How to Use AI for P&C Claims Operations

Loss adjustment expense is one of the largest controllable costs in property and casualty insurance. U.S. P&C insurers spent $86.0 billion on LAE in 2025, roughly 15.6 cents for every dollar of loss paid. Much of that money goes to work that doesn't require an experienced adjuster: taking down loss details, chasing documents, answering "where's my check?" calls, and keying data from one system into another.

That's where applying AI to P&C claims operations starts to pay for itself. The value isn't in a flashy chatbot. It comes from removing the routine work that buries your adjusters so they can focus on the claims that need judgment.

This guide is written for operations leaders who have already sat through the vendor demos. It covers where AI earns its keep across the claim lifecycle, where it creates new risk, and how to roll it out without your legal team or regulators pulling the plug six months in.

Table of Contents

• The Evolution of AI in P&C Claims Management

• Automating FNOL and Initial Data Triage

  • Natural Language Processing for Intake

  • Computer Vision for Instant Damage Assessment

• Enhancing Fraud Detection and Risk Mitigation

• Streamlining Adjudication and Straight-Through Processing

  • Algorithmic Decisioning for Low-Complexity Claims

  • Predictive Analytics for Litigation and Severity

• Optimizing the Claimant Experience Through Personalization

• Implementing an Ethical AI Framework for the Future

The Evolution of AI in P&C Claims Management

Claims automation is older than it sounds. Rules engines have flagged suspicious claims and routed files by line of business for decades. They were brittle, though. Every new scenario meant another hard-coded rule, and anything outside the rulebook landed back on someone's desk.

The next wave brought machine learning models that scored severity and fraud likelihood from historical claims data. These models were better at handling nuance, but most lived in analytics teams and fed dashboards rather than changing how work actually moved.

Two shifts define where carriers stand in 2026:

• Language models that understand unstructured input. Phone calls, emails, adjuster notes, and repair estimates can now be read and summarized reliably. That matters because most claims data has always been unstructured.

• Agents that take action. The newer systems don't just score or summarize. They can open a claim, update a record, send a document, and hand off to a human when something falls outside their rules.

The practical result is that AI has moved from the back office to the front line of claims handling. That also means its mistakes now reach policyholders directly, which is why governance has become as important as accuracy.

Automating FNOL and Initial Data Triage

First notice of loss sets the tone for the whole claim. If the intake is incomplete, the adjuster spends the first two days calling the claimant back for details that should have been captured on the first call. FNOL is also high-volume and spiky. A hailstorm can triple call volume overnight, and staffing for those peaks is expensive.

AI-driven intake handles the volume and enforces completeness, which makes it the most sensible first project for most carriers and brokerages.

Natural Language Processing for Intake

Modern NLP lets a claimant describe a loss in their own words, by phone or email, while the system pulls out the structured fields your claims platform needs. These include date of loss, location, parties involved, injuries, police report numbers, and policy number. If something is missing, the agent asks for it right away instead of leaving a gap for an adjuster to discover later.

The measure that matters here is how much of the intake lands in your core system without anyone retyping it. Arvy's AI Agents for Claims are built around this step. They capture loss details over phone and email, record the new claim inside your existing claims system, and answer follow-up status questions. When a caller mentions a serious injury or sounds distressed, the agent passes the conversation to a person along with the full context.

Computer Vision for Instant Damage Assessment

Photo-based damage estimation has matured for auto physical damage and is improving for property. A claimant uploads images, and the model identifies damaged parts, estimates repair versus replace, and flags likely total losses early.

Treat these outputs as triage, not final numbers. Use them to route files: a cosmetic bumper scrape goes to a fast-track queue, while a suspected structural frame issue goes to a field appraiser. Carriers who try to pay directly off the model's estimate without a review threshold usually end up dealing with supplements and disputes.

Enhancing Fraud Detection and Risk Mitigation

Fraud models are the most established use of AI in claims, and they're getting sharper. The real gain in 2026 comes from feeding them richer intake data. When NLP captures a claimant's full narrative, models can spot inconsistencies such as a timeline that shifts between the first call and the recorded statement, or language that matches other suspicious claims.

A few practices separate useful fraud AI from noisy fraud AI:

• Tune for SIU capacity. A model that flags 12% of claims when your SIU can investigate 2% only produces a backlog and alert fatigue.

• Score at several points. Run scores at FNOL, after the first estimate, and before payment. Fraud signals often appear mid-claim.

