The OCC's AI framework for examiners, sent with Bulletin 2026-54, has sixteen components (sections 2.1 to 2.16), from defined purpose and risk classification through human oversight, data, privacy, vendors, resilience, testing and audit to reporting. Each describes how AI can change an existing control area and gives illustrative considerations, and a summary table pairs each one with an AI-specific risk. Six pillars group them, and examiners are told not to treat AI as a separate risk category.
TRACKED DOCUMENTS: OCC Bulletin 2026-54
What are the sixteen components of the OCC AI framework?
They are sixteen control areas, numbered 2.1 to 2.16, that examiners may use when they look at how a bank uses AI, especially generative and agentic AI. Each one describes how AI can change a familiar area such as data, privacy or audit, and gives examples of what to consider.
The framework was sent to examiners with OCC Bulletin 2026-54. The OCC writes “Super Intelligence (SI)” for what everyone else calls artificial intelligence, so this page says AI. The bulletin says the framework does not set supervisory expectations for banks. It is a tool for examiners, and its verbs are mostly “may” and “could”.
Section 2 says each component describes how AI could affect a control area, names key concepts and gives illustrative considerations. It also says the components are meant to support examiners' risk-based judgment, applied in proportion to the consequences, autonomy and complexity of the bank's AI use. Put simply, a chatbot that drafts internal notes is not looked at the same way as an agent that moves money.
A footnote in the executive summary adds that the model risk principles of OCC Bulletin 2026-13 apply to AI models that are neither generative nor agentic. So the framework is more applicable to generative AI (which writes text, images or code) and agentic AI (which can decide and act on its own).
Is AI a separate risk category for OCC examiners?
No. The executive summary says examiners “should not treat SI as a separate and distinct risk category, and banks are not required to define a separate SI risk appetite.”
Instead, AI can “introduce, amplify or modify risk in existing risk categories, such as operational, compliance and credit.” Those risks, including AI's part in them, are handled through the bank's existing risk management framework. That is why the sixteen components read like a list of ordinary control areas. For example, a lending chatbot that quotes the wrong rate creates compliance risk, and it is managed as compliance risk.
Section 1.1 makes the same point from another side. It says the framework may help link AI risk to enterprise functions such as risk identification and risk appetite, “while not inferring any requirement to aggregate SI risk.” Section 2.4 still says the board and senior management may set an AI risk appetite. Put simply, and by our reading of the two together, the framework allows one but does not ask for a separate one.
How do the overarching principle and six pillars organize the components?
One principle sits above all sixteen components: the stronger the possible harm from an AI failure, the stronger the controls can be. Six pillars then group the components by the main question each one answers.
Section 1.1 says a bank's control intensity could increase with the potential consequences of an AI agent's failure or misuse. Higher consequence uses may warrant tighter limits on autonomy, more direct human oversight, more independent validation (an independent check that the system works as designed) and stronger resilience. Risk may be judged by consequences, transfer of agency and human oversight, not by how advanced the technology is. The OCC's own example: a simple rules-based agent allowed to start transactions could be riskier than a sophisticated model limited to internal drafting. The four risk tiers turn this principle into a scale.
Section 1.2 lists the pillars, with the components that mainly address each one:
| Pillar | The question it answers | Components |
|---|---|---|
| 1. Purpose | What actions is the AI allowed to take? | 2.1 |
| 2. Risk | What happens if its output or action is not fit for purpose? | 2.2 |
| 3. Control | What stops unauthorized, inappropriate or unsafe actions? | 2.3, 2.5, 2.6, 2.7, 2.8, 2.11 |
| 4. Oversight | Who watches the AI and who is accountable for it? | 2.4, 2.3, 2.12, 2.16 |
| 5. Resilience | Can the bank detect, contain and recover from a failure and keep critical work going? | 2.10, 2.11 |
| 6. Assurance | Is there evidence the controls keep working? | 2.13, 2.14, 2.9, 2.15 |
Two components appear twice. Human oversight (2.3) sits under both Control and Oversight, and cybersecurity (2.11) under both Control and Resilience. Third-party risk (2.9) is placed under Assurance, not Control. The pillars are a summary. The sections below group the components by subject instead, which is easier to read.
