All posts
  • occ
  • occ bulletin 2026-54
  • ai agents
  • agentic ai
  • banking
  • sr 26-2
  • model risk management
  • ai governance
  • ai agent testing
  • regulatory compliance

OCC Super Intelligence Framework: How Examiners Will Review AI Agents

The OCC's November 2026 Super Intelligence framework shows how examiners will assess generative and agentic AI that SR 26-2 left out. Here are the six pillars, the risk tiers, the testing expectations and a 30-day readiness plan.

By Aodit

12 min read

On 4 November 2026, the Office of the Comptroller of the Currency (OCC) published Bulletin 2026-54, "Super Intelligence: Supervisory Framework", together with a detailed examiner framework (PDF). Despite the futuristic name, the document deals with something banks are already doing: deploying AI systems and AI agents in ordinary business activities.

For US banks, timing matters. The interagency model risk guidance issued in April 2026 (SR 26-2 / OCC Bulletin 2026-13 / FDIC FIL-15-2026) explicitly placed generative and agentic AI outside its scope. The OCC framework is the first detailed signal of how examiners intend to look at exactly those systems.

This article explains what the framework says, what it does not say, and how risk, compliance and AI governance teams can prepare evidence for their next examination.

What OCC Bulletin 2026-54 Is (and Is Not)

The framework is examiner guidance, not a new rule. The bulletin is explicit:

  • It "does not establish supervisory expectations for banks."
  • Banks "are not expected to use this document to assess their use of SI or their SI risk management."
  • Examiners should "not treat SI as a separate and distinct risk category," and banks are not required to define a separate SI risk appetite.
  • Examinations should be tailored to the bank's size, complexity and risk profile, and the OCC intends to update the guide as technology and use cases change.

"Super Intelligence" is the OCC's own label. Footnote 2 of the framework says the term covers technologies previously called "artificial intelligence" under 15 U.S.C. 9401(3), plus their rapidly advancing capabilities. In practice, think of it as AI, generative AI and AI agents.

That said, examiner guidance tells you what examiners will ask for. A bank that understands the framework knows which questions are coming and which evidence it needs. That is how most risk teams will read it.

How it fits with SR 26-2 / OCC 2026-13

SR 26-2's footnote 3 states that "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance." The revised guidance is also non-enforceable and "most relevant to" banking organizations with over $30 billion in total assets.

The OCC framework addresses that gap directly. Its first footnote states that the model risk management principles in OCC Bulletin 2026-13 apply to non-generative and non-agentic SI models, and that the framework "is more applicable to generative and agentic SI."

In other words, the carve-out from model risk guidance did not mean "unsupervised." It meant "supervised through a different lens."

The OCC is not alone. In September 2026 the Conference of State Bank Supervisors released a non-binding AI Supervisory Framework for state examiners that also covers generative and agentic AI. And in October 2026, Reuters reported that the Fed, OCC and FDIC were asking banks whether they can monitor, audit and shut down autonomous AI systems. We covered those questions in Agentic AI Risk Management in Banking: 6 Questions Supervisors Now Ask.

The Six Pillars and 16 Components

The framework organizes examiner review into six pillars: Purpose, Risk, Control, Oversight, Resilience and Assurance. Under them sit 16 components:

#ComponentWhat examiners look at, in short
2.1Defined Purpose and Permitted UseWhat the system may do, for whom, with which data and actions
2.2Risk Classification and CriticalityTiering by consequence, reversibility, autonomy
2.3Human-in-the-Loop / Human-on-the-LoopWhether human oversight is real or nominal
2.4Governance & AccountabilityOwnership; accountability stays with the bank
2.5Model Risk Management ExtensionsAdapting MRM to non-deterministic, vendor-dependent systems
2.6Data GovernanceLineage, quality, feedback loops
2.7PrivacyUse and protection of personal data
2.8ComplianceConsumer protection, BSA/AML, sanctions
2.9Third-Party & Vendor RiskProvider agents, change notice, liability
2.10Operational ResilienceDetecting, containing and recovering from misfires
2.11CybersecurityAttacks on and through AI systems
2.12Performance MonitoringHallucination, override and drift metrics
2.13TestingPre-deployment, adversarial, red teaming, hallucination testing
2.14Documentation & AuditabilityDecision traceability and change logs
2.15Internal Audit & Independent ChallengeAccess to underlying evidence
2.16Reporting & TransparencyBoard and management reporting

Four of these deserve particular attention from anyone deploying AI agents.

