AI Governance & Compliance
OCC Bulletin 2026-54: What Examiners May Ask About AI Agents
OCC Bulletin 2026-54 is not mandatory. See who it covers, how it differs from 2026-13 and what examiners may ask about AI agents.
Published on


Is OCC Bulletin 2026-54 Mandatory? What Examiners May Ask About AI Agents
By Tahir Mahmood, Co-founder & CTO, OpenBox ยท Last updated November 17, 2026
In brief: OCC Bulletin 2026-54, issued November 4, 2026, is the OCCโs โSuper Intelligence: Supervisory Framework,โ a resource for examiners that does not establish supervisory expectations for banks. It describes what examiners may consider about AI agents: consequence, reversibility, human oversight and containment. OpenBox, an AI agent governance platform, turns those passages into questions and per-action evidence. |
Is OCC Bulletin 2026-54 mandatory?
No. The OCCโs bulletin says the framework โdoes not establish supervisory expectations for banks.โ It calls the framework a resource to assist examiners. Banks are not expected to use it to assess their own use of AI or their AI risk management. The OCC also says nothing in it is intended to establish new regulatory expectations or enforceable standards.
The OCC issued Bulletin 2026-54, โSuper Intelligence: Supervisory Framework,โ on November 4, 2026. The OCC uses โSuper Intelligenceโ (SI) as its term for technologies previously covered by the term artificial intelligence. This article says AI except when quoting the OCC. The OCC intends to update the framework as the technology, use cases and supervisory understanding change, and the framework describes itself as a living document.
Examiners may use the framework to identify AI-specific considerations that supplement existing model risk management, governance, technology, compliance and operational-risk practices. Reading it as a preview of the conversations examiners may have is this articleโs interpretation, not an OCC statement. OpenBoxโs July 2026 article Transatlantic AI Governance: EU vs US for Agents compares how the EU and the US approach agent governance.
Who does the OCCโs AI framework apply to?
The framework is written for OCC examiners. The bulletin is addressed to the chief executives of all national banks, federal savings associations, and federal branches and agencies, as well as department and division heads and examining personnel. Examiners may use the framework for community banks, and the OCC has instructed them to tailor examinations to the bankโs size, complexity and risk profile. This article does not assess how other agenciesโ examiners treat AI agents.
In the bulletin, โbanksโ means national banks, federal savings associations, and federal branches and agencies of foreign banking organizations. The framework covers examinersโ evaluation of banksโ use of AI in normal-course activities, including credit decisioning, payments, fraud detection, customer service, operations, risk management and market activities. It adds that examiners should not treat AI as a separate and distinct risk category, and that banks are not required to define a separate AI risk appetite.
OCC 2026-13 vs 2026-54: what changed for AI agents and model risk management?
Bulletin 2026-13 does not cover generative or agentic AI models, and the 2026-54 framework is more applicable to them. Bulletin 2026-13, dated April 17, 2026, is titled โModel Risk Management: Revised Guidance.โ It states that generative AI and agentic AI models are not within its scope. The frameworkโs first footnote says the model risk principles in 2026-13 apply to non-generative and non-agentic models. It adds that the framework is โmore applicable to generative and agentic SI.โ
Bulletin | Date | Title | Generative and agentic AI | Status |
2026-13 | April 17, 2026 | Model Risk Management: Revised Guidance | Not within scope. | Does not set forth enforceable standards or prescriptive requirements. |
2026-54 | November 4, 2026 | Super Intelligence: Supervisory Framework | โMore applicable to generative and agentic SI.โ | Does not establish supervisory expectations for banks. |
Bulletin 2026-13 names the OCC, the Federal Reserve Board and the FDIC as the agencies behind it, and says it does not set forth enforceable standards or prescriptive requirements. OpenBoxโs May 2026 article AI Compliance Frameworks for Financial Services in 2026: Mapping the Territory covers the wider set of regimes that financial services firms face, including SR 11-7 model risk management. It predates Bulletin 2026-54.
How do bank examiners evaluate AI agents under this framework?
The OCC framework says control intensity could increase with the potential consequences of an agentโs failure or misuse. It adds that risk classification may rest on potential consequences, transfer of agency and the degree of human oversight, โrather than the technical sophistication of the SI used.โ Its example is a simple rules-based agent authorized to initiate transactions, which could present greater risk than a sophisticated model limited to internal drafting assistance.
The framework groups sixteen components into six pillars: Purpose, Risk, Control, Oversight, Resilience and Assurance. Purpose covers the actions an agent is permitted to execute, and Assurance covers evidence that its controls continue to operate effectively.
Consequence, reversibility and oversight
The frameworkโs Section 3.1 sets out an illustrative methodology in which a use case may be scored on decision consequence, reversibility, customer and financial impact, autonomy level, data sensitivity, explainability and provenance. It states that direct action carries materially higher risk than decision support. Section 2.3 of the framework says human-in-the-loop generally refers to review and approval before an output or action is used or executed. It says human-on-the-loop generally refers to active monitoring with authority and practical ability to intervene.