What AI-specific risk does the OCC attach to each component?
The framework ends section 2 with a summary table that pairs each component with one illustrative AI-specific risk. Here is that table with the risk put in plain words. The OCC's exact wording is in the full reference guide.
| No. | Component | The risk, in plain words |
|---|---|---|
| 2.1 | Defined Purpose and Permitted Use | The AI drifts beyond its job, or takes actions nobody approved, because its limits were never written down in a way that can be tested. |
| 2.2 | Risk Classification and Criticality Assessment | The strength of the controls does not match how much harm the AI could do, how independently it acts, or whether its mistakes can be undone. |
| 2.3 | Human-in-the-Loop and Human-on-the-Loop Controls | A person is named as the reviewer but lacks the authority, skill, time or clear standard to actually check the AI. |
| 2.4 | Governance & Accountability | Nobody clearly owns the AI, or the people overseeing it do not understand it well enough to challenge it. |
| 2.5 | Model Risk Management Extensions | A one-time check of the model misses that it gives different answers each time, keeps changing, or produces new kinds of output. |
| 2.6 | Data Governance | The bank cannot tell where the data came from, whether it is good, or whether it may be used, especially vendor, synthetic or fast-changing data. |
| 2.7 | Privacy | The AI memorizes, guesses, keeps or shares sensitive information beyond what was approved. |
| 2.8 | Compliance | The AI makes it hard to apply the law, explain a decision, or control what customers are told. |
| 2.9 | Third-Party & Vendor Risk | The bank cannot see inside its providers, has weak contract rights, cannot easily switch, or depends on the same few models and infrastructure as everyone else. |
| 2.10 | Operational Resilience | The AI keeps running while doing the wrong thing, and there is no tested way to stop it, carry on and fix the harm. |
| 2.11 | Cybersecurity | Attackers use prompts, data, the model, its tools or its login identity to get around defenses built for ordinary systems. |
| 2.12 | Performance Monitoring | The AI's behavior slowly moves away from what was tested, and nobody notices in time. |
| 2.13 | Testing | Tests confirm the AI does what was expected but never look for unknown failures, attacks or made-up answers. |
| 2.14 | Documentation & Auditability | The bank cannot rebuild what the AI did, what changed, which tools it used or who approved it. |
| 2.15 | Internal Audit & Independent Challenge | The people giving assurance are not independent enough or not technical enough to challenge the builders and the checkers. |
| 2.16 | Reporting & Transparency | Important AI uses, changes, incidents or customer harm are not reported fully or on time. |
What do the purpose and risk classification components cover?
Components 2.1 and 2.2 come first because everything else depends on them. One sets what the AI is for, the other sets how risky that use is.
2.1 Defined Purpose and Permitted Use
Before deployment and after a material change, a bank may document what the AI is designed to do and not do, who may use it, what data and systems it can reach, what actions it could take, and when a human should review or suspend it. The inventory may link these limits to an owner and to monitoring. Section 2.1 says activity outside the approved purpose may be treated as a control failure even when the result looks reasonable. For example, a research assistant approved for internal use that starts drafting customer letters has drifted out of scope. The table's risk: scope drift or unauthorized action without a testable boundary.
2.2 Risk Classification and Criticality Assessment
A bank may assess each use before deployment and after material changes. Factors include how serious and how reversible a decision is, impact on customers and the bank, autonomy, data sensitivity, explainability, provenance (where the model comes from) and shared dependencies. Section 2.2 gives an example: a highly autonomous payment agent may present greater risk than an agent only used for internal drafting. The table's risk: control strength that does not reflect consequence, autonomy, reversibility or combined exposure. Section 3 turns this into four illustrative risk tiers.
Which components cover people, accountability, audit and reporting?
Four components deal with the humans around the AI: the reviewer (2.3), the board and management (2.4), internal audit (2.15) and the reports that reach them (2.16).
2.3 Human-in-the-Loop and Human-on-the-Loop Controls
Human-in-the-loop means a person reviews and approves before an output or action is used. Human-on-the-loop means a person actively monitors and can step in. Section 2.3 says the reviewer may need enough expertise, information and authority to reject, change, pause or halt the AI, and that review procedures may be defined so that approval is more than a nominal click. High volumes or bulk approval could cause alert fatigue. The table's risk: a named reviewer without real authority, skill, capacity or a meaningful review standard. More on the human oversight page.
