How Agentic AI Reshapes Insurance Claims Processing: Key Findings and Frameworks From Skan AI

Carriers face steady claim volume without more adjusters to handle it. The Bureau of Labor Statistics projects employment of claims adjusters, appraisers, examiners, and investigators to decline 6% between 2025 and 2035 (BLS, 2026). All projected openings come from replacing workers who leave the occupation or the labor force.

J.D. Power’s 2025 property study found satisfaction falls 167 points when repairs take more than 31 days instead of 10 or fewer (J.D. Power, 2025).

Robotic process automation (RPA, software that replays fixed scripted steps), template optical character recognition (OCR, software that reads fields from documents with a known layout), and task mining (software that records individual desktop tasks) automated clean claims, but files with conflicting documents, coverage questions, or cross-system handoffs still reach adjusters. Those tools lacked context: the decision rules, workarounds, and escalation paths adjusters apply but no standard operating procedure (SOP) or event log (a system’s automatic record of activity) records. A Context Graph of Work, covered later in this piece, is a map of that missing context drawn from how the work actually runs.

Learn more in the Insurance Claims Processing whitepaper: it covers where AI agents across the claims lifecycle fit, where adjusters keep authority, and what regulators expect.

Which claims tasks can AI agents safely handle, and which stay with adjusters?

Agents can own assembly, checking, and routing work. Adjusters keep decisions that commit money above delegated authority, assign liability, or allege fraud. The tiers below are an illustrative framework; each carrier maps agents onto its own authority limits:

Carriers should write escalation triggers into the procedure so any adjuster-only item in the table stops the agent.

How do AI agents for insurance claims processing work across the lifecycle?

An agent carries a claim through each lifecycle stage, from intake to settlement, and hands off to a human at defined points where its rules or authority end. Agentic AI is software that plans and executes multistep work toward a goal, calling tools and systems as it goes, rather than replaying one fixed script.

• FNOL intake: The agent captures loss details and opens the claim.
• Document review: The agent extracts fields from reports, estimates and photos.
• Coverage validation: The agent checks policy terms and flags conflicts.
• Triage and routing: The agent routes claims to straight-through, desk or specialist handling.
• Adjudication: The agent drafts a recommendation for the decision owner.
• Exception handling: The agent escalates files outside its rules, with evidence attached.
• Settlement: The agent prepares payment within authority and states the coverage.

From process capture to governed execution

Skan AI is an agentic AI platform for regulated enterprises that builds a Context Graph of Work, a single connected map of how claims work actually gets done, from observed process activity, so carriers can ground agent decisions and execution in how work runs. Skan VA captures desktop activity so the Context Graph of Work can map the roles, applications, tasks, decisions, policies, and exception patterns within each process:

1. Capture claims work across applications, including handoffs and workarounds.
2. Distill it into agentic operating procedures (AOPs) encoding steps, decision rules, exceptions and escalation thresholds.
3. Assemble Skan AI Agents with permissions and guardrails in Agent Studio.
4. Simulate runs against human baselines.
5. Promote to production once performance holds.

Skan AI holds a SOC 2 Type 2 attestation and ISO 27001 and ISO 42001 certifications, and is Health Insurance Portability and Accountability Act (HIPAA), General Data Protection Regulation (GDPR), and California Consumer Privacy Act (CCPA) compliant.

How does agentic AI differ from RPA, OCR, and earlier claims automation?

Agentic AI handles variance by reasoning over unstructured inputs inside set rules. Everest Group found that RPA alone fails many insurance processes because it suits rules-based, structured-data tasks, while insurance work relies on judgment and unstructured data (Everest Group, 2021). Template OCR needs a trained layout per document; an agent’s language model reads emails, narratives, and mixed packets without one.

Gartner forecasts that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls (Gartner, 2025).

Where each approach stalls:

How do agents support fraud detection, SIU referrals, and the human handoff?

Agents compile red flags, linked parties, and supporting documents into a referral package, and the special investigations unit (SIU) decides whether to open a case. In the National Association of Insurance Commissioners’ (NAIC) private passenger auto survey, insurers reported using AI fraud models mostly to refer claims for further investigation (NAIC, 2022). The insurer keeps its statutory reporting duty under laws such as California Insurance Code § 1872.4 and Texas Insurance Code § 701.051.

