OCC’s AI Examiner Framework: How Banks’ AI Use Can Trigger Legal Liability and Supervisory Exposure
What You Need to Know
Key takeaway #1
Banks’ AI failures involving lending, payments, customer communications, or anti-money laundering (AML) compliance can expose banks to liability.
Key takeaway #2
Institutions should assess AI tools based on the data and systems these tools can access, the actions they can take, and the harm they could cause.
Key takeaway #3
The Framework highlights the importance of meaningful human oversight and documentation of AI use.
Client Alert | 8 min read | 11.16.26
On November 4, 2026, the Office of the Comptroller of the Currency (OCC) issued Bulletin 2026-54 and a Supervisory Framework on Bank Use of Super Intelligence (the “Framework”), an examiner resource addressing banks’ use of artificial intelligence (AI), particularly generative AI and AI agents that can act on their own. The Framework creates no new legal requirements or supervisory expectations, and the OCC states that banks are not expected to use it to assess their own AI use or risk management. Importantly, it identifies as higher risk several AI uses common for banks that are already governed by binding law, such as moving money, making credit decisions, and communicating with customers. A failure in any of those uses could trigger legal liability. Institutions that rely on AI for these tasks should consider the following:
- Higher-risk AI failures may implicate laws already binding on the bank. Using AI does not change the bank’s existing legal obligations or the standards for a violation. Errors in AI-enabled customer communications or AML alert disposition may nevertheless result in breaches of applicable consumer protection laws or Bank Secrecy Act requirements.
- Productivity-enhancing tools may warrant higher-risk treatment. AI coding assistants and compliance tools can affect customer accounts or production systems. Their risk depends on what data they can access and change and their role in decisions.
- Banks may need to substantiate the human oversight on which assessments of their AI use cases rely. Approval and intervention records can help show whether oversight was effective. Purpose statements and system activity records may be needed to assess whether an agent stayed within its authority.
- AI control weaknesses may draw supervisory attention without necessarily warranting a matter requiring attention (MRA). Examiners may raise governance or control concerns, but whether a weakness warrants an MRA depends on the applicable agency’s standards for legal violations and safety-and-soundness deficiencies. A departure from the Framework alone does not establish grounds for an MRA.
Structure and Risk Tiering
The Framework — which employs the Administration’s preferred phrase “super intelligence,” or SI, to refer to AI — organizes 16 components under six pillars: purpose, risk, control, oversight, resilience, and assurance. It contemplates more intensive controls where the potential consequences of an agent’s failure or misuse are greater and provides illustrative risk-tier definitions. And it applies to generative and agentic AI, which the OCC’s prior April 2026 interagency model risk management guidance expressly excluded.
The examples below are illustrative, not mandatory classifications. The OCC has instructed examiners to tailor reviews to the bank’s size, complexity, and risk profile and the nature and consequences of its AI uses.
“Human-in-the-loop” generally means review and approval before an output or action is used or executed; “human-on-the-loop” means active monitoring with authority and practical ability to intervene.
Selected Examination Considerations
The Framework raises several considerations about how familiar controls apply to AI-enabled workflows.
Engineering tools. The OCC lists code assistance for “internal code or data access changes” among its illustrative high-risk (Tier 2) examples, while general code assistance appears among the low-risk (Tier 4) examples. This focus is consistent with that of the New York State Department of Financial Services (NYDFS), which noted in a May 2026 AI advisory that secure programming practices may include additional testing, validation, and human oversight of AI-generated code before production deployment. A tool’s classification should reflect its permissions and potential effect on production systems, rather than its label as a coding assistant or productivity tool.
Customer communications and compliance operations. Tools affecting customer communications, AML alert disposition, or fraud-hold release may warrant higher-risk treatment. Banks may wish to assess whether the AI merely assists an employee or effectively determines or executes the action, and whether any human review provides a meaningful opportunity to prevent an error before the action is taken.
Purpose limits and human review. Activity an AI system undertakes outside its approved purpose “may be treated as a control failure even when the resulting output or action appears reasonable,” under the Framework. Therefore, institutions should record the scope of the system’s authority and the actions it took and consider whether logs of what the system did can be compared against what it was permitted to do. Where a risk assessment relies on human review, institutions should ensure that human roles are not reduced to “a nominal click.”
External agents. The Framework distinguishes agents supplied by the bank’s providers from customer and other external agents outside the bank’s control. Institutions should assess how they identify those agents, establish the scope of their authority, limit transactions, and revoke access. Banks are taking different approaches to customer-directed agents, including restricting credential-based access and revising customer agreements to address agent activity. For customer agents initiating payments, access controls and the resulting records also bear on whether a customer authorized a disputed transfer and how the bank must handle the dispute.
Customer agents and Regulation E. Regulation E, which governs electronic fund transfers, may apply to consumer payments initiated through AI agents if the account and transfer fall within Regulation E’s scope. How Regulation E’s existing authorization and liability rules apply to particular AI-agent arrangements raises interpretive questions. Giving an agent access does not by itself resolve whether a disputed transfer was authorized or who bears liability; the customer’s instructions, access method, agent operator, and circumstances of the transfer matter. Banks should assess access arrangements, dispute procedures, and contractual terms against applicable liability, error-resolution, and Electronic Fund Transfer Act (EFTA) anti-waiver protections. Commercial payments require separate analysis.
