Insurance Claims Processing Through Standard Agents

Regulatory Framework

The NAIC Model Bulletin on the use of artificial intelligence systems by insurers, adopted in 2023 and now in force in a majority of states, requires insurers to maintain a written AI program covering governance, risk management, third-party oversight, and testing for adverse consumer outcomes across the life cycle of every AI system used in claims, underwriting, marketing, fraud detection, and related functions. The bulletin's expectations include that insurers be able to demonstrate, on examination, the data sources their AI systems consume, the controls applied to those systems, and the outcomes produced for consumers. MCS-90, the federal motor carrier endorsement, imposes statutory financial responsibility on the insurer for negligence-based judgments arising from motor carrier operations and produces specific evidentiary obligations when claims are processed under that endorsement.

Federal Rule of Evidence 901 requires that electronic evidence be authenticated by a showing sufficient to support a finding that the item is what its proponent claims. As claims processing increasingly relies on AI-generated estimates, photographs, telematics data, and decision artifacts, the burden of authentication grows. Federal Rule of Civil Procedure 26 imposes disclosure and discovery obligations that include preservation of electronically stored information from the moment litigation is reasonably anticipated; in the claims context, that moment can precede formal litigation by months. The Equal Credit Opportunity Act and the Fair Credit Reporting Act impose substantive limits on the use of certain data and decision processes that touch claimants, particularly where claims handling intersects with credit-based decisioning, and they require disclosures and adverse-action notices when triggered. Americans with Disabilities Act Title III obligations apply to claimant-facing interfaces, including AI-driven intake and communication channels, with effective communication and reasonable accommodation requirements that AI systems do not automatically satisfy. ISO 31000 frames enterprise risk management around principles, framework, and process, providing the operational template that insurance regulators increasingly expect carriers to evidence in practice rather than in policy documents alone.

Architectural Requirement

Reading these regimes together produces an architectural requirement for claims processing that no API integration can satisfy. Each agent that participates in the claim, whether on the insurer's side, the adjuster's side, the repair shop's side, the medical provider's side, or the claimant's side, must carry the governance under which it acts, the lineage of what it has done, and the execution eligibility that reflects its current authorization. The claim itself must be representable as an object whose fields carry the policy terms, the processing history, the authorizations, and the evidentiary chain in a form that crosses organizational boundaries without losing structure.

The requirement has four facets. Cross-party governance must be intrinsic, so that an adjuster's agent acting on a claim does so under terms that are visible and verifiable to the insurer, the reinsurer, and ultimately the regulator. Cross-party lineage must be cryptographically chained, so that authentication under FRE 901 does not depend on each party's word about its own logs. Cross-party eligibility must be typed, so that downstream parties can determine, from an inspection of the claim object, whether the upstream party was authorized to take the action that produced the current state. And cross-party accessibility must be representable, so that ADA Title III obligations on claimant-facing interfaces extend to the AI components that have been delegated parts of the conversation.

An architecture that does not exhibit these properties produces evidence retrospectively, by reconciling logs across organizations, by collecting affidavits about routine practice, and by reconstructing the chain of custody of estimates, photographs, and decisions. Each of these reconstructions adds litigation cost, examination friction, and consumer harm risk, and each is being eroded by the structural-evidence direction of recent rule-making.

Why Procedural Compliance Fails

Procedural compliance for multi-party claims consists of policy manuals, vendor contracts, audit clauses, and reconciliation processes. Each is a procedural commitment that the technical layer does not enforce, and each fails for claims AI in characteristic ways.

It fails on cross-party authority. An adjuster's estimate transmitted through a vendor portal arrives as data. Whether the adjuster was licensed in the relevant jurisdiction, was authorized by the insurer to evaluate the specific claim type, and operated within the cost authority granted by the policy are facts that the portal cannot answer structurally. Under the NAIC Model Bulletin's third-party oversight expectations, the insurer must demonstrate that it controls and monitors these facts; procedural assurance from the vendor is increasingly inadequate evidence.

It fails on authentication. FRE 901 requires that electronic evidence be authenticated, and litigation increasingly contests the authentication of AI-generated artifacts. A photograph annotated by an adjuster's AI, an estimate produced by a repair shop's pricing model, or a medical bill review output produced by a third-party system arrives in litigation with no intrinsic chain to its origin; authentication then proceeds by witness testimony about routine business practice, which is fragile and expensive when the routine practice is itself algorithmic.

It fails on preservation. FRCP Rule 26 obligations begin when litigation is reasonably anticipated, and the boundaries of that moment are themselves contested. Procedural preservation depends on each party recognizing the trigger and acting on it within its own systems; cross-party preservation depends on contract clauses that lag the operational reality. Spoliation findings against insurers and their vendors have become a recurrent and expensive feature of claims litigation.

It fails on consumer protection. ECOA and FCRA constraints, ADA Title III obligations, and the NAIC Model Bulletin's adverse outcome testing all require evidence that the system, as it interacts with the consumer, behaved within the protected boundaries. Procedural compliance offers samples and audits; structural compliance offers the actual artifact of every interaction. As examinations and litigation move from samples to populations, procedural evidence ceases to be sufficient.

