Public Safety & Emergency ServicesIssue 305 records
Emergency Services
Tonight's edition looks at AI arriving in emergency call handling and EMS faster than results are published. Norfolk's E911 Division switched on AWS-powered AI for non-emergency lines with a 35–40% diversion forecast, and North East Ambulance Service in the UK is testing the market for an AI overlay on its existing call-centre telephony. In EMS, an ImageTrend case study says St. Johns County Fire Rescue uses AI Assist on 85% of reports, and Tanner Health became the first U.S. commercial site for FDA-authorized AI ECG in ambulances, citing study data, not its own. A new preprint shows generic AI agents often choose invalid hazard-analysis methods unless constrained.
- Evidence records
- 5
- Cross-source patterns
- 2
- Evidence classes
- 2 government evaluation2 vendor claim1 academic research
- Outcomes
- 4 emerging1 mixed
- Source freshness
- 4 new this fortnight1 undated
- Research completed
- 2026-10-06
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Emergency call centers are adding AI at the telephony layer, with forecasts and assurance questions but no measured results yet
Norfolk's E911 Division launched AWS-powered AI on non-emergency lines with a forecast of 35–40% diversion and no baseline published; North East Ambulance Service is testing feasibility, assurance needs and cost for an AI overlay on its existing Avaya telephony before any procurement. Interpretation: both are adding AI on top of existing call-handling infrastructure, and both leave the key measures, diversion achieved, safe escalation and failover, still to be shown.
Operating questionBefore adding AI to our call-handling telephony, what baseline, escalation test and failover evidence will we require, and when will we publish results?
Supporting evidenceNorfolk Police Department Emergency 911 (E911) Division, City of Norfolk, VirginiaNorth East Ambulance Service NHS Foundation Trust (UK)
Fire/EMS AI tools report adoption and borrowed accuracy figures, not local patient or time outcomes
St. Johns County Fire Rescue's AI Assist case study reports usage on 85% of reports and 422 rejections in 20,000+ applied completions, but no measured time or accuracy change; Tanner Health's AI ECG rollout cites a prior study (553 of 601 STEMIs vs 427) rather than its own results. Interpretation: acceptance rates and vendor-cited studies are not local evidence; agencies should baseline and audit before claiming benefits.
Operating questionWhat local baseline and audit (field accuracy, false activations, time to treatment) will we collect before and after turning on an EMS AI feature?
Supporting evidenceSt. Johns County Fire Rescue (Florida) / ImageTrendTanner Health (Georgia) / Powerful Medical
Evidence ledger
Full record · every source keeps its link and limitations
Norfolk E911 Division puts AWS-powered AI on non-emergency lines, projecting 35–40% call diversion while 911 stays human
The City of Norfolk announced on Oct 5, 2026 that its Police Department's E911 Division launched an Amazon Web Services AI-powered service for select non-emergency and informational calls. E911 staff answer over 250,000 police, fire and medical calls and about 150,000 administrative calls a year. The city expects the system to divert roughly 35–40% of non-emergency calls immediately. Callers who say 'I want to speak to an agent' are redirected to a non-emergency line for the next available call taker; all 911 calls stay with trained telecommunicators.
Why it matters, evidence and limitations
- Why it matters
- A large municipal PSAP that answers fire and medical as well as police calls is adopting a cloud contact-center AI front door; included here for its 911-center operating model (cross-stream with public safety because the division sits in the police department).
- Evidence and measured results
- Official agency announcement, not an evaluation (closest schema class used). The 35–40% diversion figure is a forecast; no baseline hold times, accuracy, misrouting or post-launch data are published.
- Limitations and uncertainty
- Forecast only; launch-day announcement; the 'speak to an agent' path returns callers to the human queue rather than an immediate answer; no detail on emergency detection on non-emergency lines.
Tanner Health deploys PMcardio Queen of Hearts AI ECG across ambulances, EDs and cath lab in first U.S. commercial rollout
A Sept 30, 2026 Business Wire release says Tanner Health is the first commercial U.S. deployment of Powerful Medical's PMcardio platform, bringing the Queen of Hearts AI ECG model, FDA De Novo-authorized in September 2026, to 12-lead ECGs in Tanner ambulances and emergency departments, integrated with Epic and EMS workflows. The AI is described as secondary decision support. The release cites a prior 1,032-patient U.S. study: 553 of 601 STEMIs found from the initial ECG vs 427 by standard triage, false positives 41.8% to 7.9%. No Tanner results yet.
Why it matters, evidence and limitations
- Why it matters
- Brings FDA-authorized AI ECG interpretation to the 911 ambulance point of first contact in a rural U.S. region, which changes prehospital STEMI recognition and cath-lab activation workflows.
