Federal banking regulators aren't the only ones taking a closer look at artificial intelligence. State regulators have their own playbook, and financial institutions, including banks and nonbanks like consumer lenders, mortgage companies, and money transmitters, should proactively pay attention, as you may see these questions in your next examination.
In September 2026, the Conference of State Bank Supervisors (CSBS) released its Artificial Intelligence Supervisory Framework, giving state examiners a roadmap for evaluating how banks and nonbank financial institutions use and manage artificial intelligence (AI). The framework addresses everything from AI governance and vendor oversight to consumer protection, cybersecurity, and increasingly autonomous AI systems.
The guidance preceded the Office of the Comptroller of the Currency's (OCC) November 4, 2026, release of its own AI supervisory framework, which we discussed in our recent article, OCC Sets New Expectations for Banks' Use of AI. Although the two frameworks differ in structure and scope, their messages are remarkably consistent.
Regulators want to know:
- where AI is being used;
- what risks it creates;
- who is responsible for overseeing it; and
- whether the institution can demonstrate that it has appropriate controls and that they are working.
Neither framework creates new legal obligations. The CSBS guidance is discretionary, and individual state regulators may decide whether and how to incorporate it into their examination programs. But that doesn't mean financial institutions should ignore it. When regulators publish the questions their examiners may ask, institutions would be well served to have the answers ready.
First Question: Do You Know Where AI Is Being Used?
One of the most revealing aspects of the CSBS framework is where it begins. Before diving into sophisticated models or emerging technologies, examiners are instructed to determine whether the institution knows where AI is being used in the first place. That may sound simple. It often isn't.
AI capabilities are increasingly embedded in software applications already used for lending, fraud detection, customer service, compliance, document processing, and other routine operations. A vendor may introduce AI functionality through a software update without the institution ever purchasing something specifically marketed as an AI product (and sometimes without appropriate stakeholders at the institution being informed). Employees may also be using generative AI tools outside approved channels. See, for example, our prior alert on so-called "Shadow AI."
In fact, the CSBS framework expressly directs examiners to consider vendor inventories, software inventories, approved tools, and recent platform changes when an institution reports that it does not use AI or is uncertain about its use. In other words, asserting (or not knowing) that your institution uses AI may itself invite additional questions.
The framework also encourages examiners to determine whether institutions maintain an inventory of AI applications, identify their business purposes and owners, and evaluate their risks. Importantly, those risks should be assessed based on how the technology is actually used. As an example, an AI tool that helps draft internal communications presents a very different risk from one that influences lending decisions, handles sensitive customer information, or takes actions without human approval.
Your Vendor's AI Is Still Your Problem
For many financial institutions, the greatest AI exposure may not come from technology they develop themselves. Rather, it may come from the vendors they already rely upon. The CSBS framework specifically addresses AI embedded in third-party products and services. Examiners may review vendor due diligence, contractual arrangements, performance monitoring, change management, and contingency planning.
This raises several important questions:
- Does the institution know which vendors use AI?
- What customer or confidential information can those systems access?
- Can vendors retain that information, share it, or use it to train their models?
- What happens when a vendor changes its AI capabilities or experiences an AI-related security incident?
These are not questions that traditional technology agreements necessarily answer. In fact, the framework expressly identifies vendor contracts and service terms addressing the use, retention, sharing, and training of AI systems on customer or institutional data as materials examiners may need to request.
Financial institutions should therefore take another look at their existing vendor relationships, particularly those involving sensitive information or consequential decisions. A contract negotiated years before AI functionality was introduced may not adequately address the risks the institution faces today. And outsourcing technology does not mean outsourcing responsibility for managing the associated legal, regulatory, and operational risks.
AI Doesn't Get a Pass on Consumer Protection
AI may change how financial institutions make decisions, but it doesn't eliminate the legal obligations that govern those decisions. The CSBS framework directs examiners to consider existing consumer protection requirements when AI affects customer communications, lending decisions, servicing, or other consumer-facing activities. Areas of potential review include, for example, fair lending, adverse action notices, accuracy, explainability, privacy, and unfair, deceptive, or abusive acts or practices (UDAAP).
Consider an AI-powered underwriting tool that recommends denying a loan. Can the institution explain the decision and provide legally sufficient adverse action reasons? Can it identify whether the tool produces discriminatory outcomes or uses biased training data? Does someone within the institution understand the model's limitations, or is the institution simply accepting the tool's results (as a policy, or in practice)?
Similar concerns arise when AI generates customer communications, responds to complaints, or provides financial information. The framework specifically contemplates examiner requests for samples of AI-generated or AI-assisted customer communications, chatbot transcripts, notices, and other consumer-facing materials.
The lesson is straightforward: Financial institutions remain responsible for complying with applicable consumer protection laws, regardless of whether a human or an algorithm produces the decision or communication.
