AI Governance Is Becoming an Examination Issue

AI has moved quickly from experimentation into the operating fabric of financial institutions. AI-powered models support credit decisions, financial crime detection and risk management. Generative AI tools summarize documents, draft communications and assist employees. AI capabilities increasingly arrive embedded within third-party software. A new generation of agentic systems can retrieve information, invoke tools and take actions across applications and workflows.

Regulatory examination is beginning to catch up. Recently, the Conference of State Bank Supervisors (“CSBS”) released its AI Supervisory Framework, giving state examiners a common approach for assessing how financial institutions use and govern AI.¹ The framework applies to state-chartered banks and state-licensed nonbanks, including mortgage, money services and consumer finance companies. While it creates no new legal obligations and individual states retain discretion over how it is used, its potential reach is significant, as state regulators supervise 79% of U.S. banks.²

More important than the framework’s reach is its structure. CSBS has translated broad AI governance principles into practical examination procedures, including scoping questions, document requests, detailed work programs and risk-tiering criteria.

For financial institutions, the implication is clear. They need to know where AI is being used, who is accountable for it, what risks it creates, how those risks are controlled and whether they can produce evidence that those controls operate as intended and are appropriately governed. Financial institutions should prepare for a more evidence-driven standard of AI governance that extends from inventories and third-party tools to generative AI, agentic systems and the outcomes they produce.

The Examination Lens Is Becoming More Concrete

The CSBS framework is organized around a Core Examiner Guide, a detailed Examiner Work Program, supplements addressing nonbank AI risk, an optional AI Use Case Risk Tiering Worksheet and supporting supervisory materials. Together, these resources give examiners a structured path from initial scoping through deeper reviews of governance, third-party risk, model risk, consumer protection and operational risk.

An examination may begin with relatively basic questions such as:

Is AI being used?
Is there an inventory?
Does AI interact with consumers or support decisions?
Is AI embedded within vendor products?
Are generative AI systems being used with sensitive data?
How does the institution ingest feedback on model performance?
Is the pre-production design, test and challenge story for an AI model well-documented and sufficiently rigorous?

From there, examiners may request AI policies, board reporting, risk-tiering documentation, lists of approved generative AI tools and examples of customer-facing outputs such as chatbot transcripts. Governance procedures focus on accountability, risk approval, testing, monitoring and third-party AI. Separate procedures address risks specific to generative and agentic AI, including hallucination, prompt injection, data exposure, permissions, logging and the ability to halt a system.

One important feature of the framework is how it approaches incomplete visibility. An institution claiming that it does not use AI may lead an examiner to review vendor inventories, software inventories, approved tools and recent platform changes before accepting that conclusion. For many institutions, this means that AI discovery itself is becoming part of examination readiness.

The AI Inventory Needs to Reflect the Real Enterprise

Many financial institutions already maintain model inventories or lists of sanctioned AI tools. The emerging supervisory approach requires a broader view. A defensible AI inventory should account for internally developed AI, generative AI applications and agents, AI embedded within third-party platforms and business processes materially supported by AI. For each material use case, institutions should understand its owner, purpose, underlying data, risk profile and applicable controls.

Embedded AI may prove particularly challenging. A business unit may procure a customer service, underwriting, cybersecurity or productivity platform without thinking of it as an “AI system,” even though AI capabilities have been incorporated into the product. Governance processes that depend primarily on employees registering obvious AI projects can therefore miss meaningful sources of risk.

Inventories must also keep pace with change. Vendors add AI features, model providers introduce new capabilities, system permissions evolve and business teams find new uses for existing platforms. An inventory that is refreshed only periodically may quickly become incomplete.

The practical challenge is therefore broader than maintaining a spreadsheet of models. Institutions need processes for identifying AI as it enters the organization, connecting it to accountable owners and updating its risk classification as the technology or use case changes.

Risk Tiering Will Determine the Depth of Scrutiny

The CSBS framework also provides insight into how supervisory scrutiny may scale with risk. The framework’s optional worksheet places AI use cases into three tiers. Lower-risk applications generally involve internal uses, human-reviewed outputs, limited data sensitivity and relatively low potential harm. Moderate-risk applications may be consumer-facing, support decisions or rely on more limited human intervention. Higher-risk applications include those that directly influence consumer outcomes, use sensitive personal information, operate with limited human review or create significant operational dependence.