OpenBox has written separately on reversibility as a governance requirement and on assessing and tiering agentic AI use cases by risk.
The OCCโs illustrative tiers
The OCCโs illustrative tiers run from Tier 1 Critical, the highest-risk tier, to Tier 4 Low. Section 3.2 of the framework defines them as summarized below.
Tier | Name | Described in the framework as |
Tier 1 | Critical | Direct action, low or no reversibility, high financial or customer impact, autonomous or partially autonomous. |
Tier 2 | High | Direct action or high-weight decision support, human-on-the-loop, moderate-to-high impact. |
Tier 3 | Moderate | Decision support with mandatory human-in-the-loop, limited data sensitivity, reversible. |
Tier 4 | Low | No customer or financial data exposure, no direct action, fully human-controlled use. |
Check the direction of any tier scale before comparing frameworks.
What evidence would answer an examinerโs questions about AI agent actions?
The framework says inputs, outputs, context, retrieved sources, tool calls, approvals, overrides and actions โmay provide evidence needed to reconstruct a material decision or event.โ A per-action record could therefore help answer the questions below. The mapping is this articleโs reading of the framework, and no part of it is an OCC requirement.
Framework passage | Question an examiner may ask | Per-action evidence that could answer it | What the OpenBox docs describe |
OCC framework, Section 2.1: activity outside the approved purpose may be treated as a control failure even when the output or action appears reasonable. | Did the agent ever act outside its approved purpose, even when the result looked fine? | A per-session record of each tool call and data access that can be compared with the written purpose. | Behavioral Rules detect multi-step patterns across a session, such as PII access followed by an external API call without approval. |
OCC framework, Section 2.3: review procedures may be defined so that approval is more than a nominal click, and override authority, reasons and patterns may be logged and reviewed. | Who approved or overrode which action, when, and why? | Approval and rejection records with the reviewer, reason and timing. | The Approvals page is the organization-wide human-in-the-loop queue. Rejections require a reason, approvals and rejections are logged in the audit trail, and a request that times out does not proceed. |
OCC framework, Section 2.14: records may support reconstruction of material outputs or actions even when the model cannot produce a rule-based rationale. | Can you rebuild this decision from the record? | Ordered session events with their payloads. | Session Replay is a step-by-step walkthrough of a session, covering events, tool calls, governance decisions and JSON payloads, opened from Agent Detail, then Verify. |
OCC framework, Section 2.14: model, prompt, data, configuration, provider, tool and permission changes may be logged and linked to testing and approval evidence. | What changed in policy or permissions, and who changed it? | A change log for policy, guardrail and permission settings. | The organization audit log covers 11 event types, including policy, guardrail and role changes, and exports to CSV and Excel. |
OCC framework, Section 2.10: kill switches, circuit breakers, rate limits and least-privilege permissions could help identify and contain an agent that is operating but acting incorrectly. For more consequential or autonomous use cases, a bank may document and test containment and suspension. | Can you stop an agent that is running and wrong, and have you tested it? | Records of when a session or action was stopped, what triggered it, and the results of any containment test. | A Behavioral Rule can return HALT, which terminates the entire agent session, or BLOCK, which rejects the action while the agent continues. |
A record is useful in proportion to the confidence placed in it. The OpenBox docs describe session attestation. Each sessionโs events are hashed with SHA-256 into a Merkle tree and digitally signed. The result is a proof certificate that holds the Merkle root, the signature and the event count. The docs say the certificate is verifiable and confirms that no governance data was altered after the fact. OpenBoxโs article AI Governance Needs Proof, Not Logs makes the broader case.
What a per-action record does not show on its own
An action record shows what an agent did and who approved it. On its own it does not show whether a reviewer had the expertise, information and authority the framework describes. Nor does it show whether the purpose was well chosen, or whether controls work across a whole portfolio of agents.
The framework covers far more than action records. Its components include data governance, privacy, compliance, third-party and vendor risk, performance monitoring, testing, internal audit and independent challenge, and reporting. Those need other evidence from other teams.
OpenBox has covered the human side in Human-in-the-Loop AI, Explained the Right Way and the identification of high-impact agent outputs in What Are Consequential Events in AI Agents?.
What about AI agents run by customers, fintechs or vendors?
The OCC framework says customer, counterparty, fintech or agentic-commerce systems that a bank does not control could present channel risk. For these external agents it lists identity, provenance, delegation, authentication, monitoring, velocity limits, dispute handling, suspension and revocation, which may be calibrated to the action requested.