1. Consequences, Not Sophistication, Drive Scrutiny

The most important idea in the framework is that risk classification may rest on potential consequences, transfer of agency and degree of human oversight, "rather than the technical sophistication of the SI used."

Section 3 lists the classification factors: decision consequence, reversibility, customer/financial impact, autonomy level, data sensitivity, explainability and provenance. It then gives four illustrative tiers:

  • Tier 1 – Critical: direct action, low or no reversibility, high impact, autonomous or partly autonomous. Examples: autonomous payment agents, real-time credit decisioning, fraud-hold release, market access.
  • Tier 2 – High: direct action or high-weight decision support under human-on-the-loop oversight. Examples: AI-assisted underwriting, agent-initiated customer communications, AML alert disposition.
  • Tier 3 – Moderate: decision support with mandatory human-in-the-loop review, limited data sensitivity, reversible outcomes. Examples: credit memo summarization, agent-assisted research.
  • Tier 4 – Low: no customer or financial data exposure, no direct action, fully human-controlled. Examples: internal productivity tools, general employee Q&A.

The framework notes that a fully autonomous, customer-facing payment agent sits at Tier 1 "regardless of how well it performs in testing."

What this means in practice: a strong benchmark score does not lower the tier. A model that is 99% accurate but can move money autonomously is still critical. What reduces risk is scope (fewer permitted actions), reversibility and meaningful oversight. Teams should document why each agent sits in its tier, using these factors, not just the model's name or vendor.

2. Human Oversight Must Be Real

Component 2.3 separates two oversight modes:

  • Human-in-the-loop (HITL): a person reviews and approves before an output or action is used.
  • Human-on-the-loop (HOTL): a qualified person actively monitors and has the authority and practical means to override or halt actions.

The framework then raises the uncomfortable questions: reviewer competency and authority, meaningful review standards, span of control, override logging, escalation and reviewer independence. It also notes that high volumes or bulk approval can reduce review quality.

That is a direct challenge to "rubber-stamp" oversight, where one analyst approves hundreds of agent outputs per hour. If oversight is a control in your risk assessment, expect examiners to test whether it actually works.

A related point for agent design: oversight depends on the agent knowing when to stop. An agent that never escalates, or escalates everything, undermines HITL and HOTL alike. Escalation behavior is something you can and should test before deployment.

3. Testing: Repeated, Adversarial and Evidence-Based

Component 2.13 is the most operational section of the framework, and the one that maps most closely to behavioral stress testing. It describes testing across design, pre-deployment and live use, including:

  • Pre-deployment testing: functional, comparative, scenario, stress, adversarial, compliance, resilience, human-oversight and acceptance testing.
  • Adversarial testing: misleading, malicious or edge-case inputs designed to find weaknesses in prompt, data and tool controls.
  • Red teaming: testing whether the bank can detect, contain, respond to and recover from realistic attacks or misuse, with external expertise where appropriate.
  • Hallucination testing: measuring hallucinations under representative and adversarial prompts, defining which output categories require human verification, and sampling live outputs.
  • LLM-as-a-judge: if one model scores another, the judge's own validation, independence, bias and correlated failure modes need assessment, plus human review.

Component 2.5 (MRM extensions) adds a key principle for generative systems: repeated testing and consistency thresholds are preferred over a single successful run. It also notes that champion-challenger methods may not fit open-ended generative outputs.

Two points stand out for agent deployments.

First, a single demo or a single pass of a test suite is not evidence. Non-deterministic systems need to be tested repeatedly, and behavior under pressure (rephrased requests, escalating demands, manipulation) matters as much as baseline answers. Our prompt injection testing guide walks through banking-specific adversarial scenarios.

