AI Regulatory Compliance for Banks & Credit Unions Ahead of Exams

For the past couple of years, AI in financial services has been mostly a policy conversation inside the institution. That is starting to change. Regulators are putting together the materials examiners will use to ask banks and credit unions how they govern AI, and institutions that have not yet mapped their AI use may find the questions arrive before the answers are ready.

The clearest example so far is the Artificial Intelligence Supervisory Framework from the Conference of State Bank Supervisors. It offers a look at what examiners are being told to focus on, and it is a useful checklist for any institution that wants to get ahead of an exam.

AI Regulatory Guidance Is Taking Shape for Banks and Credit Unions

The CSBS framework was approved by two CSBS committees in August and released publicly in September. It is some of the first hard documentation that examiners can use as a basis for their assessments, and it ties into multiple AI guidance frameworks. The package includes a core examiner guide, an examiner work program with suggested controls and exam questions, a risk tiering worksheet, supplements for nonbanks and a list of source materials.

A few points on scope. The framework is written for state examiners of banks and nonbank institutions, it is discretionary, and each state agency decides how much of it to use. The guide also states that it does not create new legal obligations. Credit unions answer to their own regulators, but the framework draws on sources that are not specific to any charter, including the NIST AI Risk Management Framework, the Cyber Risk Institute Financial Services AI Risk Management Framework and the Treasury AI Lexicon. Credit unions may find it a helpful signal of where AI supervision is heading. Institutions can also use it to assess their own AI programs and prepare for examinations.

How Examiners Will Scope Your AI Use

The core guide opens with eight scoping questions. They are meant to establish whether the institution uses AI at all, and the guide warns that a negative or unclear answer may be checked against the vendor inventory, software inventory, approved tools and recent platform changes. In other words, saying you do not use AI is not the end of the conversation.

The questions cover several themes.

• Whether AI is used in products, services, operations, compliance or internal support functions.

• Whether the institution has identified its AI systems, tools, models and use cases.

• Whether AI touches customer-facing activity or supports decisions.

• Whether the institution relies on vendors for AI, and has looked for AI embedded in vendor products.

• Whether generative AI or large language models are in use.

• Whether use cases are differentiated by risk, and whether sensitive or customer information passes through AI tools.

Each answer routes the examiner toward a section of the framework, so the way an institution answers these early questions shapes the depth of the review that follows.

What Examiners May Ask an Institution to Produce

The document request list shows what examiners may ask for once AI use is identified. It groups requests into governance and oversight, AI inventory and use cases, generative AI and emerging use, and additional materials.

• Governance. AI policies, standards and procedures, committee or oversight materials, management and board reporting, monitoring dashboards, and employee guidance or training.

• Inventory. A list of AI systems, tools, models or use cases, business purposes and ownership, risk ratings or tiering, vendor products with known or potential embedded AI, and change management documentation.

• Generative AI. Policies and restrictions, a list of approved or known tools, and documentation of data controls and output review.

• Additional materials. Vendor oversight materials, model validation or monitoring records, consumer-facing compliance materials, samples of AI-generated customer content, contract terms on data use and training, and internal audit or control testing reports.

Many institutions could assemble parts of this list today. The inventory and the vendor items tend to be the harder ones.

AI Governance Framework Expectations Start With Inventory and Oversight

The governance section asks examiners to determine who is responsible for AI oversight, how use cases are evaluated for risk and approved, how information reaches management and the board, and how AI risks are documented and updated over time. It also asks whether governance covers internally built tools, externally sourced tools and AI embedded in third-party products.

That last point runs through the whole framework. Embedded vendor AI shows up in scoping, in the inventory procedures and in the governance procedures. Vendor products are one of the most common places an institution ends up using AI without having formally decided to, which is why examiners are asked to look for it.

Third-Party AI, Model Risk and Consumer Protection

The framework does not try to replace existing supervisory resources. It routes AI questions into third-party and vendor review, model risk review, consumer protection review and operational risk review. The nonbank supplements then add AI-specific considerations to each.

For vendor AI, the supplement centers on whether the institution treats it as a managed dependency rather than a black box. Considerations include AI-specific due diligence, contract terms on data use and model updates, notification of material vendor changes, ongoing monitoring and contingency planning. For model risk, the supplement asks about intended use, limitations, drift, validation, human challenge and vendor opacity. For consumer protection, it covers adverse action reasons, fair lending, UDAP and UDAAP, proxy data risks and whether human review is meaningful.

The guide notes that some existing model risk resources may not fully address generative or agentic AI, so examiners may use the framework to identify AI-specific considerations that warrant added review.

Risk Tiering Is Central to an AI Risk Management Framework

The risk tiering worksheet gives institutions and examiners a shared way to think about proportionality. It scores each use case on factors such as consumer impact, human oversight, harm potential and data sensitivity, and the highest-rated factor generally sets the preliminary tier. Controls build as the tier rises, so a higher tier carries the controls of the lower ones plus additional monitoring, independent validation, AI-specific incident response and reporting to senior management or the board. The worksheet also points to reassessing a use case when it gains more decision-making authority.

An institution that has already tiered its AI use cases will find the rest of the framework much easier to answer.

Generative and Agentic AI, and Where Logging Fits

One section of the framework looks ahead to newer tools. The generative AI and emerging use section names agentic tools directly, and procedure GE-8 covers AI systems that can take actions with limited human direction. Examiners are asked to review how the institution defines the actions the system is permitted to take, the human checkpoints, logging, reversibility, and the ability to restrict or halt the system.

Logging is one item in that list, and it is worth planning early even though many institutions have not reached it yet. Traditional access logs and change tickets show which account acted and whether a system change was approved. They do not show the instruction an agent was given or the model version behind it. Institutions that are only beginning with agents can take a simple step now by deciding what a record of agent activity should show and who owns it. The framework does not prescribe log fields, so each institution will need to define what fits its own risk tier.

Questions to Ask Before an Exam

The areas of review in the framework translate into questions an institution can use to test its own readiness.

• Do we have an inventory of AI systems, tools and use cases, including AI embedded in vendor products?

• Who is accountable for AI oversight, and what reaches management and the board?

• Have we tiered our AI use cases by risk, and do controls scale with the tier?

• Do vendor contracts and due diligence address AI-specific issues such as data use and model updates?

• Do we have policies for generative AI, including data controls and output review?

• For any AI that acts with limited human direction, have we defined permitted actions, checkpoints, logging and a way to halt it?

Where Institutions Can Start

The CSBS framework is a useful place to begin because it shows how examiners are being asked to approach AI. Building the inventory, assigning ownership and tiering use cases cover much of what the early sections ask for, and the later sections become easier once that foundation is in place.

For financial institutions working through these questions, Compass IT Compliance helps teams build the oversight that AI adoption calls for, including AI governance aligned to the NIST AI Risk Management Framework, independent AI risk assessments and audits that produce evidence for regulators and stakeholders, and AI-specific incident response planning. Organizations that would like a thoughtful outside perspective on their approach are welcome to start a conversation with our team.