Public Safety & Emergency ServicesIssue 275 records
Emergency Services
Tonight’s edition tracks PSAP automation scale-up against disclosure and executable-AI controls. Oct 2 GMToday reports Waukesha County’s Ava AI has handled 91,892 calls since January with ~53% automated under continuous dispatcher review. Oct 1 Powerful Medical–corpuls partnership aims to put Queen of Hearts OMI AI-ECG into EU/UK/CH ambulances. Vanderbilt/MNDEC’s GenAI 911 training deployment shows formal-methods assessment beating pure LLM baselines (94.3–95.9% vs 87.8%). Against that build-out, EMS1/Seattle Times documents multi-year Corti 911 medical-call AI without public disclosure or metrics, and DisasterBench shows tool-mismatch/parameter-binding errors dominating first failures in LLM disaster-response planning.
- Evidence records
- 5
- Cross-source patterns
- 3
- Evidence classes
- 2 independent reporting2 academic research1 vendor claim
- Outcomes
- 2 emerging2 cautionary1 effective
- Source freshness
- 2 new this fortnight2 older, newly relevant1 recent
- Research completed
- 2026-10-04
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
PSAP AI is absorbing non-emergency load at scale while training systems try to grow the human capacity behind it
Waukesha’s Ava reports ~53% automation across 91,892 handled calls with every action still dispatcher-reviewed, aiming to free staff for 911. Nashville’s GenAI trainer scaled to 190 users/1,120 sessions with formal protocol assessment—showing agencies are investing in both call-volume relief and trainable human oversight. SLED PSAPs should budget verifier capacity and documented escalation SOPs alongside any conversational AI go-live.
Operating questionIf Ava-like automation hits 50%+ of our non-emergency volume, what dispatcher review SLAs and training throughput must we fund so 911 answer-time gains are real?
Supporting evidenceWaukesha County Communications Center / Waukesha Freeman (GMToday)Vanderbilt University / Metro Nashville Department of Emergency Communications
911 AI assist is expanding faster than public disclosure and outcome metrics
Seattle’s Corti deployment analyzed medical 911 calls for years without caller notice, public review, or shared success metrics. Meanwhile Waukesha is openly discussing Ava automation rates and Wisconsin peer adoption. As more PSAPs automate non-emergency and assist triage, disclosure, surveillance-ordinance review, and publishable KPIs are becoming the differentiator between capacity tools and legitimacy risks.
Operating questionBefore our next 911 AI assist or conversational agent expands, what will callers be told and which outcome metrics will we publish?
Supporting evidenceSeattle Fire Department / The Seattle Times (via EMS1)Waukesha County Communications Center / Waukesha Freeman (GMToday)
Emergency AI needs formal/executable grounding—LLM fluency alone fails first at tool binding
DisasterBench’s FPoF analysis finds tool mismatch and parameter-binding errors dominate first failures across 233 disaster-planning tasks over 26 tools. Nashville’s GenAI trainer improved expert agreement to 94.3–95.9% only when formal-methods specs—not pure LLM judging—anchored assessment. For SLED EMS/EM agent pilots, require schema-validated tool calls and expert-traceable rules before trusting multi-step plans.
Operating questionWhere do our EMS/EM AI workflows validate tool parameters and escalate on schema mismatch before a plan reaches a duty officer?
Supporting evidenceTexas A&M University et al.Vanderbilt University / Metro Nashville Department of Emergency Communications
Evidence ledger
Full record · every source keeps its link and limitations
GMToday: Waukesha County Ava AI has handled 91,892 calls with ~53% automated under dispatcher review
On October 2, 2026, the Waukesha Freeman (GMToday) reports the Waukesha County Communications Center went live January 21 with Aurelian’s Ava conversational AI for call handling. Operations Manager Chris Becker states Ava has handled 91,892 calls since go-live, with about 53% automated in some way (gathering information for dispatchers, transferring callers, or providing information). Dispatchers still review every Ava action and flag transcription/location/question-order issues for vendor iteration. Staffing is approved for 46 dispatchers; Becker frames Ava as helping existing staff answer 911 faster rather than replacing dispatchers, with a satisfaction target near 60% automation versus vendor estimates of 70–75% non-emergency reduction. Fail-safes escalate disabled-vehicle and other cues to live dispatchers; Becker emphasizes humans ultimately decide. Other Wisconsin agencies (La Crosse first, Dane live, Brown preparing, Ozaukee non-emergency) are adopting Ava.
Why it matters, evidence and limitations
- Why it matters
- Direct PSAP emergency-communications AI evidence with measured call volumes and automation share under human review—core emergency-services dispatch stream.
- Evidence and measured results
- Local journalism quoting the communications-center operations manager with specific call counts and automation percentage; not an independent outcome audit.
- Limitations and uncertainty
- Single-agency local reporting; metrics agency-reported; emergency-answer-time benefits not measured here.
Powerful Medical and corpuls partner to embed Queen of Hearts AI ECG in prehospital corpuls devices (EU/UK/CH start)
On October 1, 2026, Powerful Medical announced a partnership with corpuls to integrate the Queen of Hearts AI ECG model into corpuls monitor/defibrillator and telemedicine systems used by emergency professionals in over 70 countries, starting in the EU, Switzerland, and the United Kingdom. A 12-lead ECG on a corpuls3 would return STEMI and STEMI-equivalent (OMI) AI interpretation within seconds via corpuls.mission LIVE, with hospital PMcardio displaying the same finding for cath-lab readiness. The vendor states ~40% of occlusions lack classic ST-elevation; claims Queen of Hearts was trained on >2.5M angiographically labeled ECGs, doubles sensitivity vs STEMI criteria, reduces false cath-lab activations from 42% to 8%, and cites a Belgian prospective implementation cutting time-to-treatment by 48 minutes (STEMI-equivalent ~six hours sooner). FDA De Novo approval is highlighted on the site banner.