Existing Legal Obligations and Supervisory Exposure
The Framework creates no new enforceable standards, so departing from its illustrative practices does not, by itself, establish a legal violation or grounds for an MRA. Banks’ underlying activities remain subject to existing consumer protection, fair-lending, Bank Secrecy Act, and payment requirements.
Under the OCC and the Federal Deposit Insurance Corporation’s new MRA standards, effective November 2, 2026, an MRA requires an actual violation of banking or banking-related law or regulation, or an imprudent practice meeting the rule’s material-financial-risk criteria. The possibility of a future legal violation does not alone suffice for an MRA, although the financial-risk criteria may support an MRA before a violation or loss occurs. The Federal Reserve has addressed its examiners’ approach to MRAs separately, in a statement of supervisory operating principles.
For banks, the practical question is what an identified AI weakness represents: a departure from an illustrative practice, an actual legal violation, or an imprudent practice meeting the financial-risk criteria. That distinction should guide the bank’s response to examination findings and its remediation priorities.
State-chartered banks and state-law requirements. The Framework is written for OCC examiners overseeing national banks, federal savings associations, and federal branches and agencies, but state-chartered banks face similar underlying legal and operational exposures. Reuters reported in October that all three federal banking agencies were asking about banks’ AI controls, and the Conference of State Bank Supervisors released its own AI supervisory framework in September, which each state agency may decide whether and how to use. State requirements may apply across charter types depending on the entity, activity, exemptions, and applicable preemption principles. NYDFS’s cybersecurity regulation applies to covered entities subject to exemptions. Its October 2024 and May 2026 AI letters address cybersecurity risks but do not themselves impose new requirements. Colorado S.B. 26-189 adds notice, human-review, and record-retention requirements beginning January 1, 2027, subject to its scope, exemptions, and alternative-compliance provisions. State-law coverage and federal preemption remain provision- and activity-specific; nonbank affiliates and agents do not automatically inherit National Bank Act protection.
Practical Considerations
Institutions may wish to consider the following options within their existing governance programs, with the scope and intensity of review proportionate to the significance of their AI uses:
- Use-case inventories. Record who can invoke each system, the system’s access to data and tools, permitted actions, human checkpoints, and associated legal obligations.
- Testing and human oversight. Test the detection and suspension of inappropriate activity, continuity arrangements, and reviewer capacity. Assess independence and correlated-failure risks where secondary agents or judge models provide a backstop. Keep internal audit’s assurance role distinct from management’s testing and validation.
- Incident handling and access controls. Include AI-related failures in incident-response procedures and assess whether they trigger existing notification obligations. For example, an agent that improperly alters account records or disrupts payment processing may require notification under OCC Part 53 if the incident meets the rule’s computer-security and material-disruption criteria.
- Vendor arrangements. When providers add AI features, review change notices, information rights, responsibility for errors, and exit rights.
- Recovery arrangements. Test whether the bank can identify affected customers and transactions, correct erroneous actions, and provide appropriate remediation following an AI failure. Before restarting the affected AI functionality, verify that the cause has been addressed and the relevant controls operate effectively.
- Existing governance processes. Identify where existing policies and oversight processes address material AI uses and where adjustments are needed. Board and management reporting should identify those uses, significant incidents and customer harm, and the status of corrective action. Institutions can do so within existing risk categories and risk-appetite statements. The Framework does not require a separate AI risk category or risk appetite.
The Framework creates no new legal requirements. Banks remain responsible for complying with applicable law when using AI. Institutions may wish to confirm that their controls and records address the legal risks of AI use and support their position in an examination or customer dispute.
Please contact the authors if you have questions.
Contacts
Insights
Client Alert | 8 min read | 11.16.26
OCC’s AI Examiner Framework: How Banks’ AI Use Can Trigger Legal Liability and Supervisory Exposure
On November 4, 2026, the Office of the Comptroller of the Currency (OCC) issued Bulletin 2026-54 and a Supervisory Framework on Bank Use of Super Intelligence (the “Framework”), an examiner resource addressing banks’ use of artificial intelligence (AI), particularly generative AI and AI agents that can act on their own. The Framework creates no new legal requirements or supervisory expectations, and the OCC states that banks are not expected to use it to assess their own AI use or risk management. Importantly, it identifies as higher risk several AI uses common for banks that are already governed by binding law, such as moving money, making credit decisions, and communicating with customers. A failure in any of those uses could trigger legal liability. Institutions that rely on AI for these tasks should consider the following:
Client Alert | 3 min read | 11.13.26
Massachusetts Trade Secret Preliminary Injunctions: Prove the Harm, Tailor the Relief
Client Alert | 1 min read | 11.12.26
SBA Announces Doubling Veteran Contracting Target to 10% for FY 2027, but Questions Remain
Client Alert | 6 min read | 11.12.26


