Public Safety & Emergency ServicesIssue 234 records
Emergency Services
Today's edition shifts from yesterday's wildfire cameras to 911-center assistive AI: quality-assurance scaling versus trust and data-rights constraints. A Prepared/Axon case study says St. John's County 911 moved from ~2% manual QA sampling to automated 100% coverage with 90% less manual review time, while Long Beach's city page describes a CommsCoach pilot to score 100% of calls for supervisor coaching. Against that operational push, npj Digital Medicine's DyLM-OHCA reports strong retrospective OHCA call recognition (AUROC 0.937, median 19.7 s) but low PPV and no prospective HITL outcomes, and Footnote4a documents Prepared/Axon terms asserting vendor ownership of 911 Caller Data with broad retention and unilateral modification rights.
- Evidence records
- 4
- Cross-source patterns
- 2
- Evidence classes
- 2 independent reporting1 vendor claim1 academic research
- Outcomes
- 2 mixed1 emerging1 cautionary
- Source freshness
- 2 recent2 older, newly relevant
- Research completed
- 2026-09-29
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
911 AI quality assurance is scaling toward 100% call scoring with human coaching gates
St. John's County's Prepared/Axon case study reports moving from ~2% manual QA sampling to automated 100% coverage with 90% less manual review time and human review of fails, while Long Beach's city page describes a CommsCoach pilot to score 100% of calls for supervisor coaching under an explicit no-replacement pledge. SLTT buyers should treat full-coverage QA AI as a workforce tool that still needs local agreement metrics and labor/eval policy—not as proof of safer outcomes.
Operating questionIf we turn on 100% AI call scoring, who reviews fails, how do scores enter evaluations, and what local agreement KPIs replace vendor case-study claims?
Supporting evidencePrepared (Axon 911 Intelligence) featuring St. John's County 911, FloridaCity of Long Beach Department of Disaster Preparedness and Emergency Communications (republishing ABC7)
Assistive 911 AI still needs trust, transferability, and hard data-rights controls
DyLM-OHCA shows strong retrospective OHCA discrimination (AUROC 0.937) yet low PPV, Korean-only validation, and an explicit decision-support framing after prior RCT trust failures, while Footnote4a documents Prepared/Axon terms asserting vendor ownership of Caller Data, broad improvement licenses, 10-year retention language, and unilateral term changes. Operational enthusiasm for assistive 911 AI should be gated by prospective HITL evidence and negotiated DPAs—not AUROC or feature decks alone.
Operating questionBefore live assistive alerts or cloud QA/transcription, what prospective HITL evidence and locked data-ownership/retention terms must be in place?
Supporting evidenceKorea University / Chung-Ang University / Incheon Fire Headquarters / Catholic University of Korea (npj Digital Medicine)Footnote4a (H.C. van Pelt)
Evidence ledger
Full record · every source keeps its link and limitations
Prepared/Axon case study: St. John's County 911 AI QA from ~2% sampling to 100% coverage
Prepared's August 31, 2026 case study reports St. John's County 911 (St. Augustine, FL) moved from manually reviewing about 2% of ~101,000 annual 911 calls (20–30 minutes per review) to Automated QA + Training from Axon 911 Intelligence by Prepared. Quoted floor manager Debra Smith: the center went from QAing very few calls to doing 100%. In the past 90 days the center QA'd 4,928 calls including 100% of failing calls (score <90%); manual review time is down 90%, with a weekend example of ~120 failing calls in ~3 hours versus ~30 hours manually. Auto and manual scores are said to align within 3%. Managers keep AQA up during shifts to catch issues in progress; coaching happens within days. Humans still review failures and own coaching.
Why it matters, evidence and limitations
- Why it matters
- Direct PSAP workforce-AI signal with quantified QA coverage and review-time claims for SLTT emergency communications. Why now: not in emergency-services archive; published 2026-08-31 and newly relevant as 911 AI QA spreads beside non-emergency answering coverage after 2026-09-27's wildfire/triage edition.
- Evidence and measured results
- Opened Prepared case-study HTML via curl. Date August 31, 2026; 2%→100% QA framing; 101,000 911 calls; 4,928 calls/90 days; 90% manual-time reduction; within-3% score alignment; HITL review of <90% scores taken from page.
- Limitations and uncertainty
- Vendor marketing case study with single-site self-reported metrics; no independent audit of 90% time savings or 3% score agreement; no outcome data on call quality or patient/public safety.
