Model Risk Questionnaires: The Paperwork Layer That Stalls Every AI Vendor Deal at Banks and Insurers

See also (wiki): wiki/model-risk-management.md, wiki/ai-vendor-contracts.md, wiki/vendor-security-questionnaires.md

Executive Summary

• Banks deploying AI must satisfy SR 11-7 model risk management requirements. The framework contains no standard questionnaire — each institution builds its own, typically 50–150 questions for a material AI model, drawing from the OCC Comptroller’s Handbook and now the U.S. Treasury’s new Financial Services AI Risk Management Framework (230 control objectives across seven categories, published February 19, 2026).
• Insurers face a parallel regime. The NAIC launched a 12-state AI Systems Evaluation Tool pilot in March 2026 (running through September 2026) with four structured exhibits covering AI usage inventory, governance, high-risk system detail, and data lineage. Adoption is expected at the NAIC fall meeting in November 2026.
• The common vendor failure mode is identical in both regimes: training data lineage. When an examiner or regulator asks “where did the data come from that trained this model, and do you have legal rights to use it?” — most AI vendors cannot produce a complete answer. Explainability documentation and bias-testing evidence are close seconds.
• The FS-ISAC (Financial Services Information Sharing and Analysis Center) published a two-tier generative AI vendor evaluation template: Level 1 (basic, ~15 questions for R&D/low-risk use) and Level 2 (comprehensive, covering legal, regulatory, model validation, and vendor moderation for production deployments).
• Time-to-complete for a full SR 11-7 model validation cycle on a material AI system runs 3–6 months at large banks and 6–12 months at mid-market institutions with smaller model risk management teams. The NAIC pilot is asking insurers to respond within the March–September 2026 window, effectively a six-month timeline for first-time AI governance documentation.

The SR 11-7 Questionnaire Problem

SR 11-7 does not prescribe a standard questionnaire. It prescribes three pillars — model development documentation, independent validation, and governance — and leaves implementation to each institution. The result: every bank builds its own model risk questionnaire, and every AI vendor answers a different version at every customer.

For traditional models (credit scorecards, interest rate models), this worked. The models were stable, the math was documented, and a competent model risk management (MRM) team could validate in weeks.

AI breaks this in five ways that the existing SR 11-7 corpus file (Pass 93) documents in detail. What matters for procurement is the questionnaire consequence: each of those five failure modes generates a new category of questions that vendors must answer, and most cannot.

What Examiners Actually Ask About AI

Based on the OCC Comptroller’s Handbook (August 2021) and the February 2026 Treasury FS AI Risk Management Framework, the questions cluster into seven categories:

The 230-Control-Objective Framework

The U.S. Treasury’s Financial Services AI Risk Management Framework, published February 19, 2026, is the most comprehensive AI questionnaire framework in U.S. financial regulation. Developed with 100+ financial institutions, the Financial Services Sector Coordinating Council (FSSCC), and the Cyber Risk Institute (CRI), it includes:

1. AI Adoption Stage Questionnaire — a maturity self-assessment that determines which controls apply
2. Risk and Control Matrix — 230 mapped control objectives across governance, data, model development, validation, monitoring, third-party risk, and consumer protection
3. Guidebook — implementation guidance
4. Control Objective Reference Guide — a 400+ page document with evidence examples for each control

The framework integrates with NIST Cybersecurity Framework, enterprise risk management, and SOC 2. For a mid-market bank ($1B–$50B assets) deploying its first production AI system, the 230 controls are the new floor — and most institutions have mapped fewer than 50 of them today.

Who Signs Off

At a bank, the sign-off chain for a material AI model typically runs: model developer → independent validation team (or third-party validator) → MRM committee → Chief Risk Officer → Board risk committee (for material models). The MRM committee meets monthly or quarterly. A single AI model that misses a committee cycle adds 30–90 days to the deployment timeline.

The NAIC Insurance Regime

The NAIC Model Bulletin on Use of AI by Insurers (adopted December 2023) established principles-based expectations for AI governance. In 2025–2026, the NAIC’s Big Data and Artificial Intelligence Working Group operationalized those principles through the AI Systems Evaluation Tool — a structured questionnaire regulators use during market conduct and financial examinations.

The Evaluation Tool Structure

The 12-State Pilot

The pilot launched March 2, 2026, across California, Colorado, Connecticut, Florida, Iowa, Louisiana, Maryland, Pennsylvania, Rhode Island, Vermont, Virginia, and Wisconsin. Monthly coordination calls among states will inform tool updates in September–October 2026, with formal adoption expected at the November 2026 NAIC fall meeting.

The NAIC is also developing a Third-Party Data and Models Working Group with a broad definition of “third party” encompassing any nongovernmental entity providing data, models, or outputs for insurance activities. A model law on third-party oversight is anticipated in 2026, potentially including vendor licensing requirements.

What Trips Up Insurers

The Monitaur analysis of the pilot identifies four friction points:

1. Data hygiene documentation complexity — Exhibit D requires detail on lineage and quality checks that most insurers have never assembled for their AI vendors
2. Vendor evaluation authority — Ambiguity about how far an insurer can audit its AI vendor’s underlying model
3. Jurisdictional inconsistency — States can modify questions, creating a patchwork compliance burden
4. AI vs. general data governance — Distinguishing AI-specific evaluation from existing data governance is harder than it sounds

The FS-ISAC Vendor Evaluation Template

The Financial Services Information Sharing and Analysis Center (FS-ISAC) published a Generative AI Vendor Evaluation and Qualitative Risk Assessment tool — the closest thing to a standardized AI vendor questionnaire in financial services. It uses a two-tier model:

Level 1 (lower-risk engagements — R&D, educational, sandboxed): Basic questions covering GenAI use scope, data privacy notices, foundation model identification, and information security. Approximately 15 core questions.

