Trusted AI in property and casualty insurance: how carriers make claims and underwriting AI defensible

Last updated: October 2026

Trusted AI in property and casualty (P&C) insurance is claims and underwriting AI whose outputs can be relied on when a premium is set, a reserve is booked, or a payment is released. It requires continuous evaluation against the live book, confidence-based routing to qualified experts, traceable decisions, and corrections that feed back into the model.

The question for carriers is no longer whether AI can read a claim file or score a roof. It's whether those outputs hold up at reserve review, in the regulator's file, and at renewal.

The P&C AI value gap

Investment is accelerating faster than results. BCG's March 2026 analysis, The AI-First Property and Casualty Insurer, expects insurance AI spending as a share of revenue to triple in 2026, yet finds only 38% of P&C insurers generating value at scale from AI in core workflows. BCG also estimates that operating cost per dollar of premium could fall 15% to 25% for carriers that redesign workflows around AI.

Regulators are moving in parallel. As of July 2026, 25 states have formally adopted the NAIC Model Bulletin on AI, with 8 more in progress, and non-adopting states are still applying bulletin-aligned expectations through market conduct examinations.

Why P&C AI stalls after the pilot

A claims intake agent extracts fields from first notice of loss (FNOL) submissions, ACORD forms, and police reports with high accuracy on a validation sample. Six months after expansion, adjusters re-check most of what it populates and no one can say whether it's improving. The pilot didn't fail. It never earned the right to be relied on.

The book changes continuously. A new plaintiff firm's demand letter template, a new imagery provider, or a revised policy form shifts the data. Accuracy measured at launch stops describing reality.
Uncertain outputs aren't routed. They're either accepted, creating silent leakage, or every file gets touched, erasing the expense case.
Adjuster and underwriter corrections are lost. Overrides land in claim notes and referral emails, so each new line, state, or program starts from zero.

Where unverified AI lands: loss and expense ratios

In insurance, AI errors rarely show up on a model dashboard. They show up in the financials, often long after the decision that caused them.

Component	How unverified AI shows up	When it surfaces
Pure loss ratio, current accident year	Risk priced on stale property attributes; coverage terms missed at intake	At claim, renewal, or the next catastrophe
Pure loss ratio, prior-year development	Inconsistent injury coding and reserve calls across adjusters and TPAs	One to several years later, as adverse development
Expense ratio	Re-keying, parallel manual review, and unnecessary inspections	Immediately, spread across headcount and vendors
Regulatory exposure	Missed state notices; AI-influenced decisions that can't be explained	At market conduct exam, audit, or complaint

A model that saves 20 minutes per file but shifts reserve accuracy by a few points hasn't saved money. It has moved cost from the expense ratio into future reserve development, where it's harder to see and more expensive to fix.

What the NAIC AI Model Bulletin expects

The bulletin asks insurers to maintain a written AI systems program covering how systems are selected, validated, monitored, and retired, with testing before deployment and monitoring after. The insurer remains accountable when an AI model affects a consumer decision, even when the model came from a third party. In practice, that turns AI oversight into an evidence problem: carriers need a traceable decision path for every AI-influenced outcome.

The trust loop on a roof and in a casualty file

Property underwriting

After a hail season, an imagery provider refreshes coverage in a growth territory where stone-coated steel roofing is common, and roof condition confidence drops on roofs that resemble weathered shingle. With the trust loop in place, AI Trust detects the confidence shift. CloudFactory's AI Platform routes high-value or hazard-flagged properties to CloudFactory Evaluators, who confirm the material, correct the attributes, and document the visual signature. The validated examples enter the evaluation set, then the training set, and the knowledge graph links the condition to the territory, imagery source, and roof type.

Casualty claims

A plaintiff firm adopts a new demand package format, and extraction confidence on treatment chronologies drops. The platform prioritizes files with severe injury indicators or approaching state notice deadlines. Evaluators build the verified chronology, confirm injury coding, and correct the extraction schema. The next demand from that firm is read correctly, and reserve inputs stay consistent across adjusters, offices, and TPAs.

Five moves for claims and underwriting executives

Measure AI on pure loss and expense ratios, not dashboards. Track a weekly trajectory against a governed evaluation set, not a single launch accuracy score.
Fund the exception path, not only the agent. Budget for routing rules, catastrophe surge review capacity, and converting reviewed cases into training data.
Make imagery and claim files an asset class. Inventory imagery, demand packages, medical records, and adjuster notes, and assign owners.
Write decision rights before autonomy goes live. Define what runs straight through, what needs validation, and what escalates, by claim value, severity, coverage complexity, and state.
Capture the judgment of your best adjusters and underwriters. Every expert override is knowledge the organization is about to lose unless it becomes training and evaluation data.

Key takeaways

P&C AI errors land in the pure loss ratio and expense ratio, often years after the decision.
Pilots stall because the book changes, uncertainty isn't routed, and corrections are lost.
The NAIC Model Bulletin makes traceability and ongoing monitoring an examination issue.
The unit of progress is a decision that holds up at reserve review, in the regulator's file, and at renewal.

Glossary

Combined ratio: Pure loss ratio plus expense ratio; below 100 indicates an underwriting profit.

Pure loss ratio: Accident year losses plus prior-year reserve development, divided by net earned premium.

Adverse reserve development: An increase in estimated losses for prior accident years, often traced to inconsistent early reserving.

NAIC AI Model Bulletin: National Association of Insurance Commissioners guidance requiring insurers to govern, test, and monitor AI systems used in consumer-impacting decisions.

Get the full insight paper

Trusted AI in Property and Casualty Insurance goes deeper on the trust loop, the five domains where it changes outcomes, and how Property Data Intelligence and Casualty Claim Intelligence turn imagery and claim files into verified decisions.

Download the free insight paper