The chief risk officer's view of AI claims investigation: model risk, controls, and concentration

A chief risk officer does not evaluate AI claims investigation on return on investment - the deal survives only if the tool fits inside the carrier's existing control framework. Every other seat at the table asks whether the tool creates value. The claims VP wants loss-ratio movement, the CFO wants payback, the SIU director wants a defensible workflow. The CRO asks a different question entirely: does this tool create un-owned residual risk, and can I get it inside my model-risk, third-party-risk, and control framework before it goes live.

That reframe changes how the whole procurement reads. The CRO rarely holds the budget chair and rarely champions the purchase. What the CRO holds is veto power through the enterprise-risk lens. A tool that cannot be inventoried, validated, and challenged does not get approved no matter how strong the loss-ratio case. This post walks the CRO's actual review: how the tool gets classified, the model-risk vocabulary a CRO borrows from banking, the regulatory perimeter for insurer AI, concentration and vendor risk, and the under-investigation exposure most risk functions are mispricing.

The core insight, stated up front: investigation-layer AI is a lower model-risk class than detection or pricing AI, because it produces evidence for a human adjudicator rather than making the pricing or coverage decision that the NAIC bulletin, NYDFS Circular Letter No. 7, and Colorado SB21-169 are aimed at. That distinction is what this post spends its length defending. For the seat next door - the compliance officer who files the antifraud plan - see the compliance officer's AI investigation deployment guide.

A chief risk officer classifies AI claims investigation as a model-risk and third-party-risk object that must fit an existing control framework before it can be approved. The classification is not about whether the tool works. It is about which risk categories the tool touches, which controls already govern those categories, and whether the residual risk after those controls is one the enterprise is willing to own.

This is where the CRO's evaluation diverges from every other buyer. Each seat in the buying center optimizes for something different, and the CRO's variable is the one that can stop a deal cold. A tool can clear loss-ratio, payback, and workflow review and still fail the risk review if it produces decisions no one can reconstruct. That is why a CRO reads a vendor pitch backward from the failure modes rather than forward from the value story.

The reframe matters because the backdrop is a live governance exposure in both directions. Per the Coalition Against Insurance Fraud, insurance fraud steals at least $308 billion every year from American consumers, and fraud is present in roughly 10% of P&C claims. A CRO who treats only the AI tool as the risk, and not the fraud the carrier is failing to investigate, is pricing one side of the ledger. The layered model - prevention, detection, investigation - matters here precisely because the risk profile differs by layer. To place this within the full committee, see how to map the buying center for AI investigation, which lays out where the veto seats sit; this post is the deep dive on the seat that blocks.

The Hesper-specific classification point is structural. Hesper does not add a new scoring or decisioning model. It adds an evidence-production and documentation layer downstream of a human adjudication - from fraud detection to fraud resolution. Every agent decision is logged with its sources, reasoning, and timestamps. Audit-trail-native by design is not a marketing line for the CRO; it is the property that lets an AI investigation decision enter a control framework at all.

Model risk is the risk of adverse consequences from decisions based on incorrect or misused model output. A CRO applies four controls to any model that enters the enterprise: a model inventory entry, independent validation, effective challenge, and documented governance. AI claims investigation belongs in the inventory and needs all four - but its risk class is lower than a pricing or underwriting model, and the reason is structural, not cosmetic.

The vocabulary here comes from banking. The Federal Reserve and OCC's SR 11-7, "Supervisory Guidance on Model Risk Management" (April 2011, since revised) defined the model inventory, independent model validation, and "effective challenge" concepts that insurance CROs now borrow. SR 11-7 is banking guidance, not binding on insurers, so a CRO uses it as the vocabulary source rather than the rulebook. But the four controls travel cleanly, and they are the frame through which an investigation agent gets read.

