What problem does it solve?
Emergency communications centres work under constant volume: Baltimore answers about 1.4 million 911 calls a year, and at Galt Police Department a single dispatcher is sometimes on duty alone. Call takers must understand panicked, noisy or non native callers, get an address, follow a protocol and type everything at once, often while also handling radio. Critical conditions are missed: Copenhagen EMS researchers note that dispatchers fail to identify roughly a quarter of out of hospital cardiac arrests. Callers who do not speak the local language wait for a telephone interpreter. Quality assurance often covers only a sample of calls (roughly 30% in Baltimore before automation), with feedback arriving long after the call.
The stakes are high. A wrong alert, a mistranslation or a missed cue can cost a life, which is why the EU AI Act lists emergency call classification and dispatch as high risk.
- Copenhagen Emergency Medical Services researchers report that emergency medical dispatchers fail to identify approximately 25% of out of hospital cardiac arrests.Effect of Machine Learning on Dispatcher Recognition of Out-of-Hospital Cardiac Arrest During Calls to Emergency Medical Services: A Randomized Clinical Trial (2021)
How does it work?
- Transcribe live. Speech recognition produces a running transcript of the call on the call taker's screen, with key details (address, weapons, symptoms) highlighted.
- Translate. For callers in another language, the system detects the language and shows a translated transcript, or voices the call taker's questions in the caller's language.
- Flag critical conditions. A model listens for patterns of time critical conditions, such as cardiac arrest, and alerts the call taker to consider the matching protocol.
- Summarise and document. At the end of the call, the system drafts a summary for the incident record, which the call taker edits.
- Review quality. Automated QA checks every call against protocol and flags calls for supervisor review and coaching.
- Keep humans in charge. Call takers decide the triage category, protocol and dispatch; the AI never dispatches or downgrades a call.
- Audience
- Employee facing
- Autonomy
- Assist
- Adoption
- Early adopters
- Channels
- Phone and voice, Agent desktop
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Risk and loss reduction, Speed and cycle time, Inclusion and access, Employee productivity.
Indicative value
An emergency communications centre handling 1 million calls a year
USD 333,333 to USD 1.5 million
Call taker time released, valued at cost per year
How this is calculated
Formula: calls * minutesSaved / 60 * hourlyCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Emergency calls per year calls, calls per year | 1,000,000 | 1,000,000 | The reference centre; Baltimore answers about 1.4 million a year. |
| Call taker minutes saved per call on documentation minutesSaved, minutes per call | 0.5 | 1.5 | Editorial assumption for summaries replacing manual narrative entry. No public benchmark states this yet. |
| Fully loaded call taker cost per hour hourlyCost, USD per hour | 40 | 60 | Editorial assumption. Replace with your own cost. |
What it leaves out: Values documentation time only. It leaves out the value of faster recognition of critical conditions, interpreter costs avoided, full QA coverage and the cost of the system; released time in understaffed centres usually goes to answering calls faster, not to savings.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Copenhagen Emergency Medical Services
Denmark · Government and public sector · 2018
From September 2018 to December 2019, a machine learning model listened to 112 emergency calls at Copenhagen EMS through speech recognition and flagged suspected out of hospital cardiac arrest. Because of downtime, it processed 169,049 of the 226,130 calls the service received (74.7%). It flagged 5,847 calls as suspected cardiac arrest, and the 5,242 eligible calls were randomized: dispatchers in one group saw an alert, the other group worked as usual. The model alone had higher sensitivity than dispatchers without alerts (85.0% against 77.5%) but a much lower positive predictive value (17.8% against 55.8%), and dispatchers with alerts did not recognize significantly more confirmed arrests (93.1% against 90.5%, P = .15). It shows that a model with higher sensitivity did not, in this trial, lead to better dispatcher recognition.
No outcome disclosed.
Galt Police Department
United States · Government and public sector · 2024
Galt Police Department serves 26,000 residents with eight dispatch staff, who handle nearly 30,000 calls a year, more than 73% of them non emergency. Since 2024 an AI agent from Prepared answers the ten digit non emergency line, works out what the caller needs, resolves it or routes it to the right resource, and transfers any genuine emergency immediately. On 911 calls, dispatchers get a live transcript, an AI summary and key details highlighted as the call happens. During a shooting in Galt, the agent handled the incoming non emergency calls in the background while dispatchers managed the response.
- Contact deflection: 73%, share of call volume handled before reaching a dispatcher
"With 73% of call volume now handled before it reaches a dispatcher's headset, the calls that do come through are the ones that genuinely need a human."
Claimed by: vendor
Baltimore City 911 (Emergency Communications)
United States · Government and public sector · 2022
Baltimore's emergency communications centre answers about 1.4 million 911 calls a year. Since partnering with Prepared in early 2022, call takers see a live transcript, AI summaries and highlighted key details on every call, which helps with addresses and callers who are hard to understand. Non English calls are transcribed and translated in real time, and for Spanish calls operators can dial in an automated voice translator instead of a third party interpreter. Automated QA now reviews every call, where a contractor previously reviewed roughly 30%. The case study also carries the unattributed line that the tool is "about 98% accurate", without saying what was measured or how, so it is not recorded as an accuracy figure.
