AI emergency dispatch systems are already listening in on 999 calls across parts of the UK, and the numbers behind why explain the urgency: human call handlers in Copenhagen recognised out-of-hospital cardiac arrest correctly in roughly 73 percent of cases, while an AI system listening to the same calls caught it in 93 to 95 percent, and did so around 30 seconds faster. Every minute of delay before CPR begins cuts a cardiac arrest patient’s chances of survival by roughly 10 percent, which makes that gap a matter of who lives and who doesn’t, not a marginal efficiency improvement.
That single, well-documented use case, AI listening for cardiac arrest, is now being adapted for the Welsh Ambulance Service through a research partnership with the University of York. It is also the clearest example of a much broader shift already underway: AI systems increasingly deciding, in the first seconds of a 999 call, how urgently your emergency should be treated.
How AI Emergency Dispatch Actually Listens to a 999 Call
The ASSIST project, a partnership between the University of York, the Welsh Ambulance Service, and Danish company Corti, is adapting an AI platform first piloted in Copenhagen for use inside a UK 999 control room. The system runs continuously in the background of an emergency call, using speech recognition and machine learning to analyse keywords, speech patterns, and non-verbal sounds, including irregular breathing, gasping, or silence, that indicate a caller’s own account of the emergency may not match what is actually happening.
Research behind the project found that at least 25 percent of out-of-hospital cardiac arrests go unrecognised by call handlers using standard questioning alone, often because a distressed bystander describes symptoms inaccurately or the underlying condition is genuinely difficult to identify from a verbal account.
Crucially, the AI does not replace the call handler or make the dispatch decision itself. It flags a possible cardiac arrest to the human operator, who then decides whether to act on that signal, typically by walking the caller through CPR instructions immediately rather than waiting for further confirmation. That design, AI as an early-warning layer rather than a decision-maker, is deliberate, and it reflects a lesson the industry learned the hard way from earlier, less transparent systems.
The Accountability Question Nobody Has Fully Answered
When Corti’s technology first drew wider attention, the World Economic Forum’s own coverage raised the exact question that still shadows every AI dispatch system today. Kay Firth-Butterfield, then head of AI and machine learning at the WEF, asked directly who is liable if the machine gets it wrong: the AI manufacturer, the human being advised by it, or the centre using it.
She framed the underlying dilemma precisely: these systems need to meet standards of transparency and accountability, and it remains genuinely unresolved how much better than a human an AI system needs to be before relying on it, or not relying on it, becomes the negligent choice.
That question has not gone away because the technology improved. It has become more pressing, because AI dispatch tools are now spreading into far more categories of emergency call than cardiac arrest detection alone, into general triage, incident classification, and threat assessment across entire control rooms, according to reporting on how UK emergency services are piloting these systems. The more categories of decision an AI system touches, the harder it becomes to audit any single flagged or unflagged call after the fact.
Why 999 Triage Needed Fixing in the First Place
The case for AI assistance in emergency dispatch does not rest only on cardiac arrest statistics. Nuffield Trust data tracking NHS ambulance performance shows how much strain the underlying triage system is already under. Category 2 calls, covering emergencies like suspected heart attacks, strokes, and sepsis, carried an average response time of 29 minutes 13 seconds as of May 2026, against a national target of 18 minutes.
Category 3, urgent but not immediately life-threatening calls, averaged 1 hour 38 minutes with a 90th-percentile wait of over four hours. Getting the initial triage category right the first time, rather than needing to escalate a call after a delay, matters enormously when the system as a whole is already running well behind its own targets.
A separate structural problem compounds the triage challenge: research cited by the Tony Blair Institute’s analysis of NHS AI navigation found that 38 percent of patients transported to hospital by ambulance in England were not subsequently admitted, suggesting a meaningful share of ambulance dispatches may be responding to calls that, with better initial information, could have been directed toward a different, more appropriate service.
AI tools are being pitched as a way to close that gap from both directions: catching the genuine emergencies human triage protocols miss, while more accurately routing calls that do not need an ambulance at all.
Beyond Cardiac Arrest: Where Else This Technology Is Headed
Cardiac arrest detection is the most rigorously documented application of AI emergency dispatch, but it is deliberately the narrowest one. The same underlying speech-analysis approach is being extended by Corti and competing systems toward broader incident classification, distinguishing a genuine stroke from a less urgent presentation, flagging signs of sepsis, and helping call handlers verify address and location details more quickly during chaotic calls. Each of those extensions carries a version of the same trade-off documented for cardiac arrest: a narrower, well-defined signal is easier to validate rigorously than a broad, general-purpose triage judgement, and the evidence base thins out considerably the further a system moves from the single, well-studied cardiac arrest use case.
