Public Safety & Emergency ServicesIssue 285 records
Emergency Services
Tonight's edition follows non-emergency voice agents spreading through Wisconsin PSAPs and asks what evidence backs them. Dane County reports 24,000+ AVA calls in a month with staff monitoring for misses; Ozaukee County launched Ava as 911 volume rose 17%. A new preprint shows LLMs anchoring on the first complaint and under-triaging urgent cases when judging conversations as they unfold, a warning for any mid-call AI. Two vendor releases, 911Translate at 100+ agencies and HEARTS 2.0 LLM documentation in Japan, show fast adoption with thin accuracy data.
- Evidence records
- 5
- Cross-source patterns
- 3
- Evidence classes
- 2 independent reporting2 vendor claim1 academic research
- Outcomes
- 4 emerging1 cautionary
- Source freshness
- 4 new this fortnight1 recent
- Research completed
- 2026-10-05
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Wisconsin PSAPs keep adding non-emergency voice agents, with human monitoring and 911 transfer as the stated safeguard
Dane County reports 24,000+ AVA calls in its first month with staff rotating to monitor and call back suspected misses; Ozaukee County launched Ava citing 113,538 non-emergency calls and a 17% rise in 911 volume. Both rely on emergency detection and transfer to humans, and neither has published missed-escalation counts. Interpretation: adoption is outrunning public measurement of the one failure that matters.
Operating questionBefore we add a non-emergency voice agent, how will we count and publish missed or delayed emergency transfers, and who staffs the monitoring seat?
Supporting evidenceDane County Public Safety Communications (Wisconsin) / Spectrum News 1Ozaukee County Sheriff's Office Communications Center (Wisconsin) / WRJN
Live-call AI should be judged on how it handles conversations as they unfold, not on completed records
The sequential triage preprint shows models anchoring on the first complaint and under-triaging the most urgent cases (best QWK 0.295 vs clinicians 0.887–0.929). Dane County's AVA depends on spotting emergency cues mid-call, and its leaders say misses came from programming. Interpretation: agencies should test emergency detection and acuity prompts at multiple points in a call before trusting them.
Operating questionHave we tested our call-handling AI on partial calls where the critical fact arrives late, and what is its high-acuity miss rate?
Supporting evidenceUniversity-affiliated researchers (arXiv preprint 2609.22904)Dane County Public Safety Communications (Wisconsin) / Spectrum News 1
AI translation and EMS documentation tools are arriving with adoption counts but little accuracy evidence
911Translate claims 100+ agencies in six months with no accuracy data; HEARTS 2.0 launches LLM voice summarization across three Japanese fire departments with only an OCR accuracy figure. HEARTS does publish usage limits that keep personal data out of the LLM. Interpretation: agencies should require written usage rules and their own accuracy tests before relying on either kind of tool.
Operating questionWhat accuracy test and usage rules must a translation or documentation AI pass before dispatchers or crews rely on it?
Supporting evidenceDeskOfficer, Inc. (911Translate)Himeji City Fire Bureau, Nishi-Harima Fire Union, Ako City Fire Department / TXP Medical
Evidence ledger
Full record · every source keeps its link and limitations
Spectrum News: Dane County AVA handled 24,000+ non-emergency calls in first month with no major mistakes reported
Spectrum News 1 reports (Sept 28, 2026) that Dane County's AVA voice assistant handled more than 24,000 non-emergency calls in roughly its first month. The deputy director said there were no major mistakes and attributed misses to how the county programmed the system. Staff rotate monitoring AVA so a communicator can call back if an emergency is missed; callers can ask for a human. Eleven months of the trial remain.
Why it matters, evidence and limitations
- Why it matters
- Direct 911-center evidence that conversational AI on non-emergency lines is being run as a staffed, monitored trial, with emergency detection on the non-emergency line called out as a key function.
- Evidence and measured results
- Local TV reporting quoting a communicator and the deputy director; the 24,000-call count and 'no major mistakes' are agency statements with no published error log, missed-emergency count, or 911 answer-time baseline.
- Limitations and uncertainty
- One month of operation; metrics are agency-reported; no independent evaluation; 'mistakes' defined informally by the agency.
