From the Emergency Services edition of November 5, 2026
KOSDI study: Seoul Emergency Operations Center AI 119 callbot cut average wait ~7.33% with >70% early urgency accuracy
Seoul Emergency Operations Center / Chung-Ang University · Fire / 119 emergency call-taking · Seoul, Republic of Korea
- Publisher
- Journal of the Society of Disaster Information (KOSDI)
- Original publication
- 2026-03-31
- Source retrieved
- 2026-11-06
What happened
Mun and Park analyze Seoul's AI-supported 119 call-taking: STT/NLP/urgency models assist human call-takers rather than replace them. Average wait fell about 7.33% (19.79s in 2024 to 18.34s in 2025). In an Aug. 13, 2025 110-minute rain surge, the callbot handled 2,150 of 2,898 calls; of 412 judged calls, urgency accuracy was 72.2% (emergency) and 73.4% (non-emergency), with 81% unjudged due to abandon or immediate human connect. Authors urge better training data, explainable AI and governance. Why now: the Nov. 5, 2026 Digital Times council audit of the same Seoul 119 AI program makes these ops metrics newly relevant for agencies deciding go-live gates.
Why it matters
Provides rare measured PSAP wait-time and surge-accuracy figures for fire/EMS call AI that U.S. centers can use as a benchmark—while today's Seoul audit shows why those figures must sit under independent governance.
Evidence and measured results
Open-access PDF DOI 10.15683/kosdi.2026.3.31.389; Tables 3–6 wait and surge accuracy; architecture described as assistive to existing exchange/CAD.
Limitations and uncertainty
Single-city Korean 119 context; early-model ~70% urgency accuracy; large unjudged share in surge; not a randomized outcome trial; older than freshness window except for why-now linkage to Nov. 5 audit.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-11-06; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Large 911/119 centers need to cut non-emergency load and wait time without replacing human call-takers on true emergencies.
- Stakeholders
- PSAP director
- Fire/EMS medical director
- Call-taker union or workforce leads
- City IT/security
- Elected public-safety oversight
- Discovery questions
- What share of calls are non-dispatch, and what is current average speed of answer?
- Will AI only engage above a queue threshold?
- What urgency ontology maps to local EMD?
- What independent evaluation will mirror Seoul's later audit questions?
- Value hypothesis
- An assistive callbot that screens surplus queue traffic could modestly reduce wait time if local accuracy and governance match or beat Seoul's reported pilot—value is operational, not clinical outcome proven here.
- Potential engagement
- Pilot design and measurement plan for a regional PSAP comparing pre/post wait time and AI-vs-human priority agreement on a surge sample.
- Unsupported claims
- Guaranteed 7.33% wait reduction in any U.S. center
- Clinical outcome improvement or lives saved
- That >70% urgency accuracy is adequate for unsupervised answer
Pre-sales engineering
Role takeaway
- Fit
- Fits assistive STT/NLP urgency hinting integrated to existing call-taking and CAD screens; not a standalone AI call-answerer.
- Architecture and integration
- Exchange → AI analysis server (STT/NLP/urgency) → call-taker UI + CAD notes; human issues dispatch. Separate training and production data stores.
- Prerequisites
- Labeled local call audio with priority ground truth
- Low-latency STT for local accents and place names
- CAD/API hooks for summary fields
- Constraints
- Seoul metrics from one center and one surge window
- Callbot only entered when ≥4 waiting calls
- Korean language and 119 multi-hazard mix
- Security
- Paper cites de-identification and access control; U.S. deployments need CJIS/HIPAA-aligned handling of medical 911 content.
- Proof of value
- Replicate wait-time A/B with queue-threshold gating
- Agreement rate AI vs certified EMD on held-out calls
- Surge drill measuring judged vs unjudged (abandon/transfer) share
Delivery
Role takeaway
- Work
- Implement gated assistive pilot, instrument metrics from Seoul tables, and couple go-live to an external or council-visible evaluation.
- Dependencies
- Telephony vendor
- CAD vendor
- QA/training unit
- Privacy counsel
- Ownership
- PSAP operations owns acceptance; medical director owns clinical protocol mapping; IT owns hosting.
- Skills and adoption
- Call-takers practice using AI summaries without over-trust; QA scores AI-influenced calls separately.
- Governance checkpoints
- Pre-pilot protocol approval
- Mid-pilot accuracy review
- Pre-expansion independent evaluation
- Acceptance criteria
- Documented wait-time change with confidence bounds
- Urgency agreement threshold met on sample
- Human override and transfer paths tested under surge
- Risks
- Over-trust of ~70% early accuracy
- Mis-localization of street names
- Political conflict if ops metrics diverge from later audits
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Prefer assistive AI analysis servers feeding call-taker UI/CAD with queue-threshold callbot engagement over unsupervised answer.
Governance
Who approves, reviews and stays accountable for outcomes?
Even positive pilot metrics need institutional frameworks and explainability before scale—consistent with later Seoul oversight conflict.
Security and privacy
What data, permissions and controls need testing?
Authors note de-identification and access control plus separation of training and ops data as design requirements.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Workload relief for call-takers is claimed; equity across accents/disabilities is not measured—must be tested before U.S. adoption.
Procurement
What should contracts, pricing and exit terms secure?
Buyers should contract for metric transparency (wait, accuracy, unjudged share) matching these tables, plus independent evaluation rights.
Operating model
Which teams own the service once it runs?
Keep humans as final decision-makers; use AI for surge pre-screen and documentation assist.
Publication history
- 2026-11-05Emergency Services · Issue 604 resources
Stable resource ID: kosdi-seoul-ai-119-callbot-wait-733-accuracy-20260331