2.4 Governance & Accountability
The board and senior management may set an AI risk appetite, name accountable executives and use-case owners, and receive information on material AI use. Section 2.4 notes that when vendors build or host the AI, accountability for managing the bank's risk generally remains with the bank. The table's risk: diffuse ownership or too little technical fluency for real oversight and challenge.
2.15 Internal Audit & Independent Challenge
Internal audit could give extra assurance over governance, validation and controls. Section 2.15 says effective challenge may depend on independence from both the team that builds or owns the AI and the team that validates or monitors it. Audit may need the underlying evidence, not only management summaries. The table's risk: assurance that is not independent or technical enough.
2.16 Reporting & Transparency
Internal reporting may cover material uses, owners, significant changes, control exceptions, incidents, customer remediation and concentrations. Section 2.16 says a bank may set escalation thresholds by severity, duration, customer impact and legal requirements, and may reconcile reports to the AI inventory and audit findings. The table's risk: material uses, changes, incidents or harm not reported fully or on time. The board, audit and reporting page covers 2.4, 2.15 and 2.16 together.
Which components cover validating, monitoring, testing and documenting the AI?
Components 2.5, 2.12, 2.13 and 2.14 adapt model risk management to AI. Model risk is the chance of harm from decisions based on wrong or misused model outputs.
2.5 Model Risk Management Extensions
Some AI does not fit older assumptions such as stable behavior and repeatable outputs. Section 2.5 says model risk principles may be instructive but adapted. It names drift (behavior changing over time, for example after a vendor update or a prompt change), the limits of explainability tools, and non-determinism (the same question getting different answers), where repeated testing may beat one successful run. The table's risk: a point-in-time validation that misses these features.
2.12 Performance Monitoring
Testing happens before launch. Monitoring watches the live system as data, prompts, permissions and vendor models change. Section 2.12 lists metrics such as accuracy, hallucination rate (a hallucination is a fluent but false answer), override and complaint rates, and thresholds that could trigger review, reduced autonomy or suspension. The table's risk: live behavior drifting from the tested baseline without anyone noticing in time.
2.13 Testing
Section 2.13 calls testing central and says its rigor may depend on how much is at stake. It describes testing across design, pre-deployment and live use, and names approaches such as expert review, adversarial testing (feeding misleading or malicious inputs), red teaming, using one model to judge another, and hallucination testing. The table's risk: tests that confirm expected behavior but never hunt for unknown, adversarial or hallucination failures.
2.14 Documentation & Auditability
Documentation lets people other than the builders understand and audit the AI. Section 2.14 says records may allow a material output or action to be rebuilt, through inputs, outputs, sources, tool calls, approvals and overrides, and that changes to model, prompt, data or permissions may be logged. The table's risk: being unable to reconstruct what happened. All four are on the model risk management page.
Which components cover data, privacy, compliance and vendors?
Components 2.6 to 2.9 cover what goes into the AI, whose information it touches, which laws apply to what it does, and who supplies it.
2.6 Data Governance
Section 2.6 says AI is only as reliable as the quality, lineage (where data came from and what was done to it), representativeness and permitted use of its data. It covers new data the AI creates, such as prompts, outputs and logs, and synthetic data, whose flaws can be hard to spot. The table's risk: unclear lineage, quality or permissions, especially in vendor, synthetic or fast-changing data.
2.7 Privacy
Section 2.7 says a model might reproduce training data or infer sensitive traits it was never given, and that prompts and logs could become stores of sensitive information. Contracts may settle whether a vendor can use bank data to train models for other customers. The table's risk: memorization, inference, retention or vendor use beyond approved purposes.
2.8 Compliance
Section 2.8 says AI does not change a bank's legal obligations. It names the Bank Secrecy Act (the anti-money-laundering law) and sanctions, consumer compliance, and hallucinations about fees, rates, account status or policies. In practice, a chatbot that invents a fee waiver creates the same exposure as an employee who does. The table's risk: AI that makes legal requirements, decision reasons or customer statements hard to control or explain. See the data, privacy and compliance page for 2.6 to 2.8.
