The Bad-Faith SIU Rule: Factual Cause vs. Statistical Suspicion
Under state Unfair Claims Settlement Practices Acts (NAIC Model #900), referring an insured or claimant to a Special Investigation Unit (SIU) or delaying payment based on a black-box probability score—without individual, objective evidentiary proof of material misrepresentation—constitutes an unfair claims practice and exposes the carrier to bad-faith punitive damages.
Enterprise Fraud Detection in the AI Era
Paris- and Boston-based Shift Technology is one of the most widely deployed insurtech vendors in the property, casualty, and health claims sectors. Reporting hundreds of millions of processed claims across dozens of global carriers, Shift positions its AI as a core automation layer for detecting fraud networks, scoring claim suspicion, and streamlining straight-through claims payouts.
However, as state insurance commissioners increase oversight under the NAIC Model Bulletin on AI Systems and state anti-bias statutes like Colorado SB 21-169, enterprise fraud models face unprecedented evidentiary demands.
The Three Regulatory Vulnerabilities in AI Fraud Scoring
Our audit of predictive fraud scoring models across property and casualty carriers highlights three structural failure modes that create severe statutory liability for insurers:
| Operational Mechanism | Underlying Algorithmic Risk | Governing Statute & Regulatory Standard |
|---|---|---|
| Social Network & Association Graphs | Ingesting claimant relational graphs, phone contacts, and shared addresses can penalize individuals living in densely populated or multigenerational households. | Disparate impact based on national origin and familial status (Colorado SB 21-169; C.R.S. § 10-3-1104.9). |
| SIU Referral Gatekeeping | Automating SIU referrals when an opaque composite score crosses an arbitrary threshold (e.g., “Risk Index > 75”). | Unreasonable investigation delay and harassment claims under state UCSPA (e.g., Tex. Ins. Code § 542.003). |
| Explainability Deficits in Deposition | When examiners are asked in bad-faith depositions why a claim was withheld from payment, they cite the software output without personal knowledge of the model's weights. | Inadmissibility under Rule 702 (Daubert/Frye) and waiver of the “genuine dispute doctrine.” |
The Deposition Hazard: Why Fraud Scores Collapse in Court
In bad-faith insurance litigation, the core legal question is whether the insurer had an “arguable, reasonable basis” for delaying or denying payment. When an adjuster test-fires an automated tool like Shift:
“Q: Why did you delay paying Mr. Rodriguez's legitimate water damage invoice for 90 days?
A: The Shift AI flagged the claim with an elevated fraud index of 82.
Q: What specific factual facts in Mr. Rodriguez's invoice made it fraudulent?
A: I don't know the exact algorithms. The vendor's machine learning model looks at thousands of historical patterns.”
This testimony destroys the carrier's defense. A statistical correlation generated by a third-party vendor is not admissible evidence of individual fraud. State fraud statutes require specific proof of intentional misrepresentation of a material fact. Outsourcing suspicion to an algorithm leaves the insurer completely defenseless against statutory interest, attorney's fees, and punitive bad-faith awards.
Carrier Due Diligence Scorecard for Shift Deployments
To safely deploy Shift Technology or comparable automated fraud detection tools, carrier compliance officers and claims committees must enforce five non-negotiable diligence controls:
- Mandatory Disparity Audits: Require Shift to provide certified annual disparate impact reports demonstrating that its models do not disproportionately flag protected demographic classes across auto, property, and workers' compensation lines.
- Zero Autonomous Referral: Prohibit the software from automatically freezing claim payments or routing files to SIU without an independent human examiner confirming objective red flags.
- Document-Level Provenance: Ensure that every alert generated by the system cites a specific conflicting document or date discrepancy rather than an abstracted statistical “score.”
- Cross-Carrier Data Custody Review: Clarify whether the vendor uses your policyholders' claim history to train shared fraud models that benefit competing carriers.
- Contractual Regulatory Indemnity: Require the vendor to indemnify the carrier against regulatory penalties arising from algorithmic bias or model governance failures under state DOIs.