Second, the framework flags hallucinations about fees, rates, account status or policies as a heightened-risk area alongside BSA/sanctions and consumer compliance. For customer-facing agents, factual consistency is a compliance issue, not just a quality issue.

Where AODIT fits

This is the problem AODIT was built for. AODIT is an independent behavioral risk rating and stress-testing platform for AI agents in banking and finance. Its AODIT-6 methodology evaluates agents across six dimensions (Reliability, Integrity, Confidentiality, Judgment, Resistance and Resilience) using an 8-turn adversarial protocol that moves from baseline questions through grey-zone requests, hard pushes and manipulation attempts to recovery and final judgment.

That structure lines up with several things component 2.13 describes:

  • Adversarial testing: the Resistance dimension covers jailbreaks, prompt injection, social engineering, authority spoofing and adversarial reframing.
  • Escalation and oversight behavior: the Judgment dimension assesses escalation and blast-radius awareness, which is relevant evidence for HITL/HOTL design.
  • Consistency over a single run: full evaluations run 20–100 scenarios per test, rather than relying on one successful interaction.
  • Documentation: results come as a PDF-ready report you can attach to your testing file.

AODIT's results are advisory and complement, rather than replace, your own validation, monitoring and controls. It does not require access to model weights, and on-premise and air-gapped deployment is supported for banks that cannot send data outside their environment.

You can see how it works in two minutes: the free live demo runs an 8-turn adversarial test against a sample agent, with no account needed.

4. Containment, Vendors and Evidence

Containment and recovery

Component 2.10 expects examiners to look at how a bank detects and contains a "misfire": monitoring and observability tools, control agents, deterministic rules, circuit breakers, rate limits, least-privilege permissions and kill switches. It also covers continuity (falling back to trained staff or a narrower agent), recovery (rollback, data correction, root-cause analysis) and customer remediation, including identifying the full affected population and independently verifying completion.

The practical lesson: a kill switch that has never been tested is a hope, not a control.

Third parties

Component 2.9 makes clear that accountability "generally remains with the bank" even when a vendor provides the model or agent. It points examiners toward provider-agent authority and tool-use limits, an AI incident vocabulary (hallucination, misfire, drift, manipulation), notification timelines, liability allocation, and notice of new AI features or model changes, followed by retesting.

Silent vendor model updates are a real risk. If your provider changes the underlying model, your previous test results may no longer describe the agent in production. Banks with EU operations will recognize similar themes from DORA, covered in our DORA and AI agents article.

Documentation examiners may ask for

Across components 2.1, 2.13, 2.14, 2.15 and 2.16, the framework points to a consistent evidence set:

  • An inventory with use case, owner, provider, permissions, deployment status, monitoring and fail-over arrangements.
  • Decision traceability: inputs, outputs, retrieved sources, tool calls, approvals, overrides and actions.
  • Change logs for model, prompt, data, configuration, provider, tool and permission changes, linked to testing and approvals.
  • Testing evidence: results, sampling methods, reviewer qualifications and hallucination measurements.
  • Audit access to underlying evidence, "not only management summaries."
  • Reporting to management and the board on material use cases, incidents, control exceptions and customer remediation.

A 30-Day Readiness Plan for AI Agent Owners

The framework does not create new requirements, but it shows what a good answer looks like. Here is a practical sequence we recommend for banks running or piloting AI agents:

  1. Week 1 – Inventory and purpose. List every agent and generative AI use case. For each, write down permitted actions, data access and users. Activity beyond an approved purpose is exactly the kind of gap examiners are likely to treat as a control weakness.
  2. Week 1 – Tier each use case. Apply the seven classification factors. Record the reasoning, especially for anything that can act autonomously or affect customers.
  3. Week 2 – Test oversight. For HITL and HOTL controls, check review volumes, reviewer authority and override logs. Test whether agents escalate when they should.
  4. Week 2–3 – Run repeated, adversarial behavioral tests. Prioritize Tier 1 and Tier 2 agents. Cover manipulation, prompt injection, hallucinations about fees, rates and policies, and recovery after pressure. Test more than once.
  5. Week 3 – Exercise containment. Trigger the kill switch or circuit breaker in a controlled setting. Document how long containment took and what fallback was used.
  6. Week 4 – Review vendor terms. Check for change notification, incident notification, data use and liability clauses. Define what triggers retesting.
  7. Week 4 – Assemble the evidence pack. Bring inventory, tiers, test results, change logs and oversight metrics together in a form internal audit can verify.