Subrogation, the insurer’s recovery of a paid claim’s cost from a responsible third party, follows the same split. AI models can score claims for recovery potential; recovery staff decide whether to pursue the demand.

Integrating agents with Guidewire, Duck Creek and legacy cores

Agents use the application programming interfaces (APIs) of a modern core system, the central platform a carrier runs policies and claims on, and the screens of systems without them. Guidewire’s Qusar release (August 2026) added an Agentic Framework, and its local server for the Model Context Protocol (MCP), a standard that lets AI agents connect to software tools, serves Developer Assistants (Guidewire, 2026). Duck Creek’s Agentic AI Platform supports MCP and Agent to Agent Protocol (A2A) (Duck Creek, 2026).

Skan VA captures the end-to-end claims flow across mainframes (large central computers), virtual desktop infrastructure (VDI, remote desktops hosted on a central server), and legacy desktop applications. The platform requires zero integrations, APIs (the connections that let one program call another), or logs, so those systems need no changes.

Skan AI Agents, the software agents that carry out the claims work, execute across both application APIs and the user interface (UI), so agents can act on systems that expose no API as well as modern platforms that do. The customer chooses where deployment runs: on-premises inside the customer firewall, in private cloud (a cloud environment dedicated to one customer), or as a Skan AI hosted cloud service that follows Skan AI’s security program.

What governance, compliance and auditability do agent claims decisions require?

The NAIC Model Bulletin states that AI-supported insurer decisions must meet applicable insurance laws and regulations, that carriers should retain reviewable records of agent actions, and that actions must not violate unfair claims settlement practices acts (state laws setting standards for fair claim handling) “regardless of the methods the Insurer used.”

Twenty-nine jurisdictions have adopted the NAIC Model Bulletin on AI (NAIC, 2026), which expects a written AI program covering claim administration and, where appropriate and available, vendor audit rights. For health payers, the U.S. Department of Health and Human Services (HHS) applies HIPAA’s minimum necessary standard (the rule limiting data use to what a task requires) to claims adjudication and lists a claims-processing third-party administrator as a business associate example (an outside firm handling protected health data for a payer).

Skan AI’s control layer records each input, output, action, decision, and rationale. Under privacy by design, Skan PRISM applies pixel-level redaction inside the customer environment, raw screen data stays within the customer perimeter, and email content and keystrokes are not captured.

What STP rates and ROI can carriers realistically expect?

Straight-through processing (STP, claims resolved end to end without manual handling) targets belong on simple, low-severity claims; complex claims stay human-led. Boston Consulting Group (BCG) estimates that fully automated redesign can resolve up to 70% of simple claims in real time (BCG, 2025).

BCG puts the potential reduction in loss adjustment expense (LAE, the cost of investigating and settling claims) from agentic AI-orchestrated claims at roughly 20-30% (BCG, 2026). Measure cycle time, cost per claim, and leakage (payments above what was owed) against a baseline observed with Skan Intelligence.

Claims agent projects underperform because agents built from written SOPs and partial event logs learn the clean path and can miss the exceptions adjusters handle daily. Grounding agents in work observed by Skan AI Process Intelligence, and revising AOPs when adjusters overturn an agent’s recommendation, narrows that gap.

Line-of-business differences

Agent scope shifts with claim type, documents, and regulation:

Should carriers build, buy or partner, and how should rollout proceed?

Start with one high-volume, low-severity claim type, prove the agent against a human baseline, then widen scope. Building in-house fits carriers with engineering capacity and documented processes; buying or partnering depends heavily on contract terms. Gartner advises setting deliverables, knowledge transfer, intellectual property rights, and exit responsibilities before contracting vendor forward-deployed engineering for agentic AI (Gartner, 2026).

Skan AI Blueprint identifies candidate claim types, Skan AI Process Intelligence observes those identified areas in depth and maps them into the Context Graph of Work, and Skan AI Agents execute under Agentic Operating Procedures (AOPs).

To see this applied to your claims workflow, request a Skan AI demo. In claims, widen scope from one claim type to adjacent types and lines, then the full operation.