WRJN: Ozaukee County consolidated center deploys Aurelian Ava on non-emergency line as 911 volume rises 17%
WRJN reports (Sept 30, 2026) that Ozaukee County's consolidated communications center, serving seven police departments, seven fire departments and the Sheriff's Office, put Aurelian's Ava on its non-emergency line. The center processed 113,538 non-emergency and 17,612 emergency 911 calls in 2025, and emergency volume is up 17% this year. Ava routes callers by stated reason, transfers detected emergencies to 911, and supports Spanish; the sheriff said there will be no job cuts.
Why it matters, evidence and limitations
- Why it matters
- Adds another Wisconsin consolidated PSAP, serving fire and EMS as well as police, to the regional Ava adoption cluster, with a clear volume rationale.
- Evidence and measured results
- Short local news item restating agency claims; no performance data, error rates, or post-launch metrics. 'Rigorous training' is an unspecified vendor/agency claim.
- Limitations and uncertainty
- Launch announcement only; no outcome data; volumes are 2025 figures and the 17% increase is not broken out by source.
arXiv: LLMs anchor on chief complaint in sequential triage; best model QWK 0.295 vs clinicians 0.887–0.929
A Sept 19, 2026 arXiv preprint evaluates six LLMs on predicting ESI triage acuity from growing prefixes of nurse-patient conversations (425 simulated, 50 clinician-authored). Every model drops from moderate-to-substantial agreement on completed records to fair-to-moderate agreement during the conversation. Labels anchor on the chief-complaint exchange; prompting fails to fix it; every model under-triages ESI-1 cases. Clinicians reach QWK 0.887–0.929 versus 0.295 for the best model, and ensembling worsens the failure.
Why it matters, evidence and limitations
- Why it matters
- Emergency-department triage, not 911, but directly relevant to live-call AI in dispatch and nurse-line diversion, where decisions are made on partial, unfolding conversations.
- Evidence and measured results
- Controlled evaluation with perturbation tests and clinician comparison; uses simulated and clinician-authored conversations rather than real triage audio; preprint not peer reviewed.
- Limitations and uncertainty
- ED triage with ESI labels; synthetic conversations; six specific models; transfer to 911 protocols such as MPDS is untested; preprint.
TXP Medical: Himeji and two Hyogo fire departments launch HEARTS 2.0 with LLM voice summarization of EMS records
TXP Medical's Oct 1, 2026 release says three Hyogo fire departments (about 790,000 residents, about 48,800 ambulance runs a year) began full operation of HEARTS 2.0 in October. Crews speak summaries and an LLM maps them into record fields, described as a national first. Operating rules bar LLM use for personal information and limit it to urgent cases where typing is hard. Monitor OCR reached 92.3% fully correct across six items; offline mode is included.
Why it matters, evidence and limitations
- Why it matters
- International EMS example of generative AI for field documentation and hospital-notification speed, with explicit usage limits; transferable as a design pattern, not as evidence of outcomes.
- Evidence and measured results
- Vendor press release; the only accuracy figure is OCR (92.3%); no LLM summary accuracy, time savings, or transport-time outcomes are reported. Expected effects are stated as goals.
- Limitations and uncertainty
- Vendor claims; Japan's EMS and hospital-notification system differs from U.S. structures; outcomes not yet measured.
DeskOfficer: 911Translate AI language line says it is live at 100+ agencies within six months, without published accuracy data
DeskOfficer announced Oct 1, 2026 that 911Translate, an AI interpreter line dispatchers conference into calls, is live at more than 100 agencies under six months after launch. It claims 90+ languages, language detection within a second in most cases, two-way real-time translation, per-line confidence scores cross-checked against other tools, and drop-in replacement of existing interpreter lines. It is free to start. No accuracy, error, or call-outcome data is given.
Why it matters, evidence and limitations
- Why it matters
- Fast adoption of AI translation in 911 is a material change in language access on emergency calls; the evidence is adoption count only.
- Evidence and measured results
- Vendor press release; adoption count is self-reported; no independent accuracy testing, no named agencies, no error rates by language.
- Limitations and uncertainty
- Vendor claims only; 'free to get started' may shape adoption speed independent of quality; no outcome data.
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.
- 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.