The OCC Just Published the Only AI Risk Question That Matters: What Can It Do?
AI Governance | Agentic AI | Enterprise Risk Management | Banking Regulation
The most important question about artificial intelligence is no longer how intelligent the model is. It is what the system is authorized to do, what happens when it makes a mistake, and whether anyone can stop it before damage occurs.
That distinction matters as financial institutions move from AI systems that generate recommendations to autonomous agents that execute transactions, change customer records, release fraud holds, and influence credit decisions.
The Office of the Comptroller of the Currency's November 4, 2026 examiner framework, identified as OCC Bulletin 2026-54, offers a useful way to think about that distinction: Evaluate AI according to the consequences of its actions, not simply the sophistication of its technology.
For CIOs, CISOs, risk officers, and business executives, this is an important change in how AI governance should be approached.
An AI Agent Does Not Need to Be Intelligent to Be Dangerous
A model's complexity does not determine its business exposure. Permissions, autonomy, financial consequences, and the ability to reverse an action often matter more.
This is particularly relevant to agentic AI, where a system can select tools, access enterprise applications, make decisions, and execute actions with limited human intervention.
Governance Without Another Compliance Checklist
One important aspect of the OCC framework is that it does not establish a new set of enforceable regulatory requirements or a separate risk category for superintelligent systems. Instead, it provides a way to examine existing risks through six pillars: purpose, risk, control, oversight, resilience, and assurance.
Organizations do not necessarily need another AI governance framework. They need to apply existing risk management disciplines to the authority AI systems are being given. A bank may already have established policies for financial authorization, segregation of duties, operational resilience, change management, and third-party risk. The question is whether those controls remain effective when an AI agent performs activities employees previously handled.
Five Questions Every AI Agent Owner Should Answer
Before allowing an AI agent to operate in production, ask five questions.
Can the action be reversed? An incorrect report can be corrected. An executed payment, disclosed customer record, or regulatory filing may be difficult or impossible to reverse.
What permissions does the agent have? Read-only access carries a different risk from the authority to modify records, approve transactions, change permissions, or communicate externally.
What data can the agent access? Customer financial information, personal data, credentials, and confidential business information increase both security and privacy exposure.
What is the financial, legal, or regulatory consequence of failure? An error that creates an inconvenience is different from one that causes financial loss, customer harm, or a compliance violation.
Can a human reviewer realistically intervene? An approval button is not meaningful oversight if the reviewer lacks sufficient information, authority, expertise, or time to challenge the decision.
Apply these questions to the entire AI agent inventory, not only the most advanced models.
Human Oversight Must Be More Than a Checkbox
There is an important difference between human-in-the-loop and human-on-the-loop oversight.
Human-in-the-loop requires approval before executing a consequential action. Human-on-the-loop allows the system to operate while a person monitors its behavior and can intervene.
Neither approach is automatically effective.
If an AI agent processes thousands of transactions per hour, assigning a human to approve each one may create the appearance of control without meaningful review.
Effective oversight requires evidence: what the agent decided, which information it used, why the decision was approved or rejected, what exceptions occurred, and whether the agent could be suspended.
Who has the authority to shut the agent down, and how quickly can they do it?
Auditability and Third-Party Risk Cannot Be Afterthoughts
Traditional application logging may record the final transaction. Agentic AI can require substantially more information to reconstruct what happened.
That evidence may include the original prompt, retrieved data, model version, tool calls, permissions, configuration changes, human approvals, and overrides. Without those records, investigating a disputed decision becomes difficult. Third-party AI introduces another challenge. A provider may update its model, modify agent behavior, or introduce new capabilities without the customer making a corresponding application change.
Contractual change notifications, security assessments, testing, accountability, and exit planning therefore become part of the AI governance discussion.
Existing Frameworks Still Provide the Foundation
The OCC's consequence-based approach fits naturally with the NIST AI Risk Management Framework and ISO/IEC 42001.
Framework | Application to agentic AI |
NIST GOVERN | Establish accountability, ownership, authority limits, and risk acceptance |
NIST MAP | Identify the agent's purpose, operating context, permissions, and potential consequences |
NIST MEASURE | Evaluate exposure, test controls, and collect evidence of agent behavior |
NIST MANAGE | Apply risk treatments, monitor performance, and provide intervention or shutdown capabilities |
ISO/IEC 42001 Clause 6 | Plan AI risk assessment and treatment, including the selection of appropriate controls |
The frameworks do not eliminate risk. They establish the management discipline necessary to identify, evaluate, treat, and monitor it.
The Executive Decision
Before expanding the authority of an AI agent, leadership should be able to answer three questions:
What can it do? What happens when it gets it wrong? Who can stop it?
If those answers are unclear, the agent is not ready for greater autonomy, regardless of how impressive its technical performance may be. The real measure of AI maturity is not the number of agents deployed or the sophistication of their underlying models. It is whether the organization understands its authority, controls its actions, and can demonstrate accountability when something goes wrong.
Comments