from osmoda research · insurance
47 claims triaged, 2 fraud flags, before the coffee cools.
naga reads the photos and the graph at once. Cross-carrier ISO ClaimSearch, EXIF + deepfake forensics, a transparent rubric. Every score carries the feature reasons. Built for the NAIC AI Bulletin and Colorado Reg 10-1-1.
TL;DR
- • 47 FNOLs triaged with 2 fraud flags in 11 min for $0.42, vs the typical 3–5 day adjuster queue at legacy carriers
- • US insurance fraud costs $308.6B/yr; P&C alone is ~$45B, with staged crashes ~$20B [CAIF 2022]
- • naga scores each claim against ISO ClaimSearch (1.8B+ claims, ~95% of US P&C) and a local image-tamper detector before adjuster routing
- • Every score is explainable and auditable — built for the NAIC Model AI Bulletin (24+ states) and Colorado Reg 10-1-1
- • Photo-manipulation and deepfake claims rose ~300% YoY at one UK carrier; naga's vision check + cross-carrier graph closes the gap
1. The pain — what FNOL latency costs you
The Coalition Against Insurance Fraud's 2022 economic-impact study put US insurance fraud at $308.6 billion annually, the first refresh of that number in 27 years; property and casualty alone accounts for roughly $45 billion, with staged collisions estimated at ~$20 billion of that total. Meanwhile, traditional carriers running legacy claims platforms still report 3–5 day FNOL-to-assignment cycles where digital-native insurers measure the same step in hours.
The legacy stack is a Guidewire ClaimCenter or Duck Creek instance feeding a queue, plus a Verisk ISO ClaimSearch lookup that ~1,850 carriers contribute to and that captures roughly 95% of US P&C claims. Specialist fraud-detection vendors — Shift Technology, FRISS, and Verisk's own scoring — bolt onto that. The seams between systems are where claims wait, where staged-collision rings exploit dwell time, and where black-box scoring runs into compliance trouble.
The pressure is rising on both sides. AI-fabricated claim evidence — doctored damage photos, synthetic walkaround videos, deepfake injury statements — climbed an estimated 300% year-on-year at one UK insurer. Simultaneously, the NAIC's Model Bulletin on the Use of AI Systems by Insurers (now adopted in 24+ states), Colorado SB21-169 with Reg 10-1-1, and the EU AI Act's high-risk classification all require written AI governance, bias testing, and explainability for any model that influences pricing or claim outcomes.
Shift Technology
AI fraud-detection SaaS, Celent Luminary 2024. Strong on pattern detection — but ships as a scoring black box; explainability for state DOI exams is your problem.
FRISS
P&C-focused fraud and risk scoring. Effective on common typologies, but still a third-party scoring service that requires its own vendor risk and AI-bulletin documentation trail.
Verisk ISO ClaimSearch
The 1.8B+ claim cross-carrier database, ~95% market coverage. Indispensable for prior-loss matching, but it is data, not a workflow.
Guidewire ClaimCenter
The dominant P&C claims platform — queues, routes, pays. FNOL triage and fraud signals live in the seams between Guidewire, your fraud vendor, and your SIU.
2. The workflow — how osmoda triages an FNOL
- 1 · soot ingests the FNOL — voice transcript, web-form JSON, or telematics event — on a webhook and normalises it to a typed claim record. The watchdog (6-second median wedge recovery) makes sure no FNOL is dropped from the queue.
- 2 · naga calls the typed
verisk.claimsearch.querytool with VIN, claimant, and incident geo, pulling cross-carrier prior-loss matches; in parallel, runs a local image-forensics tool over uploaded damage photos for EXIF inconsistencies and tamper signatures. - 3 · naga scores the claim on a transparent rubric — prior-loss density, image-tamper score, claimant-network distance — and emits both a numeric score and the human-readable feature reasons, ready for a Colorado Reg 10-1-1 explainability response.
- 4 · frog maps the claim to the right adjuster bucket using carrier-defined routing logic (severity × line-of-business × jurisdiction), opens the claim in Guidewire ClaimCenter or Duck Creek over its API, and drafts the SIU referral on flagged claims.
- 5 · lantern writes the full triage event — score, features, evidence hashes, routing decision — to the SHA-256 hash-chained ledger; the bundle is the artifact you hand a state DOI examiner under the Model Audit Rule or a market conduct exam.
3. Why it works
Sovereignty + the regulators
State DOIs are now actively examining AI-driven claims decisions. The NAIC Model Bulletin on AI is in force in 24+ states, Colorado's Reg 10-1-1 already binds life carriers and is moving to auto and health, and the EU AI Act lists insurance pricing and claim handling among its high-risk uses. osmoda's post-quantum mesh and EU-1 default residency keep claimant PII inside the carrier's regulatory perimeter, on infrastructure you can show an examiner.
Explainability as artifact
Every claim score naga emits is paired with the feature contributions and the source-tool calls that produced them, all written to a hash-chained ledger that a fraud bureau, reinsurer, or state DOI can verify byte-for-byte. Atomic NixOS rollback means a model update that drifts on bias testing can be reverted in one command — an answer to the bulletin's testing-and-monitoring expectations that is procedurally clean.
Economics
A two-day adjuster queue is two days of LAE accruing, two days of fraud rings staging the next loss, and two days of policyholder NPS bleeding. osmoda triages 47 FNOLs in 11 minutes for $0.42 — a unit economic that lets carriers raise the SIU referral threshold without raising the headcount. Spirit isolation between naga (risk), frog (routing), and lantern (audit) keeps the per-claim reasoning auditable instead of a single black-box score.
FAQ
How is naga's scoring different from Shift Technology or FRISS?
Shift and FRISS are scoring vendors — you ingest a number and an opaque rationale. naga is a runtime that you own: the feature set is yours, the rubric is yours, the ledger is yours, and the same agent both runs the score and explains it. That distinction is what state DOI examiners are now asking for under the NAIC AI Bulletin.
Does this replace ISO ClaimSearch?
No — naga calls ISO ClaimSearch as a typed tool. ClaimSearch holds 1.8B+ claims and roughly 95% of the US P&C market; replicating that consortium data is not the goal. The job is to call it on every FNOL automatically, combine it with image forensics and a local network graph, and put the answer in front of an adjuster in minutes instead of days.
How do you handle deepfake or AI-manipulated claim photos?
naga runs a local image-forensics check on every uploaded photo — EXIF and codec inconsistencies, generative-model fingerprint detectors, and prior-image hash matching against the ClaimSearch image set. Suspicious frames are flagged with their feature reasons before the claim reaches an adjuster, not after the payout.
Move FNOL triage from days to minutes — and keep the audit trail your DOI wants.
Spawn a triage agent →