What the OCC's new examiner framework means for AI agents at your bank
Author
Audrey
NobleCloak's AI Correspondent · AI-drafted, fact-checked against our sourced evidence before publishing.
Date Published
In short
- On November 4 the OCC published Bulletin 2026-54, a framework examiners may use to review how banks use AI, including AI agents. It is a resource for examiners, not a rule.
- It sorts AI risk by what a system can do, not how sophisticated it is. "A simple rules-based agent authorized to initiate transactions could present greater risk than a sophisticated model limited to internal drafting assistance."
- It defines human-in-the-loop as "review and approval before an output or action is used or executed," and says approval should be "more than a nominal click."
- Fed Governor Barr said on November 16 that major banks are finding AI agents "attempting to take actions in retail customer accounts."
- The comment deadline on the interagency third-party risk proposal moved to December 16.
1. OCC Bulletin 2026-54: an examiner framework, not a new rule
The OCC calls AI "Super Intelligence" (SI), following the September executive order that renamed the term. The bulletin is direct about what the document is: "This supervisory framework does not establish supervisory expectations for banks," and "Banks are not expected to use this document to assess their use of SI or their SI risk management." Examiners may use it at community banks, scaled to each bank's size, complexity and risk.
It does not create a separate AI risk category. Examiners look at AI through the risks a bank already manages: operational, compliance, credit and the rest.
Read it anyway. It shows the questions an examiner is now equipped to ask, and several are about agents.
Risk follows what the agent can do
"Risk classification may be based on potential consequences, transfer of agency, and the degree of human oversight, rather than the technical sophistication of the SI used." The framework's own example: "A highly autonomous payment agent may present greater SI risk than an SI agent only used for internal drafting."
The framework's illustrative tiers put it plainly: "a simple internal productivity tool with no customer or financial data exposure sits at Tier 4, whereas a fully autonomous, customer-facing payment agent sits at Tier 1, regardless of how well it performs in testing." An agent that can move money or change customer records sits at the top, however simple its logic.
Human-in-the-loop, defined
Section 2.3 separates two kinds of oversight:
- Human-in-the-loop "generally refers to review and approval before an output or action is used or executed."
- Human-on-the-loop "generally refers to active monitoring with authority and practical ability to intervene."
It then warns about the obvious failure: "Review triggers, evidence, and procedures may be defined so that approval is more than a nominal click. High volumes or bulk approval could reduce review quality and contribute to alert fatigue."
That's the rubber-stamp problem. An approval step that asks someone to click yes on everything doesn't count for much.
Decide in advance what needs a person
"The purpose statement may distinguish actions the system can take autonomously from actions that call for review, approval, abstention, or escalation." In practice, that's a list: which actions an agent may take alone, which need an approver, and which it may never take.
Be able to reconstruct what happened
"Inputs, outputs, context, retrieved sources, tool calls, approvals, overrides, and actions may provide evidence needed to reconstruct a material decision or event."
Treat agents like privileged users
"Agent identities, credentials, logs, and data use may be governed in a manner comparable to privileged users." And for agents a vendor runs: "accountability for managing the bank's risk generally remains with the bank."
2. Barr: agents are already acting in customer accounts
In a November 16 speech in Milan (his own views, not the Board's), Fed Governor Michael Barr said "major banks now report finding AI agents attempting to take actions in retail customer accounts." He pointed to "strict authentication, explicit consent verification for specific actions, technical controls limiting agent capabilities" as the safeguards. He also warned that agents could "autonomously move funds at high speed" during periods of stress.
3. More time on third-party risk
The OCC, Fed, FDIC and NCUA extended the comment period on their proposed third-party risk management guidance to December 16 (from November 16). The proposal doesn't mention AI. It matters here because AI vendors are third parties, and the OCC framework points back to third-party risk management for agents a vendor operates.
What didn't change this month
- Model risk guidance (OCC Bulletin 2026-13 / SR 26-2, April) still excludes generative and agentic AI. The agencies said a request for information on AI would follow, and none has been published yet.
- FDIC, NCUA, CFPB and NYDFS issued no new AI guidance between October 26 and November 17.
- Colorado: SB 26-189, signed May 14, repealed and re-enacted the 2024 AI Act as an automated decision-making technology (ADMT) law. Its first duties begin January 1, 2027. It covers decisions about "financial or lending services," and after an adverse outcome consumers can request human review.
What to do with this
The framework isn't a checklist, but it does describe what an examiner may look for when an agent can act. Four questions to answer before your next exam:
- Which agent actions run alone, which need a person, and which are never allowed? Write it down per action, not per agent.
- Who approves each one, and do they have the authority to? The framework's test is whether the approval is "more than a nominal click."
- Can you show one decision end to end: who asked, what the agent tried, who approved it, and when?
- Are agent identities and permissions managed like privileged users?
Where NobleCloak fits
NobleCloak puts the right person in the loop for the actions that need one. Policy marks each tool as allowed, needs approval, or denied. Approval goes to a named person or role. One approval covers one action. The Record of Decisions keeps who asked, which agent, which tool, the decision, who approved, and when.
It doesn't cover everything in the framework. It doesn't validate models, and it doesn't capture model inputs and outputs. It covers the part where an agent is about to act. Read how human-in-the-loop works or see the banking and credit union page.
Sources
- OCC Bulletin 2026-54 (Nov 4, 2026)
- OCC framework PDF (NR 2026-97a)
- Governor Barr, "Some Potential Financial Stability Implications of Artificial Intelligence" (Nov 16, 2026)
- Interagency TPRM comment extension (OCC NR-IA 2026-100)
- OCC model risk guidance release (NR 2026-29, Apr 2026)
- Colorado SB 26-189, Automated Decision-Making Technology (signed May 14, 2026)

