MVA White Collar Defense, Investigations & Regulatory Advice Blog

On September 16, 2026, the Conference of State Bank Supervisors (“CSBS”) released its new Artificial Intelligence Supervisory Framework, an examiner guide designed to help state regulators evaluate how financial institutions use artificial intelligence and assess related risks.

Key Takeaways

While the CSBS framework does not create new legal requirements or supervisory expectations and expressly refers to the framework as a discretionary supervisory tool, it does provide a roadmap for how state regulators are likely to approach artificial intelligence (AI) examinations going forward. Institutions should expect increasing scrutiny of:

• Enterprise-wide AI inventories and use-case mapping;
• Risk-based classification and approval processes;
• Board and management oversight of AI activities;
• Specific controls surrounding generative AI and LLM deployments;
• Vendor management practices addressing embedded and third-party AI;
• Ongoing monitoring, testing, and lifecycle governance; and
• Consumer protection and explainability considerations for AI-enabled decisions.

An Examination Framework Built Around Core Areas of Supervisory Focus

The framework directs examiners to focus on four primary areas: (1) initial AI scoping; (2) governance and oversight; (3) AI inventories and use cases; and (4) generative AI and emerging technologies. The framework also provides examiners with a document request list and direction that examiners consider other existing, applicable supervisory frameworks governing third party, model, consumer compliance, and operational risk reviews.

Examiners are encouraged to tailor supervision based on direct, and indirect, usage of AI

The CSBS framework emphasizes that examiners should tailor their review to the institution’s risk profile, size and complexity and that exams should reflect the nature and significance of its AI usage.  The framework begins with eight questions intended to aid examiners in assessing the scope of an institution’s AI usage, including exposure to AI through usage by its vendors.  The questions seek to determine the institution's AI footprint, governance framework, third-party AI exposure, use of generative AI, risk management practices, and handling of sensitive data through AI systems. Each question provides the examiner with a related prioritization of related supervisory areas of focus.

Governance Remains a Core Examination Pillar

Consistent with other recent regulatory initiatives, risk governance is a core pillar of the framework. Examiners are instructed to review whether institutions have established clear structures that define how AI use is approved, overseen, communicated, restricted, and updated. The framework repeatedly stresses the importance of identifying responsible owners for AI systems, ensuring senior management oversight, establishing clear approval and escalation processes, and ensuring that governance requirements extend to internally developed tools, vendor-provided systems, and embedded AI capabilities. Regulators also are encouraged to review board and management reporting regarding AI use, performance, incidents, and emerging risks.

Importantly, the framework takes a lifecycle approach to governance. The framework instructs examiners to consider whether institutions not only evaluate AI at inception, but also monitor systems over time as underlying models, data, vendors, business processes, and risk exposures evolve.

AI Inventories and Risk Classification Receive Significant Attention

One of the most practical messages from the framework is that regulators will place significant emphasis on understanding whether an institution knows where and how AI is being used throughout the enterprise. Examiners are directed to review whether institutions maintain a current inventory of AI systems, tools, models, and use cases, including internally developed solutions, third-party tools, and embedded AI functionality. Examiners are also directed to review any documentation describing the business purpose and role of AI for identified AI use cases.

Beyond merely identifying AI systems and documenting their business purpose, examiners are instructed to assess how institutions categorize AI use cases and evaluate them based on risk. The framework encourages institutions to distinguish between internal tools, customer-facing systems, decision-support applications, and decision-making technologies, while also assigning risk ratings based on factors such as consumer impact, human oversight, data sensitivity, and potential harm from errors. To support this effort, CSBS included an AI Use Case Risk-Tiering Worksheet that classifies AI uses of the factors into low, moderate, and high-risk categories. The risk tiering then aligns to suggested controls, with higher-risk use cases expected to need more extensive governance, testing, monitoring, validation, and senior-level oversight.

Generative AI and Agentic AI Receive Dedicated Treatment

Perhaps the most significant feature of the framework is its dedicated section addressing generative AI, large language models (“LLMs”), and emerging agentic AI capabilities. CSBS recognizes that there are unique risks presented by these technologies and therefore provides additional guidance for them.

