The Non-Delegation Doctrine
Under Section 1 of the NAIC Model Bulletin on Artificial Intelligence Systems, an insurance carrier remains directly and non-delegably liable for any statutory violations, unfair settlement practices, or unlicensed adjusting conduct caused by third-party vendor software.
Market Overview: The AI Claims Landscape Under Scrutiny
As insurance carriers and TPAs adopt artificial intelligence for intake, document classification, medical bill review, and demand package evaluation, procurement teams face a critical compliance gap: general software vendor questionnaires do not evaluate insurance-specific statutory liabilities.
Below is an objective regulatory analysis examining the operational architectures, public representations, and compliance boundaries of prominent market participants.
1. EvenUp (Plaintiff Demand Generation)
Primary Function: Generates personal injury demand packages, medical chronologies, and claimed damages for plaintiff law firms.
| Regulatory Dimension | Operational Reality & Carrier Scrutiny |
|---|---|
| Evidentiary Accuracy & Hallucinations | Public reporting and whistleblower disclosures in late 2024 revealed that automated demand generators can introduce phantom injuries, misread medical ICD codes, or duplicate chiropractic charges. Insurance examiners receiving AI-generated demands must maintain rigorous verification controls against original provider invoices. |
| Algorithmic Negotiation Anchor | Plaintiff AI platforms use proprietary settlement databases to generate elevated demand anchors. Insurers facing EvenUp demands risk bad faith if they rely on black-box counter-algorithms instead of verifiable, page-by-page document reconciliations. |
| Legal Responsibility | EvenUp disclaims legal liability in its terms of service, placing 100% of ethical and legal responsibility on submitting attorneys. Insurers must similarly ensure defense adjusters do not accept asserted summaries without primary-source substantiation. |
2. Stream Claims (stream.claims)
Primary Function: Workers' compensation and casualty document summarization, medical chronology generation, and workflow triage.
| Regulatory Dimension | Operational Reality & Carrier Scrutiny |
|---|---|
| Licensing & Discretionary Adjusting | To remain compliant with state adjuster licensing statutes (e.g., Cal. Ins. Code § 14021), platforms summarizing workers' comp files must operate strictly as clerical assistance. Any automated recommendation regarding disability ratings, causality, or settlement reserves triggers statutory adjusting oversight. |
| Data Privacy & Cloud Transmission | Workers' comp records contain detailed occupational health records, psychiatric evaluations, and Social Security numbers. Enterprise diligence requires verifying whether claimant data is isolated or transmitted to multi-tenant third-party LLMs. |
| Explainability Standards | When adjusters use AI chronologies to justify indemnity reserve changes, the chronology must provide direct visual links to the underlying doctor's narrative to survive state workers' compensation board audits. |
3. Legora (Legal AI Drafting & Prediction)
Primary Function: AI assistant for law firms and legal teams, providing drafting, demand analysis, and case valuation prediction.
| Regulatory Dimension | Operational Reality & Carrier Scrutiny |
|---|---|
| Unauthorized Practice of Law (UPL) Boundaries | Like many legal-tech providers, Legora explicitly disclaims that it is a law firm or provides legal advice. Both plaintiff and defense firms utilizing predictive valuation engines must ensure that licensed attorneys independently formulate legal theories and valuations. |
| Black-Box Settlement Predictions | Using predictive AI to recommend settlement values can violate state UCSPA statutes if used by insurers to justify lowball offers without factual, case-specific evidence. |
4. Shift Technology (Fraud Detection & Claims Automation)
Primary Function: Enterprise AI for insurance fraud detection, subrogation, and automated claims handling.
| Regulatory Dimension | Operational Reality & Carrier Scrutiny |
|---|---|
| Algorithmic Bias & Colorado SB 21-169 | Fraud scoring algorithms that disproportionately flag claims from specific demographic or socioeconomic groups face severe scrutiny under state anti-bias laws. Carriers must require vendors to provide empirical bias audits and validation studies. |
| The Right to Explanation | If a claim is investigated or denied based on a fraud score, state insurance departments require the insurer to articulate the factual basis. Carriers cannot rely on an unexplainable “fraud likelihood score” without documentary proof. |
Comparative Compliance Scorecard for Insurers
The following scorecard outlines the key statutory questions carrier compliance committees must require every claims AI vendor to answer:
| Vendor Diligence Requirement | Statutory Basis | Compliant Architectural Standard |
|---|---|---|
| Adjuster Discretion | State Licensing Codes (Cal. § 14021, Tex. § 4101) | Zero automated valuations or denial mandates. Software assists with evidence extraction; licensed adjusters retain 100% authority. |
| Explanatory Provenance | NAIC Model #900 & State UCSPA | Every single date, dollar figure, and diagnostic code links directly to an underlying Bates-numbered document page. |
| Algorithmic Fairness | Colorado SB 21-169 & NY DFS Circular 7 | No black-box regional multipliers or demographic discounting models. Transparent, reproducible factual reconciliation. |
| Data Custody & Privacy | NAIC Model #668 & HIPAA Privacy | Zero data egress to multi-tenant public APIs. Processing remains inside single-tenant, isolated infrastructure with zero model retention. |
Conclusion: The Defensible Standard
As state insurance commissioners expand market conduct examinations into claims artificial intelligence, the era of unscrutinized black-box software is ending. Carriers that align their technology procurement with transparent, source-linked, examiner-controlled standards protect themselves against catastrophic regulatory penalties, voided releases, and bad-faith liability.