Level 2 (production deployments with business-process integration or confidential data): Comprehensive assessment adding legal and regulatory compliance, vulnerability management, model validation procedures, and vendor content moderation policies.

The Atlas Systems framework (2026) proposes a similar 15-question core across seven categories: AI usage and scope, data handling, model governance and explainability, security and access controls, compliance, operational resilience, and third-party dependencies.

Key Data Points

What This Means for Your Organization

The questionnaire burden is now the binding constraint on AI vendor deployment timelines in regulated industries — not the technology evaluation, not the budget approval, not even the security review. A bank or insurer that starts an AI vendor engagement without understanding the model risk questionnaire requirements will discover them 3–6 months into the process, after the business case has been approved and the vendor has been selected.

Three things to do before the next AI vendor conversation:

1. Map your questionnaire exposure. If you are a bank or bank service provider, the Treasury FS AI RMF’s 230 control objectives are your preparation checklist. If you are an insurer in one of the 12 pilot states, request a copy of the NAIC AI Systems Evaluation Tool exhibits and begin populating them now — before an examiner asks.

2. Pre-build your vendor’s answers. The most common failure mode is sending the questionnaire to the vendor and waiting. Most AI vendors have never completed a model risk questionnaire. Draft the answers yourself based on the vendor’s published documentation, then ask the vendor to confirm or correct. This cuts months from the cycle.

3. Budget for independent validation. SR 11-7 requires independent validation of any model used for material decisions. For AI, “independent” increasingly means a third-party firm with AI-specific validation expertise — not an internal team that validated credit scorecards last year. Budget $50K–$200K per material model depending on complexity.

If the questionnaire layer is where your next AI deployment is stuck — or where you suspect it will get stuck — that is a conversation worth having. Reach out at brandon@brandonsneider.com.

Sources

1. U.S. Federal Reserve, “Supervisory Letter SR 11-7: Guidance on Model Risk Management,” April 4, 2011. https://www.federalreserve.gov/supervisionreg/srletters/sr1107.htm. Credibility: HIGH — primary regulatory source.

2. U.S. Department of Treasury / FSSCC / Cyber Risk Institute, “Financial Services AI Risk Management Framework,” February 19, 2026. Via Lowenstein Sandler analysis: https://www.lowenstein.com/news-insights/publications/client-alerts/financial-services-ai-risk-management-framework-operationalizing-the-230-control-objectives-before-the-market-wakes-up-data-privacy. Credibility: HIGH — federal framework developed with 100+ institutions.

3. NAIC, “Model Bulletin on the Use of Artificial Intelligence Systems by Insurers,” December 2023. https://content.naic.org/sites/default/files/cmte-h-big-data-artificial-intelligence-wg-ai-model-bulletin.pdf.pdf. Credibility: HIGH — primary regulatory source.

4. Fenwick & West LLP, “NAIC Expands AI Systems Evaluation Tool Pilot Program to 12 States,” 2026. https://www.fenwick.com/insights/publications/naic-expands-ai-systems-evaluation-tool-pilot-program-to-12-states-key-updates-for-insurers-and-ai-vendors-supporting-insurers. Credibility: HIGH — Am Law 100 firm analysis of primary source.

5. GARP, “SR 11-7 in the Age of Agentic AI: Where the Framework Holds – and Where It Strains,” February 2026. https://www.garp.org/risk-intelligence/operational/sr-11-7-age-agentic-ai-260227. Credibility: HIGH — independent risk management professional association.

6. FS-ISAC, “Generative AI Vendor Evaluation & Qualitative Risk Assessment,” 2024. https://www.fsisac.com/hubfs/Knowledge/AI/FSISAC_GenerativeAI-VendorEvaluation&QualitativeRiskAssessment.pdf. Credibility: HIGH — financial services industry consortium.

7. Monitaur, “NAIC AI Systems Evaluation Tool Pilot: A Guide for Insurers,” 2026. https://www.monitaur.ai/blog-posts/naic-ai-systems-evaluation-tool-pilot-a-guide-for-insurers. Credibility: MEDIUM — AI governance vendor, but substantive analysis.

8. Plante Moran, “How the NAIC AI Model Bulletin Is Evolving,” March 2026. https://www.plantemoran.com/explore-our-thinking/insight/2026/03/how-the-naic-ai-model-bulletin-is-evolving. Credibility: MEDIUM — accounting/advisory firm analysis.

9. Atlas Systems, “AI Vendor Risk Assessment Questionnaire for Compliance (2026),” 2026. https://www.atlassystems.com/blog/ai-vendor-risk-questionnaire. Credibility: MEDIUM — vendor-published template, but representative of industry practice.

10. IBM, “Cost of a Data Breach Report,” 2025. Via Atlas Systems. Credibility: HIGH — annual benchmark study (vendor-published but widely cited as independent).

Brandon Sneider | brandon@brandonsneider.com April 2026