• Never let a score deny a claim on its own. A flag should trigger an investigation, not a decision.

Documentation needs a hard look as well. A February 2026 federal ruling narrowed attorney-client privilege for purely automated AI-assisted investigative reports. If your fraud or coverage investigations depend on AI-generated reports with no attorney direction, review those workflows with counsel before you assume the reports are protected.

Streamlining Adjudication and Straight-Through Processing

Straight-through processing is the goal everyone talks about and few reach at scale. The reason usually isn't the technology. Most carriers haven't defined clearly enough which claims qualify. Start narrow and widen the scope as evidence builds.

Algorithmic Decisioning for Low-Complexity Claims

Good STP candidates share the same traits: low severity, clear coverage, a single party, no injury, and documentation that can be verified automatically. Examples include glass claims, simple water damage below a set threshold, and rental reimbursement.

Write the eligibility criteria down, set dollar caps, and have the system log why each claim passed or failed every rule. Begin with automated approvals only. Any claim that would be denied or reduced should go to a human. This keeps the risk on the right side while you build confidence, and it lines up with where regulators are heading.

Predictive Analytics for Litigation and Severity

At the other end of the spectrum, the biggest savings come from spotting the claims most likely to escalate. Litigation propensity models use signals such as attorney involvement, injury descriptions, jurisdiction, and delays in contact. They can flag a bodily injury claim within days instead of weeks.

The payoff comes from what you do once a claim is flagged. Assign it to a senior adjuster, set reserves earlier, and make proactive contact. A claimant who hears from someone quickly is less likely to hire a lawyer out of frustration.

Optimizing the Claimant Experience Through Personalization

Ask claims leaders what drives complaints, and most will say the same thing: silence. Claimants rarely leave because a claim took 30 days. They leave because they heard nothing for 12 of those days and had to call three times to find out why.

Status calls are pure cost for your team and pure frustration for the claimant, which makes them ideal for automation. An AI agent connected to your claims system can tell a policyholder exactly where things stand: the inspection is scheduled for Thursday, the estimate has been approved, or payment was issued yesterday by direct deposit. It can do this at 9 p.m. on a Sunday, in the claimant's preferred language and channel.

Personalization goes further than status updates. Some useful patterns:

• Send proactive updates at each milestone so claimants don't need to call.

• Adjust tone and pacing for vulnerable claimants, such as those dealing with a total home loss or an injury.

• Deliver documents like payment letters and repair authorizations automatically when they're ready.

• Escalate quickly with full context. Nothing angers a claimant more than repeating their story to a human after already explaining it to a bot.

Arvy's agents follow this model. You decide what they're allowed to answer, every action is traceable, and anything that needs judgment goes to your team with the conversation history attached.

Implementing an Ethical AI Framework for the Future

Regulation is moving quickly, and health insurance is leading the way. In Washington, SB 5395 bars using AI as the sole basis to deny or modify health-related claims as of June 11, 2026. Alabama's SB 63, effective October 1, 2026, requires insurers to certify annually that a human professional makes the final adverse determination. These laws target health plans rather than P&C, but they show the direction regulators are taking. Sensible P&C carriers are building to that standard now.

Liability is changing on the commercial side too. New ISO endorsements introduced in January 2026 let CGL carriers explicitly exclude bodily injury or property damage arising from generative AI. The industry is pricing AI risk for its customers, and it should apply the same scrutiny to its own operations.

A workable framework covers four things:

1. Human-in-the-loop for adverse outcomes. Denials, reductions, and coverage disputes always get human sign-off.

2. Full audit trails. Every AI action should be logged and explainable to a regulator or a plaintiff's attorney.

3. Defined authority. Write down what each agent can and cannot do, in plain language.

4. Bias testing. Check outcomes across geographies and claimant groups at least quarterly.

Across these sections, the same pattern holds. The carriers getting real results from AI aren't automating everything at once. They start with intake and status updates, where volume is high and risk is low. They put clear rules around decisions and keep humans responsible for every outcome that hurts a claimant. Over time, that LAE line shrinks and adjusters get back to the work they were hired to do.

The fastest way to prove value is to choose one high-volume workflow, such as FNOL intake or claim status calls, and measure it properly. Arvy works alongside your team to connect your systems, launch a focused pilot quickly, and track the impact from the first week. Start With One Workflow and let the results guide what comes next.