Question 1

How do you make an ML pricing or claims model defensible under the NAIC Model Bulletin on Use of AI/ML in Insurance?

Accepted Answer

The Bulletin (adopted by 20+ states by 2026) effectively requires an AI Systems Program covering governance, risk management, third-party oversight, and testing for unfair discrimination. We ship: (1) a written AI governance policy mapped to the carrier's ERM framework, (2) a model inventory with criticality tier, (3) pre-deployment disparate-impact testing against proxies for race, ethnicity and ZIP-level income — using the 80% rule plus calibration-by-group, (4) ongoing monitoring with KS / PSI drift triggers, (5) a third-party data attestation pack (FCRA, telematics-vendor SOC 2), and (6) a state-by-state filing memo so the carrier's actuarial team can drop the model into a SERFF rate filing without rework.

Question 2

How do you integrate with Guidewire PolicyCenter / ClaimCenter without breaking the nightly batch and policy-period processing?

Accepted Answer

We never write back synchronously to the in-suite databases. Reads come from the Guidewire Cloud Data Access (CDA) APIs, PolicyCenter and ClaimCenter Plugin Integration APIs, or a CDC stream off the operational store. Inference results land in a separate scoring service and are surfaced inside ClaimCenter or PolicyCenter via a Gosu Plugin Registry entry that calls our REST endpoint, with the response stamped onto a custom claim or submission entity. End-of-day batch (rating, billing, statutory exports) runs unmodified. Typical Guidewire Cloud overlay goes live in 8–12 weeks; on-prem InsuranceSuite is 10–16 weeks once a Gosu / Edge build environment is wired up.

Question 3

Can you automate IFRS 17 CSM and Risk Adjustment calculations and tie the outputs back into ledger?

Accepted Answer

Yes. We build the measurement-model engine (BBA / GMM, VFA for participating, PAA for short-duration) on top of the policy-admin extract, compute the Contractual Service Margin roll-forward, allocate Risk Adjustment using a confidence-level or cost-of-capital approach, and post the results into the actuarial sub-ledger (FIS Insurance Risk Suite, Moody's RiskIntegrity IFRS 17, SAS IFRS 17, or a bespoke Spark / Delta build). Disclosures generate as XBRL-tagged tables for the annual report. ML enters in two narrow places: lapse / mortality experience studies and cashflow projection compression — both with full reproducibility and a parallel deterministic run for audit.

Question 4

What is the dispute rate on image-based auto damage assessment, and how do you bound it?

Accepted Answer

On clean photo intake (well-lit, multi-angle, VIN visible), our auto-glass and bumper / panel models settle without adjuster touch on roughly 55–70% of claims at a sub-5% policyholder dispute rate. We bound dispute risk with: (1) a confidence threshold below which the claim auto-routes to a human appraiser, (2) a parts-and-labour estimate generated against Mitchell, CCC or Audatex pricing — not pulled from the model directly, (3) an explicit fall-back to in-person inspection for total-loss-likely vehicles (model predicts ACV-to-repair ratio), and (4) a state-by-state opt-out for jurisdictions with virtual-inspection restrictions.

Question 5

Do you replace RMS / AIR / KCC for CAT modelling, or overlay them?

Accepted Answer

Overlay, not replace. The vendor CAT models remain the system of record for treaty and rating-agency reporting — they encode peril physics and vulnerability curves the regulator and reinsurer already accept. We add an ML layer that: (1) re-weights the event catalogue using recent climate signals where the vendor model has a 3–5 year lag, (2) recalibrates vulnerability curves to the carrier's actual loss experience by construction type and ZIP, (3) flags exposure-aggregation hotspots the vendor model under-weights (e.g. inland flood adjacent to flood zones, wildfire WUI shoulders), and (4) produces an "ML-adjusted PML" alongside the headline vendor PML so the reinsurance buyer sees both. The overlay typically lands in 4–6 weeks.

Question 6

How do you keep a claims fraud / SIU referral model out of fair-insurance and state UDAP trouble?

Accepted Answer

SIU referral models in insurance are not "lending" but they are subject to state unfair-trade-practices acts, the Civil Rights Act, and increasingly state insurance bulletins on algorithmic discrimination. We treat them like a regulated model: no protected-class or proxy inputs (ZIP-only inputs are audited for race correlation), disparate-impact testing on referral rates, calibration parity (the model must be right about fraud at the same rate across groups), and a human-in-the-loop requirement — the model never denies a claim, it only routes for SIU review, with the adjuster recording the final disposition. We also keep an explainability artefact per referral (top features, comparable-claim distance) so the carrier can answer a DOI complaint.

Question 7

How do you handle telematics data privacy and consent for UBI (usage-based insurance) programs?

Accepted Answer

Three guard-rails. (1) Consent: an explicit, granular consent flow — separate opt-ins for trip collection, scoring, marketing use and data resale, with a one-tap revoke that wipes the trip store within 30 days. (2) Minor / COPPA-adjacent edge: if the insured vehicle is driven by a household member under 16 (teen-driver programs), we treat the driving data as parental-consent data and exclude marketing reuse entirely. (3) Vendor flow-down: when we ingest from Cambridge Mobile Telematics, OBD-II dongle vendors (Vector, Octo) or OEM connected-car APIs (GM, Ford, Stellantis), the data-processing agreement flows GLBA, state mini-GLBA and California CPRA terms to the vendor. The scoring model itself never sees the raw PII — it consumes a hashed driver-trip key.

Question 8

What supporting documents do you produce for a state DOI rate filing of an ML-driven rating plan?

Accepted Answer

A SERFF-ready filing pack: (1) a rate-filing memorandum describing the model in actuarial language (target variable, training window, segmentation, exposure base, credibility weighting), (2) variable-by-variable justification mapped to the NAIC Casualty Actuarial and Statistical Task Force's "Regulatory Review of Predictive Models" white paper, (3) disparate-impact and proxy-discrimination test results, (4) a side-by-side rate comparison versus the incumbent plan at the policyholder level showing dislocation by decile, (5) a monotonicity-constraints exhibit (the model is constrained so that, e.g., more prior at-fault accidents never decreases premium), and (6) the model in an inspectable form — typically a monotonic-constrained GBM exported to a scorecard-equivalent table. We have built this pack for California, New York, Florida, Texas and Washington filings; the variable-justification depth required by California CDI is the binding constraint and we design to it by default.

Carrier-grade AI: ML pricing the DOI will sign off, claims that close without an adjuster

What we build to

Insurance AI Systems

FNOL Voice + Text Automation

Underwriting Triage + ML Pricing

Image-Based Damage Assessment

SIU Fraud Referral

ACORD + Dec Page Document Intelligence

Subrogation + Litigation Propensity

Reserve Adequacy (Case + IBNR)

Customer-Service IVR + Agent Assist

Insurance Data Engineering

Policy-Admin Mirror + Event Stream

Claims Data Lake (FNOL → Close)

Third-Party Risk Data

Regulatory Reporting Pipelines

Telematics + UBI Pipeline

CAT + Exposure Aggregation

Insurance AI FAQ