Long Beach city page: CommsCoach AI to score 100% of 911 calls for supervisor coaching
A February 23, 2026 City of Long Beach news page (republishing ABC7, Feb 22) states LB9-1-1 will use GovWorx CommsCoach AI to review emergency calls and score dispatcher handling. The city reports ~600,000 911 calls/year; national standards target ≥2% QA (~12,000 calls) previously done manually. Director Reginald Harrison says the AI tool enables review of 100% of calls, with supervisors reviewing flagged calls for coaching. Contract cited at ~$68,000/year through February 2028; vendor claims ~250 other 911 centers use the tech. Harrison states AI will not replace dispatchers, take calls, or replace human judgment, and that AI listening stays within the city's secure data system after privacy/HIPAA vetting. Program described as in pilot phase.
Why it matters, evidence and limitations
- Why it matters
- Municipal primary URL (not previously archived; Hoodline secondary was) documenting 911 AI QA procurement, HITL coaching, and privacy claims. Why now: pairs with St. John's Aug 2026 AQA metrics as a second independent jurisdiction on full-coverage QA AI after recent non-emergency answering editions.
- Evidence and measured results
- Opened longbeach.gov HTML via curl. Release date 2026-02-23; 600,000 calls; 2%/12,000 standard; 100% review claim; CommsCoach/GovWorx; $68,000/yr to Feb 2028; HITL coaching and on-premises-data claim from page.
- Limitations and uncertainty
- City/press announcement of a pilot—not measured post-deployment outcomes; 100% review claim is capability framing; Hoodline/privacy package already partially covered related facts.
npj Digital Medicine: DyLM-OHCA language model detects OHCA in 60s of emergency calls (AUROC 0.937)
Open-access npj Digital Medicine article published 3 March 2026 develops DyLM-OHCA, a GPT-2-based dynamic language model for early out-of-hospital cardiac arrest recognition from 158,973 Korean 119 emergency call transcripts (2023; Seoul/Incheon/Gwangju). Within 60 seconds, DyLM-OHCA outperformed conventional ML (AUROC 0.937; AUPRC 0.456). At recall ~0.80, precision was 0.155, specificity 0.924 (FP rate 7.6%). Median recognition time was 19.7 s. Authors emphasize conversational context over keyword spotting; 52.6% of false positives later dropped below threshold. Limitations: retrospective single-country Korean-language data; no prospective dispatcher-in-the-loop outcomes; prior RCT (Blomberg) showed ML alerts did not improve recognition when dispatchers did not trust them. Framed as decision support, not automated decision-making.
Why it matters, evidence and limitations
- Why it matters
- Measured EMS/dispatch language-model evidence for life-safety call recognition with explicit HITL and transferability limits. Why now: not archived; newly relevant as U.S. PSAPs scale assistive 911 AI (QA/answering) and need outcome-grade call-AI evidence beyond vendor QA case studies after 2026-09-27.
- Evidence and measured results
- Opened nature.com HTML via curl. Publication 03 March 2026; n=158,973; AUROC 0.937; AUPRC 0.456; median 19.7 s; precision 0.155 at recall 0.80; FP 7.6%; Korean 119 setting and decision-support framing from article text.
- Limitations and uncertainty
- Korean-language retrospective internal validation; PPV 15.5% implies substantial false alarms; no measured dispatcher adoption or survival benefit in this study.
Footnote4a: Axon/Prepared 911 terms claim ownership of caller data and unilateral modification rights
July 29, 2026 Footnote4a investigation argues FlockOS 911 order forms still bind agencies to Prepared/Axon terms hosted on prepared911.com after Axon acquired Prepared. Analyzing terms versions through v8 (authored Jan 2026, published online May 14, 2026 per Webflow asset IDs), the piece highlights Axon's claim that all right, title and interest in Caller Data remain with Axon and its licensors; agencies get a limited license back. Client Data (including telecommunicator–responder audio/transcriptions in later versions) is licensed perpetually for service improvement and other Axon products. Privacy policy (still Invictus-branded/May 2024 text as of the article) states sessions may be used for business needs and retaining personal information for 10 years under a HIPAA framing, contrasting Apple's shorter emergency-data deletion expectations. Terms allow unilateral website changes applying to prior orders.
Why it matters, evidence and limitations
- Why it matters
- Independent contract/privacy scrutiny of the same Prepared/Axon 911 assistive stack featured in current PSAP QA deployments. Why now: not archived; newly relevant beside St. John's Prepared AQA case study when SLTT buyers expand Axon 911 Intelligence after 2026-09-27.
- Evidence and measured results
- Opened footnote4a.org HTML via curl. Dated Jul 29, 2026; Caller Data ownership quote; perpetual Client Data license; 10-year retention quote; unilateral modification clause; version timeline including May 2026 v8 from page.
- Limitations and uncertainty
- Legal/investigative interpretation of public terms—not a court finding; enforceability and which version binds a given agency may vary; does not measure product accuracy.
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.