- Evidence and measured results
- Vendor and health-system press release; accuracy figures come from a previously published multicenter study cited by the vendor, not from this deployment. No local outcome, time-to-treatment or false-activation data.
- Limitations and uncertainty
- Launch announcement days after FDA authorization; study figures are vendor-cited; regional health-system EMS may differ from fire-based or county EMS; no cost data.
ImageTrend case study: St. Johns County Fire Rescue uses AI Assist on 85% of fire and EMS reports, 422 rejections in 20,000+ completions
ImageTrend's case study (marked published September 2026) says St. Johns County Fire Rescue, with nearly 500 field personnel and about 67,000 incidents a year, uses ImageTrend Elite AI Assist (voice-to-text, image scanning of medication lists and insurance cards, auto-population) on 85% of fire and EMS reports, about 2.9 uses per incident. AI Assist populated 472,948 fields (44.7 per incident); more than 20,000 completions were applied with 422 rejected. Shorter completion time, less QI review and faster billing are stated without figures. A CQI Check feature is next.
Why it matters, evidence and limitations
- Why it matters
- One of the few fire/EMS AI documentation deployments reporting usage and acceptance counts at agency scale, relevant to any department weighing AI-assisted ePCR.
- Evidence and measured results
- Vendor case study with agency quotes. Counts (fields populated, completions, rejections) are system telemetry; time savings, accuracy and billing effects are asserted without baselines or numbers. Measurement period not stated.
- Limitations and uncertainty
- Vendor-authored; month-level date only; no stated timeframe for the totals; no independent accuracy audit; 'rejected completions' undefined.
arXiv HazardWeaver: generic LLM agent designs reach 22.7–41.1% on hazard-analysis route selection vs 89.4% with evidence-linked constraints
An Oct 2, 2026 arXiv preprint from Florida State University and Nokia tests whether LLM agents can choose valid scientific methods for natural-hazard analysis. On a 141-instance benchmark across seven single-hazard domains and four multi-hazard classes, adapted agent systems and control strategies such as ReAct, Reflexion and Plan-and-Solve scored 22.7–41.1% overall; HazardWeaver, which adds evidence-linked applicability conditions and typed tool compatibility, scored 89.4%. Removing either constraint cut Llama results from 93.8% to about 46–48%. Flood was the hardest shared track.
Why it matters, evidence and limitations
- Why it matters
- Emergency managers are being offered AI agents for hazard and impact analysis; this shows that general-purpose agent designs often pick invalid methods unless scientific applicability is encoded, and the authors say operational use needs local calibration.
- Evidence and measured results
- Controlled benchmark with Wilson confidence intervals, ablations across five model backbones and McNemar tests; authors' own system and benchmark; preprint, not peer reviewed; no live emergency operations.
- Limitations and uncertainty
- Frozen capability catalog and evidence index; benchmark designed by the system's authors; authors state operational forecasting and emergency decision-making need updated evidence, local calibration and domain review.
North East Ambulance Service opens £200,000 market engagement on an AI overlay for its existing Avaya Elite telephony platform
North East Ambulance Service published a preliminary market engagement notice on Sept 29, 2026 to test feasibility, capability, assurance needs, delivery complexity, indicative cost and market appetite for an AI-enabled overlay to its existing Avaya Elite telephony. It is valued at about £200,000 with responses due Oct 30, 2026. NEAS says the work will shape requirements for a future full telephony replacement aligned with emerging national standards, and that an options appraisal, business case, pilot or procurement may follow, but none is guaranteed.
Why it matters, evidence and limitations
- Why it matters
- International signal that ambulance call centers are approaching AI as an overlay on legacy telephony, with assurance and feasibility tested before buying; transferable to U.S. PSAPs on aging call-handling equipment.
- Evidence and measured results
- Official procurement notice as listed by an aggregator (wording cross-checked against the Find a Tender notice). States intent and scope only; no AI functions, vendors or results.
- Limitations and uncertainty
- Pre-tender engagement only; scope of the AI functions not described; UK NHS structures and standards differ from U.S. PSAPs; aggregator listing of an official notice.
How to read this edition
Source findings, measured results and limitations come from the cited publications. Patterns, operating questions, role takeaways and implementation considerations are Lighthouse Advisory interpretation, stated as questions to validate locally rather than guaranteed outcomes. Vendor and operator claims are labeled as claims. Full research method.
- Government evaluation
- A public body’s measured evaluation or documented pilot.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.
- Academic research
- Research produced through an academic institution or peer-reviewed venue.