These are our recommendations, not OCC requirements. They are designed to make the framework's questions easy to answer.

Key Takeaways

  • OCC Bulletin 2026-54 is examiner guidance. It sets no new expectations for banks, but it tells you how examiners will think about AI and AI agents.
  • It fills the space left by SR 26-2 / OCC 2026-13, which excluded generative and agentic AI from model risk guidance.
  • Consequences, reversibility, autonomy and oversight drive risk tier, not model sophistication or test scores.
  • Testing should be repeated, adversarial and documented, including hallucination testing on fees, rates and policies.
  • Accountability stays with the bank, including for vendor-provided agents.

Ready to build the testing evidence examiners will look for? Start with the free AODIT demo, review the AODIT-6 methodology, or request an evaluation of your production or pilot agents. Plans and report options are on the pricing page.

FAQ

What is the OCC Super Intelligence Supervisory Framework?

It is an examiner framework issued with OCC Bulletin 2026-54 on 4 November 2026. It helps OCC examiners evaluate how national banks and federal savings associations use AI ("super intelligence") and AI agents, organized into six pillars and 16 components.

Is OCC Bulletin 2026-54 binding on banks?

No. The bulletin states that it does not establish supervisory expectations for banks and that nothing in it is intended to establish new regulatory expectations. Existing laws and safety-and-soundness standards still apply.

Does the framework cover generative AI and AI agents?

Yes. The framework says it is more applicable to generative and agentic SI, while the model risk management principles in OCC Bulletin 2026-13 apply to non-generative and non-agentic models.

What makes an AI use case Tier 1 (critical)?

Direct action, low or no reversibility, high impact and autonomous or partly autonomous operation. The framework's examples include autonomous payment agents, real-time credit decisioning, fraud-hold release and market access.

What kind of AI testing does the framework describe?

Pre-deployment functional, scenario, stress, adversarial, compliance, resilience and human-oversight testing; adversarial testing and red teaming; hallucination testing; and repeated testing with consistency thresholds rather than a single successful run.

Does it apply to community banks?

Examiners may use it when supervising community banks, tailored to the bank's size, complexity and risk profile and to the nature, consequences and autonomy of its AI use cases.


Sources

This article is for general information and does not constitute legal or regulatory advice.

  1. 11 min read

    Excessive Agency in Banking AI Agents: Permissions and Blast Radius

    Once an AI agent can issue refunds or move money, its permissions become your biggest risk. Here's what OWASP, the OCC, FSB and DORA say about excessive agency, and a 6-step framework to limit an agent's blast radius.

    • excessive agency
    • ai agent permissions
    • least privilege
    • blast radius

    Read

  2. 14 min read

    LLM Model Change Management for Banking AI Agents: A Retest Playbook

    Vendor model retirements, alias moves and prompt edits can quietly change how a banking AI agent behaves. Here's what the OCC, DORA, the EU AI Act and the PRA say about change, and a 7-step playbook to retest agents before the switch.

    • model change management
    • llm model updates
    • ai agents
    • banking

    Read

  3. 13 min read

    EU AI Act Credit Scoring: What Banks Must Do Before December 2027

    Creditworthiness AI is high-risk under the EU AI Act, with obligations from 2 December 2027. Here's what's in scope (including LLM agents), who does what, what the Omnibus changed, and a 10-point readiness checklist.

    • eu ai act
    • credit scoring
    • creditworthiness assessment
    • high-risk ai

    Read