When AI Starts Acting on Its Own, the Stakes Get Higher
Like the OCC's more recent guidance, the CSBS framework pays particular attention to generative AI and emerging agentic AI systems. Unlike traditional tools that merely analyze information or generate recommendations, agentic AI can take actions with limited human direction. Depending on its permissions, an AI agent might access internal systems, communicate with customers, initiate workflows, or execute transactions. This introduces risks that conventional cybersecurity and model risk management programs may not fully address.
The CSBS framework specifically identifies hallucinations, inaccurate outputs, prompt injection, and data exposure as areas for examiner consideration. For systems capable of acting autonomously, it goes further, directing examiners to evaluate permission boundaries, human checkpoints, logging, reversibility, and the ability to restrict or halt the system.
Those last considerations are particularly important. An AI agent does not necessarily have to be compromised to cause significant harm. It may function exactly as designed while taking an action that should never have been authorized. Institutions deploying these technologies should be able to explain what their AI systems are permitted to do, what information and systems they can access, when human approval is required, and how harmful or unauthorized activity can be stopped.
Simply having a human somewhere in the process is not enough if that person lacks the information, authority, or opportunity to intervene.
What Financial Institutions Should Do Before Their Next Exam
The CSBS framework offers a useful opportunity for banks and nonbanks to evaluate their existing AI governance programs before regulators begin asking questions. Six practical steps deserve attention:
- Know where AI is being used. Develop or update an inventory of AI applications, including employee-facing tools and AI capabilities embedded in vendor products. Identify business owners, purposes, data access, and whether each application supports or makes decisions.
- Evaluate and document the risks. Establish a risk-based process for reviewing AI use cases. Consider customer impact, sensitive information, financial exposure, operational dependencies, and the consequences of inaccurate or unauthorized actions. Apply controls proportionate to those risks.
- Review vendor agreements and data practices. Determine how vendors use, retain, share, and potentially train AI models on institutional information. This may require updating your due diligence requirements to evaluate whether contracts adequately address security, oversight, changes to AI functionality, incident notification, liability, and termination rights.
- Strengthen oversight of generative and agentic AI. Address hallucinations, prompt injection, data leakage, privacy controls, and excessive permissions. Establish meaningful human review, appropriate restrictions on autonomous actions, and procedures for detecting and stopping harmful activity (aka a "kill-switch").
- Don't overlook consumer protection. Review AI applications involved in lending, servicing, customer communications, and other consumer-facing activities for compliance with applicable laws. Ensure consequential decisions can be explained and that AI-generated communications are appropriately reviewed. Give particular attention to customer-facing AI chatbots, which have drawn both regulatory scrutiny and new regulation.
- Be prepared to show your work. Review AI policies, governance records, risk assessments, training materials, testing results, and management reporting. The framework includes a document request list, giving institutions a useful preview of what examiners may want to see.
Importantly, not every institution needs the same AI governance program. The CSBS framework expressly calls for examinations to be proportionate to an institution's size, complexity, risk profile, and AI use. A community bank experimenting with internal productivity tools should not necessarily face the same level of scrutiny as a financial services company deploying autonomous systems to make consequential customer decisions.
But institutions of every size should understand their AI exposure and be prepared to explain how they manage it.
The Bottom Line: Different Regulators, Same Message
The CSBS and OCC frameworks are not identical, and neither imposes new mandatory requirements. Together, however, they provide an increasingly clear picture of how regulators are approaching AI oversight.
Financial institutions do not necessarily need to build entirely new compliance programs from scratch. Many of the relevant risks already fall within established cybersecurity, privacy, vendor management, model risk, operational risk, and consumer protection programs. The challenge is determining whether those programs adequately address AI's distinctive capabilities and risks.
For state-regulated banks and nonbanks, the CSBS framework provides a particularly useful starting point. Institutions should consider reviewing their AI governance programs against the framework's examination questions and document requests, identifying gaps, and addressing them before their next examination.
Experienced legal counsel can assist institutions in assessing their AI-related regulatory obligations, developing defensible governance programs, reviewing vendor agreements, and addressing the cybersecurity, privacy, and consumer protection risks associated with emerging AI technologies. Baker Donelson's Financial Services Privacy, Cybersecurity, and Technology Team regularly advises banks and other financial institutions on AI governance, regulatory compliance, technology transactions, cybersecurity, and incident response. If you have questions about the CSBS or OCC frameworks, or need assistance evaluating or strengthening your institution's AI program, please contact the authors, Matt White, AIGP, CIPP/US, CIPP/E, CIPT, CIPM, PCIP and Alex Koskey, CIPP/US, CIPP/E, PCIP, leaders of the Team.
The bottom line? Regulators aren't expecting financial institutions to have all the answers about AI. But they are increasingly expecting them to know what questions to ask – and to have an appropriate plan for managing the risks.