Banks also frequently obtain AI capabilities from third parties. The framework says a bank may define the scope of authority, identity, permitted tool use, data use and escalation limits for provider agents. Records from agents a bank runs itself do not cover agents it does not run. Evidence for external agents is more likely to come from channel controls such as identity, delegation, suspension and revocation.
OpenBoxโs article on DORA compliance for AI agents notes that DORA has no separate AI-agent regime, so agents fall under existing ICT risk rules.
What OpenBox documents for per-action evidence
OpenBox, an AI agent governance platform, documents several features that produce per-action records. This article does not claim that any of them maps to Bulletin 2026-54 or ensures any supervisory outcome. What follows is what the docs describe.
โข Behavioral Rules: the docs describe โstateful authorization rules that detect multi-step patterns across an agentโs session.โ Among its verdicts, a rule can return REQUIRE_APPROVAL, BLOCK or HALT. REQUIRE_APPROVAL sends the action to the human-in-the-loop queue. BLOCK rejects the action while the agent continues. HALT terminates the entire agent session. See the Behavioral Rules docs.
โข Approvals: the organization-wide queue where a reviewer approves or rejects a pending request. The default timeout is five minutes, configurable per behavioral rule from five to 60 minutes. See the Approvals docs.
โข Session Replay: a walkthrough of a session in which the docs say you can inspect each event, tool call, governance decision and full JSON payload. See the Session Replay docs.
โข Attestation: SHA-256 event hashing into a Merkle tree, signed with AWS KMS (ECDSA P-256) by default or through an external attestation endpoint. See the attestation docs.
โข Compliance & Audit: the docs state that OpenBox โrecords all AI agent governance decisions in an immutable audit trail.โ Governance events carry a timestamp, agent, event type, verdict, reason and workflow or run ID. See the Compliance & Audit docs.
OpenBoxโs July 2026 announcement with Temporal describes five governance decisions: ALLOW, CONSTRAIN, REQUIRE_APPROVAL, BLOCK and HALT, with CONSTRAIN running an action inside a sandbox.
Where to start with OCC Bulletin 2026-54
One practical way to use OCC Bulletin 2026-54 is to pick an agent that can take a consequential action. Then try to answer the five questions above from its records alone. That is the authorโs suggestion, not an OCC instruction. Each question you cannot answer from records marks a gap worth examining.
For the wider foundation, see What Is AI Agent Governance? The Complete Enterprise Guide, and for implementation detail start with the OpenBox docs.
Frequently asked questions
Why does the OCC call AI โSuper Intelligenceโ?
The OCC uses โSuper Intelligenceโ (SI) as its term for the technologies and systems previously covered by the term artificial intelligence, as defined in 15 USC 9401(3), together with their advancing capabilities. A footnote cites Executive Order 14434 (91 Fed. Reg. 63129, October 2, 2026). This article says AI except when quoting the OCC.
Does the OCC framework require a kill switch for AI agents?
No. The bulletin says the framework does not establish supervisory expectations for banks. The operational resilience section says kill switches, circuit breakers, rate limits and least-privilege permissions could help identify and contain an agent that is operating but acting incorrectly, and that a bank may document and test containment and suspension arrangements for more consequential or autonomous use cases.
Can a bank use the OCC framework to assess its own AI?
The bulletin says it is a resource for examiners and that banks are not expected to use it to assess their own use of AI or their AI risk management. Reading it to understand what examiners may consider is a reasonable step, but that is this articleโs suggestion, not an OCC instruction.
Sources |
OCC Bulletin 2026-54, Super Intelligence: Supervisory Framework Office of the Comptroller of the Currency, November 4, 2026. Accessed November 17, 2026. Supervisory Framework on Bank Use of Super Intelligence (PDF) Office of the Comptroller of the Currency, attached to OCC News Release 2026-97. Accessed November 17, 2026. OCC Bulletin 2026-13, Model Risk Management: Revised Guidance Office of the Comptroller of the Currency, April 17, 2026. Accessed November 17, 2026. OCC 2026 bulletins list Office of the Comptroller of the Currency. Accessed November 17, 2026. OpenBox docs, Behavioral Rules: https://docs.openbox.ai/trust-lifecycle/authorize/behaviors Accessed November 17, 2026. OpenBox docs, Approvals: https://docs.openbox.ai/approvals Accessed November 17, 2026. OpenBox docs, Session Replay: https://docs.openbox.ai/trust-lifecycle/session-replay Accessed November 17, 2026. OpenBox docs, Attestation & Cryptographic Proof: https://docs.openbox.ai/administration/attestation-and-cryptographic-proof Accessed November 17, 2026. OpenBox docs, Compliance & Audit: https://docs.openbox.ai/administration/compliance-and-audit Accessed November 17, 2026. OpenBox, announcement with Temporal, July 13, 2026: https://www.openbox.ai/blog/temporal-and-ai-agent-governance Accessed November 17, 2026. |

