Banks preparing for exams under the Office of the Comptroller of the Currency’s new artificial intelligence framework can start by documenting what their AI tools are allowed to do, what customer data they can reach and who can shut them down. In a Tuesday (Nov. 17) analysis, Polsinelli attorney Alexander D. Boyd urged banks to review those controls alongside vendor contracts and incident response plans, particularly where AI agents can change account records or execute transactions.
The OCC’s Nov. 4 bulletin gives examiners a framework for evaluating generative and agentic AI, technologies not included in the agencies’ revised model risk management guidance in April. It covers an AI system’s purpose, its risks, and banks’ control, oversight, resilience and assurance.
The OCC says the framework creates no new regulatory requirements or supervisory expectations and calls for examiners to tailor their reviews to the materiality of each use case. It applies to national banks, federal savings associations and federal branches and agencies of foreign banks. Polsinelli sees the guide as a way for banks to assess whether their existing controls hold up as AI takes on more authority.
Polsinelli recommends cataloging AI already in use and tools under consideration, including features embedded in third-party products. For each use, banks should document the authorized purpose, the data and systems the tool can access, the actions it can take on its own and the decisions that require human approval. The firm also recommends setting approval procedures for new deployments and material changes to existing ones.
Banks can then weigh the expected benefits against the potential for customer harm. An AI agent that can move money or access sensitive records may warrant tighter permissions, more extensive testing and human signoff for consequential actions. Polsinelli advises banks to assign clear responsibility for AI activity and monitor systems after deployment.
Privacy reviews should trace how a tool accesses, processes and retains customer information and whether it can reproduce or disclose that information in its outputs. Security teams should test whether existing controls can withstand AI-specific threats, including attempts to manipulate an agent through instructions embedded in the content it reads. They should also check whether agents have more access than they need and whether models can expose sensitive data.
An existing vendor can add AI features, switch underlying models or introduce agents that interact with bank systems under a contract that predates those capabilities. Polsinelli recommends updating due diligence to identify the models and providers involved, understand how they handle bank and customer data, and review their security controls.
Depending on the risk, banks should consider an artificial intelligence addendum that defines permitted uses, limits the use of customer information for model training and addresses downstream providers. The firm also recommends provisions covering notice of material changes and incidents, cooperation with investigations, audit rights and liability for errors or unauthorized actions that harm customers.
Polsinelli advises banks to supplement existing incident plans with procedures for identifying, reporting internally, escalating, investigating and remediating AI incidents. An agent could send inaccurate information to customers or execute an unauthorized transaction without triggering conventional alerts for an intrusion or outage. Response procedures should also cover unauthorized AI use, privacy violations, harmful outputs and compromised models.
Banks should test whether they can suspend an AI agent, revoke its access and preserve the prompts and activity logs needed to reconstruct what happened. They should also determine how to reverse unauthorized actions where possible, remedy customer harm and keep critical services running through fallback processes. Polsinelli recommends working those scenarios into tabletop exercises and business continuity tests. An exercise can reveal whether staff have the access and authority to contain an agent’s actions while keeping the underlying banking service available.