That expansion pattern is worth watching closely, because the strength of the evidence behind AI 999 systems currently rests almost entirely on the cardiac arrest results. A system marketed as an “AI triage assistant” broadly may be leaning on the credibility of that narrow, well-validated capability while quietly extending into categories of judgement that have not been tested with anything like the same rigour, a distinction easy to lose once a product ships as a single integrated platform rather than a specific, audited feature.
What This Means for the Call You Might Make
If you call 999 in a region using an AI-assisted triage system, the practical reality is that a machine-learning model may be listening alongside the human call handler you are speaking to, analysing your voice, word choice, and any background sounds for signals you are not consciously providing.
Current deployments are explicit that the AI does not make the final dispatch decision and cannot override a trained call handler’s judgement, but it can, and does, surface a concern that changes what the handler does next, sometimes before the caller has finished describing the situation.
That listening extends beyond the words you choose to say. Systems like Corti are built specifically to pick up on non-verbal cues, gasping, laboured breathing, silence where a response was expected, precisely because a caller in genuine crisis is often too distressed, or too far from the patient, to describe what is actually happening accurately. The technology’s core value proposition is built on capturing exactly the information a stressed human caller cannot reliably articulate.
The Same Human-in-the-Loop Debate Playing Out in a Different Emergency Setting
The tension in AI 999 triage, a genuine, measurable safety improvement paired with unresolved questions about accountability, over-reliance, and audit, mirrors almost exactly the pattern LiveAIWire found in AI lifeguard systems monitoring pools and beaches. Both technologies exist because a human alone, however well trained, reliably misses a meaningful share of genuine emergencies under real-world conditions, glare and blind spots at a pool, panic and inaccurate description on a phone call.
Both are explicitly designed as a supplementary layer rather than a replacement for trained staff. And in both cases, the industry itself acknowledges the open risk that widespread deployment could quietly erode the underlying human skill, or the staffing levels, the AI was only ever meant to support. That same pattern of technology genuinely helping in a narrow, well-validated task while raising broader oversight questions runs through LiveAIWire’s coverage of the AI doctor dilemma, where identical diagnostic technology improved outcomes substantially in some deployments and missed critical conditions in others, depending entirely on how much genuine human oversight surrounded it.
The always-listening design at the core of AI dispatch systems also connects to a broader concern LiveAIWire has traced in the expansion of AI-enabled monitoring across public life, where technology installed for one clearly beneficial purpose, in this case genuinely saving lives during a medical emergency, can create a broader capability, continuous audio analysis of some of the most private and vulnerable moments a person will ever experience, that operates with less public scrutiny than its intrusiveness would warrant in almost any other context.
A 999 call is not a conversation most people expect to be analysed by a machine learning model in real time, even when that analysis genuinely improves their chances of survival.
Who This Technology Serves Best, and Who It May Not Yet Reach
AI dispatch tools trained primarily on adult cardiac arrest calls carry the same population-mismatch risk LiveAIWire has documented in other AI safety deployments. Our coverage of AI elderly care technology found that safety systems are frequently designed and validated around a narrower population than the one that ultimately depends on them, with the specific needs of older adults, who make disproportionate use of emergency services and often present with atypical or harder-to-describe symptoms, added to the design only after the fact, if at all.
A model trained predominantly on clearer-cut cardiac arrest presentations may perform less reliably on the more ambiguous, multi-symptom calls that older callers and their families are statistically more likely to make.
What This Means for You
The practical takeaway from the evidence gathered here is not that AI-assisted 999 triage should be treated with suspicion, since the underlying cardiac arrest detection numbers are genuinely strong and independently documented.
It is that the technology’s presence does not change what you should do on your end of the call: describe symptoms as clearly and specifically as you can, follow the call handler’s instructions immediately rather than waiting to be asked twice, and understand that a machine listening in the background is there to catch what a distressed human voice might miss, not to replace the judgement of the trained person you are actually speaking to.
If your local ambulance service publishes information about AI use in its control room, and a growing number now do, that disclosure is worth reading, since it is currently the only way most callers will ever know a system like this was part of their own emergency.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity, and the social impact of emerging technology. He publishes daily at LiveAIWire.com.