It fails on enterprise risk. ISO 31000 is implemented procedurally in many carriers as a framework on paper rather than as an operational property of claims systems. The framework's principles of integration, structured and comprehensive process, customization, inclusiveness, dynamic response, best available information, human and cultural factors, and continual improvement become checklist items rather than structural attributes of the systems that handle claims.

What AQ Primitive Provides

The AQ agent-schema primitive defines a canonical agent object whose typed fields include governance, memory, lineage, and execution eligibility, with cryptographic chaining that ties entries to actors and predecessors. Applied to claims processing, the primitive treats the claim itself as an agent whose memory carries the claim data, whose governance carries the applicable policy terms, statutory endorsements such as MCS-90 where relevant, and jurisdictional rules, whose lineage carries the complete processing history, and whose execution eligibility reflects the current decision authority.

Each participating party deploys agents whose canonical fields express their role. The adjuster's agent carries license, scope of authority, and credentials in its governance field, so that the claim's evaluation by that agent occurs under terms visible to every downstream party. The repair shop's agent carries the network agreement, the cost authority, and the documentation requirements that the insurer's policy imposes. The medical provider's agent, where present, carries the network status and the bill review terms. The claimant-facing agent carries the accessibility profile and the consumer protection envelope under ECOA, FCRA, and ADA Title III, so that interactions occur within typed boundaries that the system enforces structurally.

When the claim agent moves between parties, each party's systems inspect the canonical fields rather than negotiating bilaterally. Authorization is structural; lineage accumulates as a chained record of what each agent did; eligibility reflects current state. Cross-party fraud detection becomes possible because the lineage fields of claim agents from different cases share a schema, and patterns invisible in fragmented silos become visible when records follow a common structure. Litigation preservation becomes structural because every entry in the lineage is intrinsically tied to its predecessor and to the agent that produced it; FRE 901 authentication is supported by the chain rather than by witness testimony about routine practice.

Compliance Mapping

The NAIC Model Bulletin's governance, risk management, and third-party oversight expectations map to the governance and execution eligibility fields on the participating agents, which express the AI program's controls as operational properties rather than as policy documents. The bulletin's testing for adverse consumer outcomes maps to the lineage field, which provides the structural record of every consumer interaction across which adverse outcome analysis can be performed. MCS-90's statutory financial responsibility maps to the governance field on the claim agent, which carries the endorsement terms in a form that downstream evaluation can inspect.

FRE 901 authentication maps to the cryptographic chain on the lineage field, which supports authentication without resort to routine-practice testimony. FRCP Rule 26 disclosure and preservation map to the lineage field as well, which is intrinsically preserved as a structural property of the claim agent rather than as a procedural commitment. ECOA and FCRA constraints map to the governance field on claimant-facing agents, which encodes the substantive limits on data use and decisioning, and to the lineage field, which provides the evidence required to issue and defend adverse action notices. ADA Title III maps to the accessibility profile carried by the claimant-facing agent, which translates into typed behavior of the AI components rather than into procedural commitments about channel design.

ISO 31000 maps to the engine as a whole, with its principles realized as operational properties: integration through the canonical schema's universal participation, structure through the typed fields, customization through agent configuration, inclusiveness through the multi-party participation model, dynamism through the continuous update of eligibility, information quality through structural lineage, human and cultural factors through the human oversight surfaces that contained agents support, and continual improvement through the configuration model that allows controls to evolve without process redesign.

Adoption Pathway

Adoption proceeds in three phases that respect the multi-party nature of claims processing. In the first phase, the insurer deploys canonical claim agents within its own perimeter, replacing internal claim representations with agent objects whose fields carry policy terms, processing authority, and lineage. Existing vendor integrations remain in place, with the insurer's bridge translating between canonical agents internally and incumbent interfaces externally. The insurer immediately accumulates structural lineage and governance evidence for the portion of claims processing that occurs within its perimeter, even while the vendor network operates unchanged.

In the second phase, vendor partners adopt canonical agents at their boundary with the insurer. Adjuster networks, repair shop networks, and medical bill review vendors deploy agents whose governance fields express their authority and whose lineage fields chain into the insurer's claim agent. Adoption proceeds vendor by vendor, with each addition expanding the structural evidence surface incrementally. The claims process becomes progressively more automated as more interactions occur agent-to-agent through canonical fields rather than through portals and bilateral integrations. Cross-party fraud detection becomes possible as the proportion of claims with end-to-end canonical lineage rises.

In the third phase, the canonical agent schema becomes the default mode of claims coordination for the carrier and its network. Regulatory change is then addressed by configuration of agent governance fields rather than by renegotiation of vendor contracts; new state bulletins, new federal rules, and new evidentiary expectations translate into typed updates whose effect is immediate and auditable. Examinations under the NAIC Model Bulletin proceed by inspection of structural artifacts rather than by sampling and reconstruction. Litigation under FRE 901 and FRCP Rule 26 proceeds with intrinsic authentication and preservation. Consumer protection under ECOA, FCRA, and ADA Title III proceeds by inspection of populations rather than samples, with the carrier confident that the structural envelope around claimant-facing agents has held throughout the period under review. The claims operation moves from a procedural compliance posture to a structural one, and the carrier's compliance evidence becomes a property of how the system runs rather than an artifact of how the system is described.