OCC Sets New Expectations for Banks' Use of AI

The framework makes clear that regulators are less concerned with how impressive a bank's AI technology may be than with what the bank allows it to do, what could go wrong, and whether anyone can step in before it does.

The OCC's message is straightforward: The more authority banks give AI, the more important it becomes to demonstrate that appropriate safeguards, meaningful human oversight, and effective risk management controls are in place.

The framework does not impose new regulatory requirements, and the OCC states it is an examiner resource rather than a bank self-assessment tool. Much of the framework is written in permissive terms – "may" and "could" – and is described as supporting examiners' risk-based judgment. Nonetheless, it offers banks something they should take seriously: a preview of the questions examiners may soon be asking. It also fills a notable gap. The revised interagency model risk management guidance issued in April (OCC Bulletin 2026-13) expressly excludes generative and agentic AI, making this framework the primary supervisory reference for those systems.

Not All AI Is Created Equal

Perhaps the most important takeaway from the OCC's new framework is that AI risk is not necessarily determined by how sophisticated the technology is, but by what it is allowed to do and what happens when something goes wrong.

An AI tool that summarizes internal documents presents a very different risk profile from an AI agent authorized to initiate payments, approve loans, release fraud holds, or interact directly with customers. The OCC's framework establishes four illustrative risk tiers, ranging from low-risk internal productivity tools (Tier 4) to critical applications such as autonomous payment agents, real-time credit decisioning, and fraud-hold release (Tier 1). Notably, a simple agent authorized to move money or take direct action can rank higher than a sophisticated model used only for internal drafting, regardless of how well it performs in testing. As the stakes increase, so do the supervisory considerations. Higher-risk applications may warrant stronger testing, tighter restrictions on autonomy, more meaningful human oversight, and greater operational safeguards.

Importantly, the framework does not establish new regulatory requirements or create a separate category of AI risk. Instead, it explains how AI can introduce or amplify risks that banks are already expected to manage, including operational, cybersecurity, privacy, credit, and compliance risks. In other words, banks do not necessarily need to reinvent their risk management programs. But they must be prepared to demonstrate that those programs adequately address how AI is actually being used in their organization. And while the framework applies to OCC-supervised institutions, state-chartered banks should expect the Federal Reserve and FDIC to take note.

Six Pillars, Six Questions Banks Should Be Asking

The OCC organizes its framework around six pillars: Purpose; Risk; Control; Oversight; Resilience; and Assurance. While not designed as a self-assessment tool, the pillars offer a useful lens for anticipating what examiners will look for:

• Purpose: Do we know where AI is being used, what actions it is authorized to do, and what systems and information it can access?
   
• Risk: Have we assessed the potential consequences of an AI error, including financial losses, customer harm, and regulatory violations, and classified that risk based on such consequences?
   
• Control: Are appropriate safeguards in place to prevent unauthorized actions, inaccurate outputs, data exposure, or misuse?
   
• Oversight: Who is accountable for each AI application, and do the people overseeing it have the knowledge and authority to intervene?
   
• Resilience: Can we detect when an AI system is behaving improperly, stop it before additional harm occurs, and continue critical operations?
   
• Assurance: Can we demonstrate through testing, documentation, monitoring, and independent review that our controls actually work and operate effectively?

These are not merely theoretical concerns. The OCC specifically addresses risks involving prompt injections, hallucinations, data leakage, model manipulation, vendor oversight, and AI agents operating with excessive permissions.

One particularly important point: An AI system does not have to be hacked or even malfunction in the traditional sense to create a serious incident. An agent may remain fully operational while making unauthorized decisions, exposing customer information, or executing transactions it should never have been permitted to make. That distinction has significant implications for cybersecurity controls, incident response plans, vendor agreements, and business continuity procedures.

What Banks Should Do Now

Banks should not wait for their next examination to find out whether their AI governance programs measure up. Several practical steps deserve attention now:

1. Develop an AI inventory and purpose statements: Inventory AI applications, including capabilities embedded in existing vendor products. Identify their owners, authorized purposes, access permissions, and degree of autonomy. Document what each application is and isn't permitted to do, because the framework treats activity outside an approved purpose as a control failure even when the result looks reasonable. Don't overlook employee use of unapproved AI tools or AI functionality quietly added through vendor updates. Notably, according to the OCC, activity outside the approved purpose may still be treated as a control failure even when the output appears reasonable;
    
2. Classify AI applications by risk: Distinguish low-risk productivity tools from applications that affect customers, move money, access sensitive information, or make consequential decisions. Apply controls proportionate to the potential harm, not simply the sophistication of the technology;
    
3. Revisit human oversight: Having a human "in the loop" means little if that person cannot meaningfully evaluate or stop an AI decision. For higher-risk uses, ensure reviewer competence, and establish clear approval requirements, escalation procedures, and mechanisms to suspend automated activity;
    
4. Take another look at vendor agreements: Determine what AI capabilities vendors are deploying, how they use and retain bank information, and whether contracts adequately address notice of model changes, security, AI-specific incident notification, audit rights, exit planning, and responsibility for AI-related errors or losses. Responsibility for the risk remains with the bank even when a vendor hosts the system, and traditional technology agreements may not adequately address these risks;
    
5. Update cybersecurity and incident response planning: Consider threats unique to AI, including prompt injection, manipulation of AI agents, excessive agent permissions, and unauthorized use of connected systems. Govern agent identities and credentials as you would privileged users. Test whether the bank can identify and contain harmful AI activity, including through tested kill switches and fallbacks, in situations where the system is functioning but making the wrong decisions; and
    
6. Document, test, and be ready to explain: Ensure governance decisions, risk assessments, testing, monitoring, and material changes are documented. Because generative AI does not always produce the same output twice, a single successful test proves little; consider repeated testing, adversarial testing, and red teaming. Banks should be able to explain not only why an AI application was approved, but also how they know its controls remain effective.

The Bottom Line: Don't Let AI Get Ahead of Your Controls

The OCC's framework is not a warning against innovation. In fact, the agency expressly recognizes AI's potential to improve efficiency, strengthen fraud detection, and enhance banking services. But it is a reminder that the ability to automate a decision does not eliminate the responsibility for that decision.

Banks that are already using AI should evaluate whether their existing policies, vendor contracts, risk assessments, and incident response procedures adequately address the OCC's supervisory considerations. Those considering more autonomous applications should make that assessment before deployment, not after something goes wrong.

Experienced legal counsel can play an important role in translating the OCC's framework into practical, defensible governance measures, particularly where AI intersects with financial services regulation, consumer protection, privacy, cybersecurity, and third-party risk management. Baker Donelson has a dedicated Financial Services Privacy, Cybersecurity, and Technology Team that regularly works with financial institutions to assess AI-related legal and regulatory risks, develop governance programs, negotiate technology agreements, and prepare for and respond to cybersecurity and other technology-related incidents. If you have any questions concerning these issues or need any assistance, please contact the authors, Matthew G. White, AIGP, CIPP/US, CIPP/E, CIPT, CIPM, PCIP and Alexander F. Koskey, CIPP/US, CIPP/E, PCIP, the leaders of Baker Donelson's Financial Services Privacy, Cybersecurity, and Technology Team.

And remember: AI may be making more decisions for banks. But when regulators come asking questions, the bank will still need to have the answers.