Insurance Claims Processing Automation With AI Agents

Only a third of insurers run claims AI at scale. A survey of 302 insurers by HFS Research and Xceedance found 33% with AI deployed at scale or end to end in claims, while a slightly larger share remain in pre-production (HFS Research and Xceedance, 2026).

Many stalled pilots share a gap: the agent never learned how claims work runs across the screens, documents, and judgment calls adjusters handle every day.

What is AI claims processing?

AI claims processing, also called claims processing automation, moves a claim from first notice of loss (FNOL) to payment. Optical character recognition (OCR) and intelligent document processing (IDP) extract intake documents, natural language processing (NLP) reads loss descriptions, rules engines validate coverage, machine learning (ML) scores fraud and severity for triage, computer vision estimates damage from photos, and rules settle clear claims.

Straight-through processing (STP) covers claims that clear every check, while adjusters take denials, injuries, exceptions, and complex losses.

Two earlier automation layers came before claims agents. Robotic process automation (RPA) bots follow pre-defined scripts across claims screens, and log-based process mining rebuilds workflows from system event data. Neither sees the full claim when work crosses a mainframe, a Citrix session, email, and Excel.

Skan AI captures processes across claims applications, stitches the activity into end-to-end claim flows and exception paths, and models the result as a Context Graph of Work: a living map of the steps, decisions, exceptions, and handoffs behind each claim.

The platform holds SOC 2 Type 2, ISO/IEC 27001, and ISO/IEC 42001 certifications and is compliant with the Health Insurance Portability and Accountability Act (HIPAA), the General Data Protection Regulation (GDPR), and the California Consumer Privacy Act (CCPA).

How does claims automation work in insurance, from FNOL to payout?

AI handles extraction, scoring, and rule checks at every stage, while adverse decisions, liability, and negotiation stay with people. The split looks like this:

Deloitte says simple, low-severity claims, such as auto losses without injuries and straightforward property damage, can flow straight to payment under human oversight. Celent places the next wave of value in the roughly 15 percent of claims that are complex, multi-party, document-heavy, or litigation-prone (Celent, 2026).

A worked example: a minor auto collision

In this illustrative workflow, another driver rear-ends a policyholder at low speed, with no injuries:

• Intake, triage, and coverage: At FNOL the customer uploads photos through an app link, and NLP extracts the date, location, and vehicle from the description. The carrier’s triage model scores the claim as low severity, and its rules engine confirms active collision coverage.
• Damage assessment: Computer vision converts the photos into estimate lines. CCC reports its AI can write 80% or more of estimate lines, aligned to each insurer’s rules, as a preliminary estimate for desk-appraiser review, and its configuration guide lets carriers set confidence thresholds between 70% and 99%.
• Adjudication and payment: Under this example’s carrier-approved rules, the carrier issues payment automatically when the estimate clears the confidence threshold and dollar limit and the fraud score is low; otherwise a desk appraiser reviews it. West Bend Mutual told Insurance Journal it reviews AI photo estimates 100% of the time.
• Subrogation: The recovery team uses subrogation scoring to flag the other driver’s carrier for recovery.

Which technologies power AI claims processing?

AI claims processing relies on five techniques, and each one does a single, distinct job in the claim, as the list below shows.

• OCR and IDP: These tools turn police reports, medical records, and invoices into fields; carriers use confidence scores to route uncertain extractions to a person.
• NLP: It reads FNOL emails and adjuster notes for dates, amounts, and parties.
• Machine learning: It scores severity, fraud risk, and recovery likelihood.
• Computer vision: It converts damage photos into assessments and estimate lines.
• Rules engines: They apply coverage terms and authority limits deterministically.

CIO cites IBM’s explanation that RPA bots require manual intervention when they hit an exception. Traditional ML predicts a score but does not act on it. Agentic AI plans and executes multi-step work, and Gartner analyst Anushree Verma advises using AI agents where decisions are needed, automation for routine workflows, and assistants for simple retrieval (Gartner, 2025).

Why do claims agents need process context captured from real work activity?

Claims agents need observed process context because claims work rarely lives in one system: a single claim can touch the core claims platform, a mainframe for policy verification, Citrix-hosted portals, email, and Excel reserve sheets, with undocumented handoffs between them.

Adjusters often apply payout rules through judgment, and exceptions appear only in real cases. Gartner warns that integrating agents into legacy systems can be technically complex, often disrupts workflows, and can require costly modifications (Gartner, 2025). Carriers need this process context for property and casualty (P&C) automation with AI.

Skan AI is an agentic AI platform for regulated enterprises. It builds a Context Graph of Work, a map of the steps, decisions, exceptions, and handoffs observed in process activity so carriers can ground agent decisions in how claims work runs.

Skan AI reports that 3 in 5 of the top US insurance carriers use its platform.

Skan Virtual Assistant (Skan VA), a lightweight desktop sensor, captures screens, clicks, and application switches across mainframe, Citrix, virtual desktop infrastructure (VDI), web, and office tools, with privacy by design. The sensor captures this work with zero integrations, application programming interfaces (APIs), or logs, reading the desktop directly rather than connecting to backend systems.

Skan PRISM redacts sensitive data at the pixel level, masking it in the screen image itself, on premises as part of privacy by design, and the platform’s privacy controls pseudonymize identities, replacing them with stand-in identifiers, so analysis stays on work patterns.

The graph works as a digital twin of operations that updates as procedures change.

How do AOPs keep an agent on the approved procedure?