Why an investigation agent is a lower risk class

The distinction that lowers the class: a pricing or underwriting model makes a decision about who gets coverage and at what cost. An investigation agent makes no such decision. It runs evidence-gathering phases - 15+ of them in parallel on each flagged claim - and outputs a reviewable report. A human SIU lead adjudicates. The model's output is an input to a human decision, not the decision itself.

That design maps onto effective challenge by construction. Effective challenge, in the SR 11-7 sense, is the critical review of a model's output by informed, independent parties who can act on findings. When an SIU lead reviews every investigation package before any action - and can override the agent's reasoning - the effective challenge is not a bolted-on control. It is the operating model. The investigator's role shifts from execution to decision-making, which is exactly the human-in-the-loop posture a CRO wants to see documented.

The AI-governance perimeter for US insurers is set by three frameworks: the NAIC Model Bulletin (adopted December 2023, in roughly two dozen states since), NYDFS Insurance Circular Letter No. 7 (2024), and Colorado SB21-169 - all of which center on governance, third-party oversight, and testing for unfair discrimination. Their scrutiny concentrates on AI that influences underwriting, pricing, and eligibility. That is the strict end of the spectrum, and it is where the contrast with investigation AI gets drawn.

The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers requires a documented AI Systems Program: governance, risk management, and internal controls across the AI lifecycle, model validation and testing (including for unfair and proxy discrimination), third-party AI vendor and data oversight, and production of that documentation in market-conduct exams. Per state adoption trackers, roughly 24 states had adopted it or substantially similar guidance by early 2025, with Wisconsin the 24th in March 2025 and Alaska first in February 2024. This is the broadest of the three - it applies across AI uses, not only pricing.

NYDFS Insurance Circular Letter No. 7, finalized July 11, 2024, sets three principles - fairness, governance and risk management, and transparency. It requires board and senior-management oversight, model risk management with "independent review and effective challenge," and a three-step disproportionate-adverse-effect test: identify any disparate impact, establish a legitimate rationale, then search for a less discriminatory alternative, all documented and available to regulators. Critically, CL-7 governs underwriting and pricing. It is the strict boundary case, not a claims-investigation rule - but its effective-challenge and vendor-responsibility language is exactly what a CRO still wants satisfied downstream.

Colorado SB21-169 (signed July 6, 2021) requires insurers to test external consumer data, algorithms, and predictive models for unfair discrimination against protected classes and take corrective action. Colorado's Reg 10-1-1 (effective November 2023 for life insurers) operationalized it into a governance framework: catalogue and inventory models, evaluate data sources, remediate unfairly discriminatory outcomes, with explicit third-party vendor oversight duties. It is the model-inventory and vendor-oversight precedent now spreading across lines.

On the claims side, the documented-decision anchors are concrete. An audit-trail-native output maps to California 10 CCR 2698.36's documented-decision requirement and to the antifraud-plan filing obligations under NAIC Model Act 680, adopted in 48 states. When a market-conduct examiner pulls a Hesper-investigated case, the reconstructable record is already there. The evidence-chain substance behind that auditability is covered in the AI fraud investigation defensibility standard.

Concentration risk is the exposure created when a single vendor, model, or data source becomes a load-bearing dependency the carrier cannot quickly replace. It is a distinct risk from model risk, and it is the second thing a CRO checks. Both the NAIC Model Bulletin and NYDFS CL-7 are explicit that the insurer retains full responsibility for third-party tools and may not rely solely on a vendor's claim. The vendor's assurances do not transfer the risk.

So a CRO asks directly whether an AI investigation tool creates a single point of failure. The mitigants for investigation-layer AI are structurally stronger than for a decisioning model. First, the tool is complementary rather than a replacement for the detection stack - complementary to FRISS, Shift Technology, and Verisk, not a replacement - so adopting it does not swap out an existing dependency and can also run standalone. Second, the output is a portable, human-readable evidence record, not an opaque score the carrier cannot reconstruct if the vendor relationship ends. Third, the standard vendor-risk controls apply cleanly: contractual audit rights, documented exit and data-portability terms, verified security posture such as SOC 2, and confirmation the vendor does not train on the carrier's data.