Why it matters, evidence and limitations
- Why it matters
- Prehospital EMS cardiac AI-ECG is squarely in emergency-services; U.S. SLED transferability limited until local device/regulatory paths and independent outcomes are shown.
- Evidence and measured results
- Vendor partnership press page with cited performance statistics and prior implementation claims; not an independent trial report.
- Limitations and uncertainty
- Vendor claims; EU/UK/CH start; Belgian/retrospective citations not re-audited here; FDA banner is not the same as this partnership’s geographic launch evidence.
arXiv: GenAI 9-1-1 call-taking trainer at Metro Nashville—formal methods beat pure LLM assessment (94.3–95.9% vs 87.8%)
Chen & Ma (arXiv:2602.13241v3, cs.CY, 9 Jul 2026) report a six-month GenAI-powered 9-1-1 call-taking training deployment with Metro Nashville Department of Emergency Communications (MNDEC). Context: U.S. call-takers handle >240M calls/year; many centers face >25% staffing shortages; new-hire training can require up to 720 hours of one-on-one instruction. The system generates scenarios across 57 incident types and assesses trainees against 1,651 protocol requirements translated to formal specifications, with foundation models role-playing callers via telephony. Deployment scaled from an 18-workstation pilot to 190 operational users and 1,120 training sessions; logs include 98,429 user interaction events. Evaluation on 1,244 real-world calls: formal-methods hybrid assessment achieved 94.3–95.9% agreement with expert assessments vs 87.8% for the strongest pure LLM baseline (ICL+CoT+RAG). Authors emphasize HITL co-design, verifiable assessment traces, and four lessons on delivery, rigor, resilience, and human factors.
Why it matters, evidence and limitations
- Why it matters
- Actionable U.S. PSAP workforce/AI training evidence with measured agreement metrics—supports emergency-services dispatch capacity without claiming live AI call-taking.
- Evidence and measured results
- Peer-style arXiv research paper with deployment logs and agreement evaluation against expert assessments; not a randomized clinical outcomes trial.
- Limitations and uncertainty
- Training only; single jurisdiction; agreement ≠ field clinical outcomes; arXiv HTML preprint lineage.
EMS1/Seattle Times: Seattle Fire used Corti AI on 911 medical calls without public disclosure or outcome metrics
EMS1 republishes Seattle Times reporting (Daniel Beekman, June 15, 2026) that for more than two years Denmark-based Corti’s AI has analyzed Seattle residents’ 911 medical calls without caller disclosure or public review, confirmed by the Fire Department in late May after Times inquiries. Corti prompts dispatchers and can nudge routing of some patients to a nurse-staffed Texas call center rather than immediate ambulance dispatch. Officials reportedly cannot explain success metrics or current contract details publicly; no public website disclaimers; no shared outcome metrics; use not assessed under Seattle’s surveillance ordinance. Why now for 2026-10-03: same-week Waukesha-scale PSAP automation (53% of ~92k Ava-handled calls) makes undisclosed 911 AI assist a live governance comparator for SLED agencies expanding conversational or triage AI.
Why it matters, evidence and limitations
- Why it matters
- Cautionary U.S. fire/EMS 911 AI governance case directly in emergency-services; newly relevant as PSAP automation metrics become public elsewhere.
- Evidence and measured results
- Investigative journalism with department confirmation of non-disclosure and missing public metrics; not a technical performance evaluation of Corti.
- Limitations and uncertainty
- June 2026 reporting; Seattle-specific law; clinical accuracy of Corti not measured here; why-now rests on contemporaneous PSAP automation scale-up.
arXiv DisasterBench: LLM disaster-response planners fail first on tool mismatch and parameter binding
Chen et al. (arXiv:2605.27957v1, cs.CL, 27 May 2026) introduce DisasterBench for LLM planning under typed tool-interface constraints in disaster-response multi-agent workflows. From expert-designed workflows they derive 233 planning tasks over a pool of 26 tools spanning perception to multi-stage branching (e.g., satellite analysis, nowcasting, flood modeling, damage assessment). Models receive tool specs/schemas without underlying workflow structure. First-Point-of-Failure (FPoF) localizes earliest root causes. Findings: planning-method gains depend strongly on model capacity; tool mismatch and parameter-binding errors dominate first failures (semantic grounding vs execution consistency); verbose intermediate reasoning can clash with structured-output requirements. Frontier example: Gemini 3.1 Pro Preview ToT reaches 73.39 exact-match in reported table cells, with lower scores for weaker models/methods. Why now: newly relevant as agencies consider agentic EM orchestration amid expanding PSAP/EMS AI tooling.
Why it matters, evidence and limitations
- Why it matters
- Provides measurable failure modes for agentic disaster-management AI—directly informative for emergency-management stream boundaries (not policing).
- Evidence and measured results
- Academic benchmark paper with task counts, tool pool size, FPoF taxonomy, and model/method score tables; not an operational deployment study.
- Limitations and uncertainty
- Synthetic/expert tasks; May 2026 preprint; scores are plan exact-match not lives saved; why-now interpretive relative to concurrent PSAP/EMS AI expansion.
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.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.
- Academic research
- Research produced through an academic institution or peer-reviewed venue.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.