Examiners are instructed to determine how institutions use generative AI, whether tools are publicly available, privately deployed, internally developed, or vendor-provided, and what controls govern the use of sensitive information. The framework specifically highlights risks associated with hallucinations, inaccurate outputs, prompt injection attacks, data leakage, overreliance on AI-generated content, and information security vulnerabilities.

The framework also addresses agentic AI systems, which CSBS describes as systems capable of taking actions with limited human direction. For these tools, examiners are directed to evaluate permission structures, human checkpoints, logging capabilities, reversibility of actions, and an institution’s ability to restrict or shut down automated activity when necessary.

Third-Party Risk Management and Consumer Protection Remain Central Themes

A recurring message throughout the framework is that institutions remain responsible for AI-related risks even when capabilities are provided through vendors. CSBS repeatedly references existing third-party risk management guidance and directs examiners to evaluate AI-specific due diligence, contractual protections, vendor monitoring, change-management practices, and contingency planning.

Similarly, the framework emphasizes that AI use does not excuse institutions from existing consumer protection obligations. Examiners are instructed to evaluate how AI influences consumer-facing communications, underwriting, pricing, servicing, collections, fraud review, and other activities affecting consumer outcomes. Particular attention is directed to explainability, adverse action notices, fair lending concerns, proxy discrimination risks, and the accuracy of AI-generated communications. Notably, the framework reminds examiners that AI-assisted decision-making cannot undermine legal requirements to provide specific reasons for adverse actions where required by law.

The Framework Reflects Broader Regulatory Trends in AI Risk Management

The framework reflect an emerging supervisory consensus that existing risk management guidance (whether for use by examiners or institutions) may not fully cover the risks presented by novel AI technologies, particularly generative and agentic AI. In April, the federal banking regulators carved generative and agentic AI out of their revised Model Risk Management guidance, noting that banks should establish governance and controls over this technology as part of their general risk management practices. Prior to the CSBS’ release of the framework, the Financial Stability Board (FSB) released a consultative report outlining practices for effectively governing and managing the development and deployment of AI. While the CSBS framework and FSB consultative report target, at least superficially, different actors – bank examiners and supervised institutions, respectively – each stresses the novelty and rapid adoption of AI usage and presents a largely similar approach, including an emphasis on tailoring the level of supervision and controls based on the materiality of activities covered by, and the risks posed by, individual use cases.  Both the CSBS framework and the FSB consultative report acknowledge the substantial benefits provided from the use of AI but seek to promote an institution’s responsible use of AI through appropriate governance structures, risk controls and oversight. State legislatures are also actively placing parameters around use of AI, with requirements in California and Colorado around use of automated decision-making technology, including in financial services related decisions, scheduled to become effective in January.

Conclusion

The CSBS framework is notable because it reflects how regulators increasingly view some AI risk through existing safety and soundness, consumer protection, third-party risk management, and model risk management principles rather than through an entirely new regulatory regime. Also, because it relies heavily on sources like the NIST Artificial Intelligence Risk Management Framework, financial institutions that have already begun implementing AI governance programs will recognize many of the framework’s concepts, including AI inventories, risk tiering, governance oversight, vendor management, and lifecycle monitoring. At the same time, the framework provides insight into the questions examiners are likely to ask during future examinations.

For state chartered financial institutions, the CSBS framework is best viewed as an examination blueprint. Organizations that can identify their AI footprint, demonstrate risk-based governance, maintain effective oversight of vendor-provided technologies, and implement meaningful controls around generative AI will likely be best positioned as regulators continue to incorporate AI oversight into routine supervisory activities. 

Finally, while formal risk management expectations have not been officially set, the groundwork for them is being established. Both the framework and the consultative report explicitly disclaim that they establish risk management requirements for banks.  The reality, however, is that institutions that lack the reflected governance structures and risk governance tools will likely be subject to enhanced scrutiny of the effectiveness of their risk programs.