How Can AI Agents Review Workers' Comp and Liability Claim Files?

Raji Chadarevian of the National Council on Compensation Insurance (NCCI) reports that closing a workers’ comp claim takes an additional 9 to 25 weeks after all medical services have been delivered, depending on jurisdiction (Triple-I, 2026). Those weeks are where cycle-time and leakage (claim dollars paid beyond what the file supports) pressure build, and AI agents for workers’ comp and liability claims now assemble the file, extract data, and draft a cited summary while coverage, settlement authority, and special investigations unit (SIU) referral stay with people.

The only way to limit these agents is with real-world context: how your adjusters, nurses, and defense counsel work a file.

What AI agent claim file review means

AI agent claim file review is software that reads every document in a file, reconciles facts across those documents, and drafts work product that an adjuster or attorney accepts, edits, or rejects. The adjuster audits a cited draft and makes the call instead of reading hundreds of pages and re-keying data.

How AI agents review a claim file, step by step

An AI agent for workers’ comp and liability claims works a file in six steps, from ingesting every document to routing a cited summary to the adjuster, and a person owns the last one:

1. Ingest every document in the file.
2. Index each document by type, date, provider, and body part.
3. Extract structured data from forms and narrative records.
4. Cross-reference treatment, causation, liability, and reserve facts.
5. Summarize findings with a source citation for each assertion.
6. Route the summary to the adjuster for sign-off.

Ingesting and indexing the file

The agent pulls first notice of loss (FNOL) data, medical records, bills, adjuster notes, depositions, and demand packages into one index. A multi-year file can contain repeat surgeries and overlapping independent medical examinations (IMEs) so the agent orders records by date of service, removes duplicates, and tags each page by provider and body part to separate the claimed injury from prior injuries.

Extracting data from unstructured medical and legal records

Optical character recognition (OCR) converts scanned and handwritten pages into text, and natural language processing (NLP) and large language model (LLM) extraction read the narrative text. On a CMS-1500, the standard medical bill form for professional services, the agent reads International Classification of Diseases, 10th Revision, Clinical Modification (ICD-10-CM) diagnosis codes in item 21, Current Procedural Terminology (CPT) or Healthcare Common Procedure Coding System (HCPCS) procedure codes in item 24D, and the item 24E pointer that ties each service line to a diagnosis (CMS); on a UB-04, the standard hospital bill form, it reads HCPCS codes in form locator (FL) 44 and the principal diagnosis in FL 67. From demand letters, it pulls the amount, the acceptance deadline, and the release terms.

Cross-referencing and analyzing

The agent builds a medical chronology (a dated timeline of every treatment) and tests each treatment request against the governing guideline: California’s Medical Treatment Utilization Schedule (MTUS), the Official Disability Guidelines (ODG) that Texas names for non-network care (TDI), or New York’s Medical Treatment Guidelines (MTG) (NY WCB, 2026). The agent flags prior injuries to the same body part, diagnoses that do not fit the reported mechanism, and care beyond guideline frequency or duration.

Synthesizing a reviewable summary

The hand-off contains a chronology, guideline exceptions, open questions, and suggested next actions, with every assertion linked to its source page. The agent writes nothing to the claim record until the adjuster signs off.

What the agent analyzes in a workers’ comp file

In a workers’ comp file, the agent tests bills against the state fee schedule and tracks the facts that drive indemnity (wage-replacement payments to the injured worker) and future medical cost, including the date of maximum medical improvement (MMI), when the worker’s condition has stopped improving:

• Medical bill review: Texas requires Medicare payment policies, including CCI edits and modifiers, for professional services (TDI). The agent flags a component code billed with its comprehensive code and no modifier.
• Treatment-guideline compliance: The agent applies the MTUS, ODG, or New York MTG checks described above to each request.
• MMI and impairment: The agent extracts the MMI date and rating and records which AMA Guides edition the state uses: California the 5th Edition (California DIR) and Colorado the 3rd Edition (Revised) (7 CCR 1101-3).
• Return-to-work timing: The agent tracks restrictions and release dates.
• Future care cost: The agent assembles treatment plans and life expectancy inputs for reserves and Medicare Set-Asides. CMS reviews a WCMSA when a Medicare beneficiary settles for more than $25,000 (CMS, 2026).

What the agent analyzes in a liability file

In a liability file, the agent reads depositions, demand packages, and defense counsel correspondence to frame fault, limits exposure, and value. The negligence rule and the demand deadline drive the review:

• Comparative negligence: Alabama, Maryland, North Carolina, Virginia, and D.C. apply contributory negligence (D.C. carves out pedestrians and bicyclists). Florida moved to modified comparative fault in 2023 and bars recovery by a party found more than 50 percent at fault, outside medical negligence actions (Fla. Stat. 768.81).
• Policy-limits demands: A Texas Stowers duty attaches when a covered claim’s demand is within limits and an ordinarily prudent insurer would accept it (Tex. Sup. Ct., 2021). The agent surfaces each deadline and condition on the first day of review.
• Settlement valuation: The agent compiles specials, wage documentation, and deposition testimony on mechanism and injury. The adjuster and counsel set the number.

Fraud, subrogation, and reserve flags

Agents score files against the carrier’s documented red flags and models early in the claim, then route each flag to a person. California requires every SIU to keep written procedures that list the red flags its staff use (10 CCR 2698.45), so claims teams can encode those red flags as the agent’s rule set. The agent routes each settlement and litigation risk score to the adjuster.