- Quality score uplift: 12%, QA scores since automated QA was deployed; the QA method changed at the same time (sampling of about 30% of calls replaced by automated review of all calls), so before and after scores may not be comparable
"Since deploying Automated QA, Baltimore has seen a 12% improvement in QA scores."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Recorded calls with outcomes for testing (for example confirmed cardiac arrest)
- Current protocols (medical, fire, police) and local address data
- Language mix of callers, to prioritise translation
Systems to integrate
- Call handling system and audio stream
- Computer aided dispatch (CAD) for incident records
- Location services and maps
- QA and training systems
Complexity: High
Integration with call handling and CAD systems, real time latency, noisy audio and many languages, plus the need for clinical and operational validation of every alert, make this a demanding public sector deployment.
- 1
Start with transcription and summaries
Live transcripts and draft summaries support every call and change no decision; staff in Baltimore and Galt describe both as practical help with addresses and documentation.
- 2
Add translation with a fallback
Offer machine translation for the most common languages, with a clear way to bring in a human interpreter when the call is complex or the translation is doubtful.
- 3
Treat alerts as clinical interventions
Before switching on alerts for conditions such as cardiac arrest, run a controlled trial on your own calls. Copenhagen's randomized trial showed a model that beat dispatchers on sensitivity did not significantly improve dispatcher recognition.
- 4
Design the alert for the call taker
A low positive predictive value floods call takers with false alarms. Tune thresholds with dispatchers and measure whether alerts are acted on.
- 5
Use automated QA for coaching
Review every call against protocol and feed findings into training. Baltimore moved from a third party reviewing roughly 30% of calls to automated QA on all of them.
Guardrails
- The AI never dispatches, downgrades or closes a call; call takers decide
- Alerts are advisory, logged and evaluated against confirmed outcomes
- Human interpreter always available as a fallback to machine translation
- Transcripts and summaries edited and approved by the call taker before they enter the record
- Fail safe design, so an outage of the AI never blocks call handling
KPIs to instrument
- Recognition rate of target conditions with and without alerts, on confirmed outcomes
- Alert positive predictive value and the share of alerts acted on
- Time to address confirmation and to dispatch
- Transcription and translation accuracy on a sampled set of calls
- QA coverage and protocol compliance scores
Human in the loop
Call takers and dispatchers make every triage and dispatch decision. Medical directors approve any clinical alert and its thresholds; supervisors review AI assisted QA findings before they reach staff files; every alert and translation is auditable against the call recording.
Common failure modes
- Better model, same outcome
- In Copenhagen the model had higher sensitivity for cardiac arrest than dispatchers (85.0% against 77.5%), but alerting dispatchers did not significantly improve their recognition. Measure the human outcome, not the model.
- Alert fatigue
- Alerts with low positive predictive value (17.8% in the Copenhagen trial, against 55.8% for dispatchers) are easy to discount. Tune thresholds and track actions on alerts.
- Mistranslation under pressure
- A wrong word in a translated address or symptom can send help to the wrong place. Show the original and translation, confirm critical details and keep interpreters available.
- Dependence during outages
- Staff who rely on AI transcripts can lose practice in manual call taking. Keep manual procedures practised and the system fail safe.
What are the risks and rules?
EU AI Act
High risk
Annex III point 5(d): AI systems intended to evaluate and classify emergency calls or to dispatch or set priority for emergency first response services (police, fire, medical aid) are high risk. Pure transcription that performs a narrow procedural or preparatory task may fall outside it under the Article 6(3) exceptions, but alerts that influence triage are in scope. An AI agent that speaks with callers directly, for example on a non emergency line, must also tell them they are interacting with AI (Article 50).
Guidance
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 5(d) lists emergency call classification and dispatch as high risk.
- Article 27, fundamental rights impact assessment for high risk AI systems (European Union, Europe). Deployers that are bodies governed by public law, or private entities providing public services, must assess the impact on fundamental rights before first using a high risk system such as one covered by point 5(d).
- NIST AI Risk Management Framework (NIST, North America). A framework for mapping, measuring and managing risk that US public safety agencies can apply to call taking AI.
Controls to put in place
- Controlled evaluation or trial on local calls before any alert goes live
- Conformity assessment, risk management and logging as required for high risk AI in the EU
- Continuous monitoring of alert performance against confirmed outcomes
- Documented fallback procedures and regular drills without the AI
- Clear records of which AI outputs the call taker saw on each call
Frequently asked questions
- Does AI improve emergency call triage?
- The best public evidence is mixed. In Copenhagen's randomized trial on 112 calls, a model flagged cardiac arrest with higher sensitivity than dispatchers (85.0% against 77.5%), but dispatchers who received alerts did not recognize significantly more cases. For transcription, translation, summaries and automated QA, the evidence on this page comes from vendor case studies from US centres such as Baltimore, which report practical benefits but no controlled comparison.
- Is emergency call AI high risk under the EU AI Act?
- Yes, when it evaluates or classifies emergency calls or sets dispatch priority (Annex III point 5(d)). That brings risk management, logging, human oversight and conformity requirements, and public bodies must also carry out a fundamental rights impact assessment (Article 27).
- Where do centres start?
- With live transcripts, summaries, translation and automated QA, as Baltimore did, and with AI on the non emergency line, as Galt Police Department did to keep routine calls away from dispatchers.
How to cite this page
Blits.ai AI Use Case Library, "AI support for emergency call triage (112 and 911)", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/emergency-call-triage-support. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published