.
Sources: KAIT 8 News, The Cool Down, 911.gov / Arlington County.