2.9 Third-Party & Vendor Risk
A third party is any outside company the bank relies on. Section 2.9 says AI third-party risk management may supplement the existing interagency guidance. The framework does not name it, but in our reading that is OCC Bulletin 2023-17. Its examples are limits on provider agents that touch bank systems, an AI incident vocabulary with notice timelines, contract liability for model errors, and notice when a vendor adds AI to a product already in use. The table's risk: opaque provider chains, weak contract rights, switching limits or concentration in common models. More on the vendor page.
Which components cover AI failures and cyberattacks?
Components 2.10 and 2.11 cover keeping the business running when AI goes wrong, and defending AI from attack.
2.10 Operational Resilience
Section 2.10 says resilience for AI covers both loss of service and incorrect operation: an agent can stay online while doing harm. Tools it names include circuit breakers, rate limits, least-privilege permissions and kill switches (a way to shut the AI off at once). It also covers fallback processes, customer remediation and outside AI agents acting through bank channels. The table's risk: a system that keeps acting wrongly with no tested way to contain it, carry on, recover and fix the harm.
2.11 Cybersecurity
Section 2.11 says AI can be both a target and a tool for attackers. It names prompt injection (hidden instructions in an input that hijack the AI), data and model poisoning, model extraction and data leakage, and agents with too many permissions. It says agent identities may be governed like privileged users. The table's risk: attacks through prompts, data, models, tools or identities that bypass defenses built for ordinary systems. Both are on the resilience and cybersecurity page.
For every other OCC AI document, start at the series hub or the OCC tracker.
How many components does the OCC AI framework have?
Sixteen, numbered 2.1 to 2.16. They run from Defined Purpose and Permitted Use to Reporting & Transparency. A summary table at the end of section 2 pairs each with one illustrative AI-specific risk.
Do banks have to implement all sixteen components?
The framework is a resource for examiners, and Bulletin 2026-54 says it does not establish supervisory expectations for banks. The components describe what examiners may consider, in proportion to the bank's AI use. Their wording is mostly 'may' and 'could'.
What is the difference between the components and the pillars?
The components are the sixteen detailed control areas. The six pillars (Purpose, Risk, Control, Oversight, Resilience, Assurance) are a high-level grouping of them in section 1.2. Some components sit under two pillars.
Does the framework create a new AI risk category?
No. Examiners are told not to treat AI as a separate risk category, and banks are not required to define a separate AI risk appetite. AI is seen as changing existing risks such as operational, compliance and credit risk.
Which component is most relevant to a customer-facing chatbot?
Several apply, but compliance (2.8) names hallucinations about fees, rates, account status and policies. Purpose (2.1) also notes that an internal tool may warrant extra review before it becomes customer-facing. Privacy (2.7) and monitoring (2.12) are also relevant.
Does the framework apply to traditional machine-learning models?
A footnote says the model risk principles in OCC Bulletin 2026-13 apply to AI models that are neither generative nor agentic. The framework is therefore more applicable to generative and agentic AI.
Where does vendor AI fit in the components?
Component 2.9 covers third-party and vendor risk, and vendors also come up in data (2.6), privacy (2.7), governance (2.4) and monitoring (2.12). Section 2.4 says accountability for managing the bank's risk generally remains with the bank even when vendors build or host the AI.
- Super Intelligence: Supervisory Framework (full text, PDF) — Office of the Comptroller of the Currency, Nov 4, 2026 (occ.gov) · our page on OCC Bulletin 2026-54
- OCC Bulletin 2026-54: Super Intelligence: Supervisory Framework — Office of the Comptroller of the Currency, Nov 4, 2026 (occ.gov) · our page on OCC Bulletin 2026-54
- OCC Bulletin 2026-13: Model Risk Management: Revised Guidance — Office of the Comptroller of the Currency, Apr 17, 2026 (occ.gov) · our page on OCC Bulletin 2026-13
- OCC Bulletin 2023-17: Third-Party Relationships: Interagency Guidance on Risk Management — Office of the Comptroller of the Currency, Jun 6, 2023 (occ.gov) · our page on OCC Bulletin 2023-17
When this rule moves, the next morning's brief says so
when one of these regulators moves, the next morning's brief says so · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning