Insurance Claims AI Governance Case Study

The situation

Context and stakes

Cedarline Insurance is a regional property & casualty carrier writing personal auto, homeowners, and small-commercial policies across seven states, handling roughly 380,000 claims per year. Under pressure to shorten cycle time after a run of severe-weather seasons, three AI capabilities had moved from pilot into production readiness — each owned by a different team, each subject to a different regulator's attention:

FirstLook: an NLP-driven triage model scoring inbound first-notice-of-loss (FNOL) reports on complexity, severity, and expected reserve, and routing them to a service track.
PanelView: a computer-vision system estimating auto-collision repair cost from claimant-submitted photos and pre-authorizing payouts below a set threshold.
Sentry: a suspicious-claim detector that scores every claim for potential fraud and refers high-scoring cases to the Special Investigations Unit (SIU).

A wrongly denied claim can trigger a market-conduct exam and a bad-faith lawsuit. A biased fraud score can push protected communities into deeper scrutiny for the same loss. And an over-generous computer-vision estimate can leak millions across a storm season. The board wanted one program that held all three systems to the same evidentiary standard while respecting how differently they fail.

“A denial letter is a legal document. If a model helped write it, we need to be able to defend every word — to the policyholder, to the department of insurance, and to a jury.” — General Counsel, Cedarline Insurance

Risk map

What could go wrong?

The Enterprise Risk group catalogued the failure modes that carry the greatest legal, financial, and reputational weight in claims handling:

Unfair claims practices

Model-driven denials or delays that violate state Unfair Claims Settlement Practices Acts.

Disparate treatment

Fraud or triage scores producing worse outcomes for policyholders in protected classes or specific ZIP codes.

Estimation leakage

Vision model systematically over-estimating repair cost, either overall or for certain vehicle segments.

Adversarial photo submission

Claimants gaming the vision model with staged or AI-generated damage images.

Silent denial

Triage or fraud model effectively closing a claim without a documented human decision.

Regulator surprise

A state DOI discovering an AI-driven decision path Cedarline had not disclosed in its rate or form filings.

Implementation

Step-by-step governance rollout

Charter a Claims AI Governance Council

Cedarline stood up a cross-functional Claims AI Governance Council rather than lodging AI in a single function. Voting members included the Chief Claims Officer, Chief Data Officer, General Counsel, Head of SIU, Head of Compliance, a customer-advocacy officer, and — a deliberate choice — a rotating frontline adjuster.

Key decisions

Any system whose output changes the treatment of a claim is in scope, even if a human clicks the final button.
Adopt the NAIC Model Bulletin on AI as the baseline; add jurisdiction-specific overlays for Colorado, New York, and California.
Three tiers: Tier 1 advisory routing, Tier 2 monetary pre-authorization, Tier 3 fraud or coverage-affecting scoring.

Deliverable: Council charter, tier rubric, and an intake process every business unit must use before an AI system touches a live claim.

Inventory every model touching a claim

A 60-day discovery sweep across Claims, SIU, and Digital surfaced 31 candidate "AI" systems. Six turned out to be static rule engines misdescribed as AI, four were unauthorized vendor pilots running on live claim data, and two were browser extensions summarizing claimant emails through a consumer LLM. Everything unauthorized was frozen until reviewed.

The refreshed inventory captured, for each surviving system: business purpose, tier, training-data provenance, decision boundary, human override point, retention policy for claimant PII, and the two named owners (business and technical).

Assign paired ownership and independent validation

Every model got a business owner accountable for outcomes and an independent validator who could not report to the model's developer:

FirstLook: owned by the VP of Claims Operations; validated by Enterprise Analytics.
PanelView: owned by the Auto Claims Director; validated by an external actuarial consultancy given the vendor's proprietary architecture.
Sentry: owned by the Head of SIU; validated by a Model Risk analyst outside SIU to prevent conflicts of interest.

Compliance signed off on customer-facing language for each system, and the Council approved production release for Tier 2 and Tier 3 systems.

Make every decision reviewable — in the claim file

Explainability was treated as a claim-handling requirement, not a data-science report.

FirstLook writes a structured note into the claim file: complexity score, top three drivers, chosen track, and any override reason. Adjusters can see and rebut the drivers before the routing sticks.
PanelView attaches the annotated images, per-panel repair-or-replace calls, parts and labor breakdown, and confidence intervals. Any estimate near the pre-authorization threshold is auto-referred to an appraiser.
Sentry emits a ranked list of red-flag indicators. SIU investigators cannot open a formal investigation on model score alone — the note must cite at least one indicator that a trained investigator would recognize as material.

Test disparate impact across claimants and geographies

Compliance and Data Science jointly ran fairness reviews on the two consequential models:

Sentry referral rates and downstream denial rates compared across age, sex, race/ethnicity proxies (BISG), preferred language, and ZIP code.
PanelView estimate accuracy compared across vehicle age, trim segment, and dealer-density regions to catch under-estimation in rural areas.
FirstLook complexity assignments compared across policyholder tenure to make sure new customers were not being routed to slower tracks.
Challenger versions retrained without the highest-risk features to test whether accuracy loss was material.