The expected controls increase accordingly. Across use cases, institutions should expect scrutiny of their AI inventory, acceptable-use policies, vendor arrangements, input and output quality controls and documented risk classification. Higher-risk applications may warrant independent validation, more extensive monitoring and direct oversight through formal AI governance structures.

This creates a useful discipline for management teams. Control requirements should reflect the actual risk presented by an AI use case. An internal productivity assistant does not necessarily require the same level of validation as an AI system influencing lending or other consumer outcomes. The institution should, however, be able to explain why each use case has been classified as it has and how the associated controls are proportionate to that risk.

AI Risk Extends Beyond the Model

Traditional model risk management provides an important foundation, but modern AI applications operate within a much broader technology ecosystem. Their behavior and risk can depend on enterprise data, cloud infrastructure, cybersecurity controls, third-party providers, application programming interfaces (“APIs”), foundation models and the applications through which users interact with them. As a result, effective AI governance increasingly requires coordination across technology, model risk, compliance, cybersecurity, legal, data governance, vendor management and the business.

Agentic AI makes these dependencies particularly important. As AI systems gain the ability to retrieve information, invoke tools and execute actions, institutions need to understand what those systems are permitted to do. They also need visibility into where human approval is required, what actions are logged, whether an action can be reversed and how the system can be stopped if something goes wrong.

Those questions can become especially significant when AI moves from generating information to acting on it. A poorly grounded response from a chatbot presents one category of risk. An autonomous system that changes information, communicates with a customer or initiates a business process creates a very different risk profile. Governance frameworks need to evolve accordingly.

Documentation Is Becoming a Control in Its Own Right

For many institutions, the greatest examination challenge may be connecting existing governance activities into a coherent, defensible record. A company may have an AI policy, a model validation process, cybersecurity controls, vendor diligence, business approvals and monitoring procedures. During an examination, however, those activities need to connect clearly to individual AI use cases and to the risks those systems create.

The CSBS framework points to several areas where gaps may surface: incomplete inventories, insufficient controls over generative AI data and outputs, limited logging of agentic AI activity, weaknesses in monitoring consumer outcomes and governance processes that exist on paper but cannot be demonstrated through testing records, approvals, audit evidence or board reporting. This elevates documentation from an administrative exercise to an important element of the control environment.

Institutions should be able to demonstrate ownership, risk assessment, approvals, testing and validation, human oversight, change management, monitoring and escalation. They should also be able to show how those activities continue after deployment as systems change and lessons from actual performance feed back into governance.

Five Questions Institutions Should Be Able to Answer

A practical readiness assessment can begin with five questions:

What AI do we have?
Establish an enterprise view that includes internally developed systems, generative AI, agents and AI embedded within third-party platforms.
Who owns it?
Identify accountable business, technology and risk owners for material use cases.
What could go wrong?
Assess potential consumer, operational, data, cybersecurity, model, compliance and third-party risks.
What controls are in place?
Connect those risks to approvals, testing, validation, access controls, human oversight, monitoring and escalation mechanisms.
Can we produce the evidence?
Create an examination record that ties policies and governance requirements to actual approvals, testing results, monitoring, exceptions and remediation.

One central theme emerges in the above questions: there is a core requirement for organizations to demonstrate governance and rigor in the design and setup of models as well as the post-launch maintenance, oversight and enrichment. Without evidence that demonstrates rigor in the design, testing and ongoing challenge of model inputs and outputs, organizations will be at a regulatory disadvantage.

The institutions best positioned for greater regulatory scrutiny will make these activities part of day-to-day AI management. A current inventory improves visibility. Risk tiering focuses control investment where it matters most. Testing and monitoring can identify problems before they become larger issues. Strong documentation makes accountability clearer and allows management to demonstrate that governance is operating as designed.

Broadly, these five questions provide a practical sequence for AI supervisory readiness: inventory the environment, assess the risks, test the controls, document the evidence and prepare to explain how the system operates in practice. As AI becomes more deeply embedded across financial services, the ability to produce that evidence will become an increasingly important measure of effective governance.

Footnotes:

1: Conference of State Bank Supervisors, “The CSBS Artificial Intelligence Supervisory Framework” (Sept. 16, 2026).

2: Conference of State Bank Supervisors, “2025 Annual Report” (Aug. 2026).

The views expressed herein are those of the author(s) and not necessarily the views of FTI Consulting, Inc., its management, its subsidiaries, its affiliates, or its other professionals.