The concentration story is why complementary positioning matters to the CRO specifically, not just to procurement. Because Hesper sits downstream of detection rather than replacing it, the detection dependency the carrier already governs stays where it is. The CRO is adding a documented layer, not concentrating more decisioning into fewer vendors. The CIO reviews the same surface from the integration and SOC 2 side; that review is laid out in the CIO checklist for AI investigation rollout.

Under-investigation is a regulatory and operational risk in its own right: the NAIC Unfair Claims Settlement Practices Act (Model #900) prohibits, among 14 enumerated practices, "failing to adopt and implement reasonable standards for the prompt investigation of claims" and "refusing to pay claims without conducting a reasonable investigation." A carrier that pays or denies flagged claims without adequate investigation carries exposure on top of the direct loss leakage from paying fraud.

This is the risk most CROs under-weight, because the status quo does not present as a risk - it presents as normal operations. Manual SIU teams typically investigate only about 25% of flagged claims; the rest are paid, denied without full work, or queued indefinitely. That 75% is an accepted but unpriced risk sitting in two places at once: the loss-leakage tail and the Unfair Claims Settlement Practices Act (Model #900) exposure. A risk function focused only on AI model risk can miss that the low-tech option is not the low-risk option.

Reframed as the CRO would see it, AI investigation is risk reduction, not only efficiency. Moving flagged-claim coverage from roughly 25% to 100% narrows both tails. The cycle-time and cost figures make the shift feasible at scale: a manual SIU investigation runs 14+ days per case at roughly $2,500, against 2-4 hours at roughly $150 per case, and throughput moves from about 10 investigations per investigator per month to 800+. The residual-risk math changes when full, uniform investigation of every flag becomes operationally possible.

There is a fairness dimension the CRO checks here too, and it points the same way. The disparate-impact scrutiny in CL-7 and SB21-169 is aimed at systems that decide coverage and price. A claims investigation agent operates downstream of those decisions. The residual concern a CRO tests is whether investigation is triggered and resourced consistently across claimants. Full, uniform coverage - the same 15+-phase investigation on every flag - is more consistent than selective manual investigation of roughly a quarter of flags. Consistency is itself a control.

A CRO can approve AI claims investigation once six controls are in place: a model-inventory entry, independent validation, an audit trail, human-in-the-loop adjudication, third-party and concentration mitigation, and documented regulatory mapping. The checklist turns the enterprise-risk veto into a set of pass conditions - and each maps to a property of the tool rather than a promise.

1. Model-inventory entry - the investigation agent is registered as a model with a defined owner, purpose, inputs, and outputs, classified at the appropriate (lower, evidence-producing) risk tier.
2. Independent validation - the agent's outputs can be independently reviewed and reproduced, because every phase is logged with sources, reasoning, and timestamps.
3. Audit trail - the output is reconstructable end to end, satisfying California 10 CCR 2698.36 and NAIC Model Act 680 documentation expectations and market-conduct exam production.
4. Human-in-the-loop adjudication - an SIU lead reviews and can override every investigation package before any action, which is effective challenge by construction.
5. Third-party and concentration mitigation - contractual audit rights, exit and data-portability terms, SOC 2 posture, no training on customer data, and a complementary (not replacement) footprint.
6. Documented regulatory mapping - the tool's scope is mapped against the NAIC bulletin, CL-7, SB21-169, and UCSPA Model #900, showing which apply squarely and how each is satisfied.

Cleared against those six, the enterprise-risk veto becomes an approval. The point a CRO ends on is the same one that started this post: the risk is bidirectional. The tool must fit the control framework, and the status quo of 25% coverage is itself a risk the framework should be pricing. The legal-risk cousin of this analysis - discovery, privilege, and admissibility of AI-produced evidence - is covered in the general counsel's view of AI fraud investigation and litigation.