California’s Bulletin 2022-5 cites allegations that insurers flagged claims from certain inner-city ZIP codes for SIU referral (CDI, 2022). Claims teams should test fraud models for bias.

The same review flags evidence of a liable third party for subrogation, the carrier’s right to recover costs from that party. For reserves, the money set aside for expected claim costs, the agent surfaces MMI status, impairment rating (a physician’s measure of permanent loss), treatment plan, demand amount, and litigation status, and the adjuster sets the reserve.

Which claim decisions stay with a human adjuster?

Licensed adjusters, nurses, physicians, and attorneys keep coverage decisions, settlement authority, SIU referrals, reserve changes, and medical-necessity denials, and the agent prepares the evidence for each.

State law sets several of these lines. California Labor Code §4610 bars anyone other than a licensed physician from modifying or denying a workers’ comp treatment request for medical necessity. Texas SB 815, a health utilization review law, bars utilization review agents from using an automated decision system to make an adverse determination, wholly or partly. The National Association of Insurance Commissioners (NAIC) states that “Existing state insurance laws apply regardless of whether decisions are made by humans, algorithms, or third-party systems” (NAIC, 2026).

Claims leaders encode those lines as governed autonomy: agent permissions, stopping thresholds, and escalation points. The agent acts within that authority and escalates with its evidence attached. Adjusters and nurse case managers keep the conversations with injured workers.

Compliance, privacy, and audit trails

As of August 6, 2026, 24 states plus D.C. had adopted the NAIC Model Bulletin, the model guidance on how insurers should govern their use of AI systems (NAIC, 2026). It lists “claim administration and payment, and fraud detection” among the uses an insurer’s artificial intelligence systems (AIS) Program, its written governance program for AI, should address (NAIC, 2023).

• Exams and vendors: The bulletin says insurers should maintain a written AIS Program, and examiners can ask about an AI system’s development, use, and outcomes. It also calls for vendor audit rights where appropriate and available.
• PHI: Many WC carriers are not HIPAA covered entities, and 45 CFR 164.512(l) permits disclosures that state workers’ comp law authorizes; California Labor Code §3762 still limits what insurers and TPAs tell employers about a claimant’s medical information.
• Privilege: In United States v. Heppner, the S.D.N.Y. held in February 2026 that a defendant’s consumer-AI documents were not privileged (Harvard Law Review, 2026). Whether counsel-directed enterprise AI use preserves privilege remains unsettled, so route defense material through counsel before any AI tool touches it.
• Audit trail: Record every input, output, decision, and plain-language rationale at the case level, including what the adjuster changed.

Integrating with claims systems and running a pilot

Claims teams should plan approved note and activity write-back, the saving of the agent’s output into the claims system, so the audit trail sits where adjusters and examiners already look. Guidewire’s Cloud API, an application programming interface (API) through which other software exchanges data with its claims system, supports creating activities on a claim (Guidewire). Sedgwick reports using API integration to deliver medical document summaries directly into claim files. A pilot moves from shadow mode to assist mode:

• Shadow mode: The agent drafts summaries and flags while adjusters work normally, and reviewers compare the two outputs by error type.
• Assist mode: Adjusters receive the draft and edit or reject it. Each reversal updates the procedure the agent follows.

Measuring agent accuracy and ROI

Measure the agent on the KPIs claims leadership already reports: cycle time, cost per claim, leakage, and STP rate. Build your leakage baseline from your own closed-file audits; the sources reviewed include no current, line-specific leakage benchmark for workers’ comp or commercial liability.

Clinical summarization studies are the nearest proxy, because the sources reviewed include no AI-versus-human study on claim files. In a BMJ Open study, physicians took 7 minutes per summary and an LLM took 15.7 seconds, with no significant difference in completeness or correctness (BMJ Open, 2025).

Score your pilot against human reviewers by error type, and count verification time alongside drafting time. Clinical summarization research tracks error types that include:

• Hallucination: The summary states content the source does not support.
• Omission: The summary leaves out facts a reviewer would have included.

Why do claims agents need operational context?

Agents grounded only in documents and event logs (the timestamped records that systems write when something happens) lack operational context, meaning knowledge of how people actually move a claim through their tools. Those inputs miss how a claim file moves across mainframe policy screens, Citrix-hosted portals, email, and Excel reserve worksheets. Process mining, which reconstructs workflows from those logs, needs integration into every system that emits event logs, and task mining, which records what a worker does on screen, captures single tasks without the end-to-end flow.

Skan AI builds the Context Graph of Work, a connected map of how claims work moves across applications and people, from observed claims activity. Skan Virtual Assistant (Skan VA) captures that work across applications under privacy by design, with zero integrations, APIs, or logs.

Skan PRISM, Skan AI’s privacy layer, applies pixel-level redaction, which masks sensitive on-screen content, inside the customer environment before metadata reaches analytics. Deployment options are on-premises inside the customer firewall, private cloud, or Skan AI hosted, so the enterprise chooses where its operational data stays. Skan AI holds SOC 2 Type 2, ISO/IEC 27001, and ISO 42001, and its assurance documentation also covers PCI DSS, HIPAA, GDPR, CCPA, and TrustArc.

• Skan AI Blueprint identifies the claims areas to automate first.
• Skan Intelligence observes those areas in depth, including exceptions and rework, and teams see first results in weeks.
• Skan AI Agents then execute approved steps under agentic operating procedures (AOPs), which set permitted actions, stopping conditions, and escalation points.