When someone calls an emergency number, the most important decisions of the whole incident happen in the first ninety seconds — before any paramedic arrives. AI in ambulance dispatch is now sitting inside that window: listening to the call, flagging what the call-taker might miss, and helping decide which vehicle goes where. It is one of the most consequential and least discussed uses of artificial intelligence in public services today.
This article looks at what these systems actually do, what the evidence says about whether they work, and why one American city’s quiet deployment became a controversy in 2026.
What does AI actually do in an emergency dispatch centre?
AI in ambulance dispatch does three distinct jobs: it listens to live emergency calls to help recognise critical conditions, it forecasts where calls will come from so vehicles can be pre-positioned, and it optimises which unit is sent and by what route. These are separate systems solving separate problems, and they succeed to very different degrees.
Breaking that down:
- Real-time call analysis. Speech recognition transcribes the caller in real time while a machine learning model scores the conversation for signs of specific emergencies — most commonly out-of-hospital cardiac arrest. The system prompts the dispatcher rather than deciding anything itself.
- Demand forecasting. Models trained on years of historical call data, plus weather, day of week, events and population density, predict call volume by area and hour. Services use this to park ambulances near where the next call is statistically likely, instead of at a station.
- Dispatch and routing optimisation. Algorithms weigh live traffic, vehicle location, crew skill level and receiving-hospital capacity to pick the unit that will genuinely arrive first — which is often not the geographically closest one.
- Triage and diversion. The newest and most contested use: identifying low-acuity callers who could be routed to a nurse line or urgent care instead of an ambulance.
The core idea: AI in ambulance dispatch does not replace the call-taker, it changes what information the call-taker has in the seconds that matter.
How does AI detect cardiac arrest during an emergency call?
The AI transcribes the call as it happens and analyses word choice, the caller’s description of breathing, and audible background sounds to estimate the probability of cardiac arrest. If the score crosses a threshold, it alerts the dispatcher on screen, prompting them to ask confirming questions and start telephone CPR instructions sooner.
This matters because cardiac arrest is time-critical in a way almost nothing else is: survival odds fall sharply for every minute without CPR, and callers rarely say “cardiac arrest.” They say the person is snoring, gasping, or “breathing funny” — agonal breathing that is easy to mistake for breathing.
The best-known system is Corti, built by a Copenhagen company of the same name. In a 2019 study published in Resuscitation, Blomberg and colleagues found the machine learning model recognised out-of-hospital cardiac arrest in archived Copenhagen emergency calls with about 84% sensitivity, compared with roughly 73% for the human dispatchers who had handled those same calls. The European Emergency Number Association (EENA) ran pilots of the technology with emergency centres in France and Italy from 2018.
The result that gets left out
Here is the part that rarely makes the headlines, and it is the most important finding in the field. The same research team ran a randomised clinical trial and published it in JAMA Network Open in 2021. Across 654 confirmed cardiac arrest patients, dispatchers supported by machine learning alerts recognised 93.1% of cases, versus 90.5% for dispatchers using standard protocols — a difference that was not statistically significant.
The model itself still outperformed humans in that trial (85.0% versus 77.5% sensitivity). But giving its alerts to dispatchers did not measurably improve the dispatchers’ decisions. The authors concluded plainly that the trial “did not find any significant improvement” in recognition.
The lesson generalises well beyond ambulances: a model being more accurate than a human does not mean a human plus that model is more accurate than the human alone. The handoff — how and when the alert appears, and whether the person trusts it — decides everything. It is the same design problem faced in AI decision support for air traffic control, where the tool has to earn a trained professional’s attention without drowning them in prompts.
Why did Seattle’s AI dispatch system cause a controversy?
Seattle became the test case for the governance side of this technology in June 2026, when reporting revealed that the Seattle Fire Department had been running Corti’s AI on every incoming medical 911 call since December 2023 without telling the public.
The system did not just listen. It prompted dispatchers in real time to route some callers to a nurse consultation line — operated in another state — instead of sending an ambulance. According to reporting summarised by GeekWire, the department never disclosed the deployment, never submitted it for review under Seattle’s own surveillance ordinance, and never sought city council approval.
The effectiveness numbers were shaky too. The department’s medical director was quoted in a 2024 Corti press release saying nurse-line routing had risen 50% since deployment; a department spokesperson later corrected that to 32%. Neither figure has been independently verified.
Nothing here suggests the underlying technology is bad. What it shows is that a triage tool which changes whether an ambulance is dispatched is a policy decision wearing a software costume — and it was treated as a procurement detail. Callers had no idea an AI was scoring their emergency.
Where is AI ambulance dispatch actually being used?
Adoption is far patchier than the vendor marketing suggests, and it varies enormously by country.
- Denmark and continental Europe: the origin point for call-analysis AI, via Corti and the EENA pilots in France and Italy.
- United States: Seattle is the highest-profile medical-call deployment; predictive demand modelling and routing optimisation are more widely and less controversially used across US EMS agencies.
- United Kingdom: strikingly low uptake. In a Freedom of Information exercise covering the UK’s ambulance service trusts, 10 of the 11 trusts that responded said they were not using any AI for call handling, and roughly two-thirds had no plans to.
- India: the direction of travel is infrastructural rather than acoustic. India’s 2026 National Ambulance Services Guidelines recommend Integrated Command and Dispatch Centres with GPS-enabled tracking, digital call management and intelligent dispatch — the plumbing that AI routing needs before it can do anything useful.
- Australia and elsewhere: demand forecasting and ambulance-positioning models are the common entry point, since they improve response times without touching clinical judgement.
The pattern is consistent: services adopt the boring logistics AI first, because it carries far less clinical and legal risk than an algorithm that influences whether a patient is seen at all.
What are the real risks?
Three stand out. Under-triage is the serious one — an algorithm nudging a genuinely sick caller toward a nurse line delays care for exactly the person who could least afford it. Automation bias is the subtle one: dispatchers may start deferring to the model’s silence as reassurance, which is precisely the failure the JAMA trial’s flat result hints at. And consent and disclosure is the one Seattle stumbled on: people calling an emergency number are in no position to opt out of anything.
Accuracy is not the whole safety argument. As with AI screening in medical imaging, what determines patient outcomes is the workflow wrapped around the model — who reviews it, what happens on a disagreement, and who is accountable when it is wrong.
Frequently Asked Questions
Does AI decide whether an ambulance is sent?
In current deployments, no — a human dispatcher makes the final call. But AI systems that prompt dispatchers to route callers to a nurse line, as in Seattle, clearly influence that decision. The distinction between advising and deciding gets thin under time pressure.
Is AI better than humans at spotting cardiac arrest on a call?
On archived call recordings, yes — models have shown roughly 84-85% sensitivity versus about 73-78% for dispatchers. However, a randomised trial found that giving those alerts to dispatchers in real time did not significantly improve their recognition rate.
Do callers know an AI is listening to their emergency call?
Usually not. Seattle ran AI on every medical 911 call for over two years without public disclosure. Disclosure rules vary by jurisdiction and are largely unsettled, which is why the Seattle case attracted so much scrutiny.
What is ambulance demand forecasting?
It is the use of machine learning on historical call data, weather, and local events to predict how many emergency calls will come from each area and hour. Services use the forecast to position ambulances closer to likely incidents before calls arrive.
Can AI reduce ambulance response times?
Indirectly, yes. Demand forecasting and routing optimisation shorten travel time by pre-positioning vehicles and selecting the unit that will genuinely arrive first. The gains are typically incremental percentage improvements rather than dramatic drops.
