A Guth Labs publication ↗
Back to News

OCC's AI examination framework focuses on consequences and human control

AI-generated Conceptual illustration of human control over AI-assisted financial activity in a bank vault.
guthlabs.ai · AI-generated conceptual editorial illustration

The US Office of the Comptroller of the Currency issued a framework on November 4 to help examiners assess banks' use of AI in ordinary business activities. Published under the title “Super Intelligence: Supervisory Framework,” the bulletin emphasizes the consequences of a system's actions, its autonomy and the strength of human oversight. OCC bulletin

The agency is explicit about the document's status: it does not create new enforceable standards, prescriptive requirements or supervisory expectations for banks. Banks are not expected to use it as a mandatory self-assessment. It is an examiner resource to be tailored to the institution and its use cases. Scope of the framework

What a system can do matters

The framework's examples distinguish autonomous actions with financial consequences from lower-impact internal assistance. It suggests examining reversibility, access permissions, data sensitivity and the breadth of possible harm. Shared vendors or infrastructure can also connect risks across otherwise separate deployments. Supervisory framework

That approach makes the workflow a more useful unit of analysis than the model name alone. The same model could draft an internal explanation in one setting and initiate an external action in another. A capability benchmark does not describe the full risk created by the permissions, data and operating process surrounding it.

The bulletin says examiners should not treat AI as a separate risk category, and banks need not define a separate AI risk appetite. It also calls for proportionality in community-bank examinations. The term “Super Intelligence” is used broadly for technologies previously encompassed by artificial intelligence, not as a claim that every bank system has reached a new scientific threshold. Risk classification and terminology

Human review must be usable

The longer guide distinguishes approval before an action from monitoring with the ability to intervene. It asks whether reviewers have enough expertise, information and authority, and whether one person can realistically monitor the assigned volume of agents and alerts. It warns that nominal approvals or bulk processing can weaken review. Human-oversight discussion

For a bank assessing its own process, the practical question is what the reviewer can actually see and stop. A dashboard is useful only if the relevant evidence arrives in time, the person understands it and intervention is possible before an irreversible outcome. Those are operational questions rather than a count of approval buttons.

An evolving examination aid

The guide covers established areas including governance, vendor risk, privacy, cybersecurity, resilience, testing and auditability. It is intended to evolve as technology and supervisory experience change. Its sixteen components and illustrative risk tiers provide structure, rather than a universal implementation checklist. Framework structure

The immediate development is greater visibility into how examiners may think about AI-enabled banking activity. Institutions can use that visibility to improve the clarity of their own explanations: what a system is allowed to do, who is responsible, how failures are detected and what happens next. None of that turns this bulletin into a new rule or guarantees that a particular deployment is compliant.

Sources and citations

Revision history

Published .