Two Sentry features were retired because their marginal lift was small relative to the disparate-impact signal they carried. PanelView added a rural-region calibration layer after a systematic 6% under-estimate showed up outside major metros.

Perimeter claimant data and defend against adversarial inputs

Claims data is uniquely sensitive — medical records, photos of homes, driver information — and PanelView's photo intake is a soft target:

All model traffic routes through a Cedarline-hosted gateway that strips PII (names, VINs, exact GPS) from any payload sent to third-party model providers.
The vendor contract for PanelView forbids training on Cedarline claim images and grants Cedarline the right to a third-party audit of that clause.
Every uploaded image is scored for signs of manipulation (EXIF anomalies, generative-image fingerprints, duplication against a hash database of prior claims). High-risk images are flagged for human appraisal rather than auto-processed.
A 7-year retention policy on model prompts and outputs, matching Cedarline's underlying claim-file retention.

Define human override and escalation for each tier

Governance must work at the moment a decision is made, not in a post-hoc report:

FirstLook (Tier 1): an adjuster can re-route any claim in one click and must document the reason. Overrides are trended weekly to catch systematic model errors.
PanelView (Tier 2): auto-pay is capped at a dollar threshold and blocked for total-loss candidates, salvage-title vehicles, and any claim with injury coding. Above the cap, an appraiser must sign.
Sentry (Tier 3): a model score alone cannot delay a payment. Any hold requires an SIU investigator's documented reason, and the policyholder receives a written notice within the state-mandated window.

Stand up continuous monitoring and a rollback path

Production monitoring runs alongside the models, not on a monthly report:

Performance: triage precision for FirstLook; estimate-versus-final-cost variance for PanelView; SIU confirmation rate and false-positive rate for Sentry.
Fairness: disparities in referral, delay, and denial rates across protected classes recomputed weekly, with alerts when adverse-impact ratios cross defined thresholds.
Drift: population stability indices on inputs; a spike in adversarial-image flags triggers an immediate review of PanelView vendor updates.
Rollback: every Tier 2 and Tier 3 model can revert to the prior version or a rules-based fallback within 30 minutes by the business owner or the on-call risk officer.

Rewire the claimant experience around AI

Governance shapes what the policyholder sees and can act on:

Cedarline's claim portal discloses when AI is used to estimate damage or route a claim, in plain language, alongside a phone number for a human adjuster.
PanelView estimates come with a one-click "request an in-person appraisal" option that cannot be denied without a supervisor's approval.
Any claim delayed or denied with an AI-assisted decision includes the specific human decision-maker's name, the reasons in plain language, and the state-appropriate appeal path.
The complaint-handling team codes AI-related complaints separately so trends can reach the Council within one reporting cycle.

Report to the board — and file with the regulator

The Council reports quarterly to the board's Risk Committee and maintains a filing posture matched to each state's expectations under the NAIC AI Model Bulletin:

Inventory by tier, with new, retired, and materially changed systems highlighted.
Fairness metrics and any threshold breaches with remediation status.
Consumer-complaint counts attributed to an AI-assisted decision.
Third-party validation findings and vendor audit results, including for GenAI subcomponents.
An annual attestation from the Chief Claims Officer that no in-production model has bypassed the Council's process.

Results

Outcomes after 12 months

FirstLook reduced average FNOL-to-first-touch by 2.1 days at parity with prior triage accuracy, with adjuster overrides trending down each month as feedback loops caught early misclassification patterns.
PanelView processed 34% of eligible auto claims through pre-authorized payout, cutting cycle time on those claims by more than half, with estimate variance held within 4% of final settled cost after the rural calibration fix.
Sentry raised SIU confirmation rate on referred claims by 12 points and reduced disparate referral rates across BISG-derived groups by a factor of three after two proxy features were retired.
Market conduct posture improved: a scheduled DOI inquiry closed with no adverse findings, and examiners specifically cited the tiered rubric, explainability notes in claim files, and the human-review requirement on Sentry referrals.

“The point of the governance program was never to slow the models down. It was to make sure that when a policyholder called, we could show them, in their own claim file, exactly why we did what we did.” — Chief Claims Officer, Cedarline Insurance

Takeaways

Lessons for other insurers

Treat the claim file as the artifact of record. If a model influenced a decision, its reasoning belongs in the claim file — not in a data-science notebook.
Never let a model create a silent denial. Delays and denials require a named human decision-maker and a plain-language reason regardless of what the model recommended.
Test the vision pipeline against adversarial inputs. Photo-based estimation is uniquely exposed to staged and AI-generated imagery; screening belongs before, not after, auto-pay.
Rural and low-density regions are a known blind spot. Under-estimation and slow routing tend to cluster where training data is thin; measure by geography, not just by protected class.
Bring the frontline into the room. A rotating adjuster on the governance council catches practical failure modes that neither compliance nor data science will see on their own.
Match the regulator's vocabulary. Align documentation to the NAIC AI Model Bulletin and each state's unfair-claims-practices language, so a filing or exam response is a copy-paste, not a translation exercise.