A hand-built agent tends to deviate from current practice when its builders encode only the documented procedure and a few interviews. Exception paths and site-level workarounds are never encoded, and the real procedure keeps changing after launch.

Agentic operating procedures (AOPs) address this. AOPs are machine-readable process definitions generated from observed claims work, encoding steps, decision logic, exceptions, escalation thresholds, and success criteria.

• Skan AI simulates and evaluates each agent workflow against the human baseline before production.
• Deviation alerts flag when observed work departs from the AOP, and human judgment stays in the loop where required.
• Audit trails and case-level playback record what each agent did, and each agent runs with defined permissions and escalation paths.

How does AI detect claims fraud?

ML anomaly scoring routes claims to the Special Investigations Unit (SIU); it does not deny them. The “2024 Insurer SIU Benchmarking Study” from Aon and the Coalition Against Insurance Fraud found automated referrals made up 45% of SIU referrals and 29% of accepted ones (Aon and the Coalition Against Insurance Fraud, 2024).

The National Association of Insurance Commissioners (NAIC) unfair claims model lists “Refusing to pay claims without conducting a reasonable investigation” as an unfair practice, so a score starts an investigation and an investigator decides.

How does automation differ for auto, health, and property claims?

The evidence each line produces changes intake and assessment:

• Auto: Adjusters use photos to assess damage, and some regulators preserve in-person inspection: New Hampshire’s Insurance Department issued guidance on April 14, 2025 emphasizing that consumers retain the right to request in-person inspections, as reported by CollisionWeek.
• Health: Claims teams assess medical records. California SB 1120 says a determination of medical necessity “shall be made only by a licensed physician or licensed health care professional,” and HIPAA’s minimum necessary standard limits PHI exposure.
• Property: Low-value contents claims suit STP. Lemonade reported roughly 55% of its claims across all lines automated as of December 31, 2024.

What results are realistic?

Realistic results are line-specific and proven against the carrier’s own baseline. Fix four KPIs before a pilot:

• STP rate: Measure it by line and claim type, never as one blended figure.
• Cycle time: J.D. Power data reported by Carrier Management puts property FNOL-to-payment above 44 days.
• Cost per claim: Agree the formula before the pilot, because carriers define it differently.
• LAE: The NAIC splits loss adjustment expense into defense and cost containment, and adjusting and other.

Slow claims carry operational cost and declining customer satisfaction: the same J.D. Power property claims study scores claims completed within 10 days at 762, against 595 when repairs take more than 31 days. Capgemini’s “World Property & Casualty Insurance Report 2025” cites 20-30% lower loss adjustment expenses with generative AI claims augmentation (Capgemini, 2025), and Datos Insights notes that STP is effective for simpler claims but is not a realistic goal for every claim. Published STP figures rarely share a basis: Insurance Thought Leadership reported in 2021 that, on average, fewer than 10% of claims are processed straight through in any line (Insurance Thought Leadership, 2021), while one vendor claims up to 80% STP for auto claims, so measure any vendor rate against your own observed baseline in a pilot.

Will AI replace claims adjusters?

AI takes intake, extraction, routine validation, and payment on clear claims, while adjusters keep adverse decisions, liability, and complex judgment.

Forrester predicts that most carriers will not let an unsupervised AI agent bind or decline coverage, set prices, deny claims, or authorize payouts. The Bureau of Labor Statistics (BLS) projects a 6% employment decline for claims adjusters, appraisers, examiners, and investigators from 2025 to 2035, with about 21,600 openings a year from replacement.

What do regulators require of automated claim decisions?

Regulators hold insurers accountable for AI they build or buy, and the expectations cover governance and documentation, state adoption of the NAIC model bulletin, and claimant rights in the European Union, as follows:

• Governance and documentation: The NAIC Model Bulletin, adopted December 4, 2023, expects insurers to maintain a written AI program covering claim administration and examinable documentation. Insurers may include vendor contract terms with audit rights where appropriate and available.
• State adoption: The NAIC’s August 2026 map lists 24 states plus DC. Faegre Drinker reports that South Carolina adopted on September 24, 2026. Insurers must check each adopting jurisdiction’s version for its expectations.
• EU claimant rights: GDPR Article 22 gives EU claimants a right not to be subject to solely automated decisions with legal or similarly significant effects and, where the contract or explicit-consent exceptions apply, a right to human intervention.

Skan AI’s neurosymbolic models pair neural pattern recognition, which learns patterns from data, with symbolic rules, which are explicit written logic, for deterministic-first execution. Case-level playback and continuous controls monitoring support compliance teams with evidence gathering and control verification across the deployed scope. They do not guarantee compliance, and the carrier remains accountable for it.

How do you start automating claims processing?

Start automating claims processing by observing the work before you automate it, working through three steps in order: scan, observe, and pilot.

1. Scan. Skan AI Blueprint maps opportunities for AI transformation and separates work to agentify, augment, or keep human.
2. Observe. Skan Intelligence builds case-level evidence of cycle time, rework, and exceptions without backend integrations or changes to core claims systems.
3. Pilot. Skan AI Agents run pilots under AOPs, measured against the observed baseline. As first-party deployment evidence from Skan AI, not independent research or a universal outcome, a Fortune 200 P&C carrier achieved a 14% productivity gain in claims in a documented deployment.

Legacy claims platforms often lack APIs and event logs, which stalls log-based tools. Data-driven insurers start on the adjuster’s desktop.

See how observed claims work becomes governed agent execution. Request a Skan AI demo to explore how observed workflows become context for governed AI execution.