Generative AI claims workflows have moved from experiment to operating model. Bain & Company estimates the technology represents a $100 billion annual opportunity in P&C claims handling, with potential 20-25% reductions in loss-adjustment expense and 30-50% reductions in claims leakage. The use cases driving that value are concrete: LLM claim summarization, FNOL automation, fraud-signal scoring, and the adjuster AI copilot that drafts letters and reserve recommendations. This whitepaper explains how Quick Silver Systems delivers these capabilities inside the Mercury Policy and Claims Administration System with a documented governance posture aligned to NAIC and state-level GenAI insurance guardrails.
Claims is the moment of truth in P&C insurance. It is also the cost center where margin compression, talent shortages, and rising severity collide. Bain's October 2024 analysis sized the opportunity bluntly: generative AI applied across the claims value chain represents roughly $100 billion in annual value worldwide, anchored by 20-25% reductions in loss-adjustment expense and 30-50% reductions in leakage. Those numbers assume disciplined deployment of a handful of well-understood use cases.
The story on fraud is equally striking. Deloitte's 2025 predictions project $80-160 billion in cumulative savings from AI-driven fraud detection by 2032, against a backdrop in which roughly 10% of P&C claims are fraudulent — about $122 billion in annual losses. In the same survey, 35% of executives named fraud detection as their top GenAI use case. Carriers that operationalize these capabilities now convert a one-time investment into a permanent unit-economics advantage.
Not every GenAI use case is mature. The honest map separates three tiers. Production-grade capabilities include file summarization, conversational FNOL intake for low-complexity lines, document classification, and copilots that draft correspondence under human review. Emerging capabilities — agentic workflows, zero-touch settlement for narrow scenarios, automated coverage analysis — are working at leading carriers but require tighter guardrails. Experimental use cases such as fully autonomous reserve setting on complex bodily-injury claims remain unsuitable for production without expert-in-the-loop oversight.
The pattern that distinguishes value-creating deployments from stalled pilots is integration depth. A standalone tool produces a summary; a tool integrated into the Mercury platform produces a summary that updates the file, populates diary entries, drafts the next letter, and writes a complete audit trail. Bain calls this the gap between "feature" and "operating model."
| Use Case | Maturity | ROI Mechanism | Guardrail |
|---|---|---|---|
| FNOL conversational intake | Production | Cycle-time reduction; deflection from call center | Scripted fallback to human agent |
| Claim-file summarization | Production | 20-25% LAE reduction; faster reassignment | Source-cited extracts; adjuster sign-off |
| Adjuster AI copilot (drafting) | Production | Desk-time reduction; consistency | Human review before send; template library |
| Fraud signal detection | Production | 10-15% detection lift; leakage reduction | Explainable scores; SIU referral workflow |
| Auto-decision (low-complexity) | Emerging | Zero-touch settlement on narrow scenarios | Coverage and severity caps; opt-out path |
| Coverage analysis | Emerging | Faster declination decisions; consistency | Counsel review on declinations |
First Notice of Loss is where claim economics are set. A poorly captured FNOL drives rework downstream; a well-captured FNOL routes the claim to the right adjuster with the right authority on the first touch. FNOL automation built on conversational LLMs now handles the full intake conversation for low-complexity scenarios — windshield, baggage, simple property — including coverage verification and direct routing to a vendor network. Newgen's 2026 analysis of agentic deployments documents zero-touch settlement on these narrow lines, with cycle times measured in minutes rather than days.
AI claims triage sits immediately downstream. A triage model reads the intake, classifies severity and complexity, identifies coverage and SIU flags, and recommends an adjuster queue with confidence scores a supervisor can override. The result is fewer reassignments, fewer coverage errors, and a measurable reduction in time-to-first-contact — the strongest predictor of severity and litigation rate on bodily-injury claims.
Adjusters spend an estimated 30-40% of their day reading. A complex liability file routinely exceeds 200 pages of medical records, police reports, witness statements, vendor estimates, and prior correspondence. LLM claim summarization compresses that file into a structured brief — facts of loss, coverage position, parties, reserves, open issues, and next-best actions — with every assertion linked back to the source page. The output is reviewable in minutes, not hours.
Two design choices separate working summarizers from impressive demos. The first is retrieval grounding: the model summarizes only what is in the file, with citations, and refuses to speculate on missing facts. The second is workflow integration. A summary that refreshes when a new document is indexed, and that converts into a diary note or reserve-change recommendation with one click, is what produces the Bain-projected LAE improvement. Mercury treats summaries as first-class file artifacts, version-controlled and auditable.
AI fraud detection insurance is the most-funded GenAI use case in P&C, and the production evidence is compelling. Emerj's 2026 review of Allianz documented £37.7 million in fraud savings during the first half of 2024 alone, a roughly 10% lift in detection rates on motor claims, and a 150% increase in detected application fraud after deploying AI scoring at quote time. These results come from layering classical anomaly detection, network analysis on parties and providers, and LLM-based narrative scoring that reads loss descriptions for inconsistencies.
The economics compound when scoring is applied at FNOL rather than only at payment review. Catching a fraudulent claim at first notice avoids the indemnity payment, the LAE of investigation, and the SIU caseload bottleneck. Mercury surfaces signals as explainable scores — never opaque verdicts — with the underlying features and the recommended SIU referral path.
The adjuster AI copilot is where the other use cases converge into a daily workflow. In Mercury, the copilot lives in the claim file: it summarizes new documents as they arrive, drafts correspondence in the carrier's voice, suggests reserve changes with reasoning, surfaces fraud and subrogation signals, and prepares the next-best-action list before the file is opened. Every output is human-reviewed before any external action is taken.
The regulatory perimeter around GenAI insurance guardrails tightened materially in 2024 and 2025, and carriers should expect it to keep tightening. ZwillGen's 2025 analysis walks through the implications of the Colorado AI Act and the NAIC Model Bulletin on Use of Artificial Intelligence Systems by Insurers: carriers must maintain a written governance program, document the training data and intended use of every model that affects a consumer, monitor for adverse outcomes, and produce that documentation on request. A copilot built without those artifacts is a market-conduct exam waiting to happen.
Per ZwillGen's guidance, carriers must document at minimum: (1) intended use and consumer impact, (2) training data sources and known limitations, (3) testing for bias and accuracy, (4) human-review checkpoints, and (5) ongoing monitoring for adverse outcomes. Mercury records all five automatically as a structured governance log on every AI-assisted claim — turning a regulatory burden into a defensible audit artifact.
The Bain $100 billion estimate is not a forecast — it sizes value that already exists, waiting to be captured. Summarization, intake automation, triage, and fraud scoring all have production evidence at scale, with measurable LAE, leakage, and cycle-time impact. What separates carriers that capture the value from carriers that pilot indefinitely is depth of integration into the claims operating system and rigor of governance documentation.
Quick Silver Systems built the copilot directly into Mercury for exactly that reason. It shares the same data model, audit trail, and role-based access controls as the rest of the platform, and every AI-assisted action is logged against the schema that satisfies the NAIC Model Bulletin and the Colorado AI Act. Carriers that adopt this posture turn capability into a durable cost and quality advantage — without inheriting the regulatory exposure that follows ungoverned deployments.
Quick Silver Systems, Inc. makes the Mercury Policy and Claims Administration System. Contact us for a working session on use-case prioritization, governance design, and ROI modeling for your book.