OCC Bulletin 2026-13, the April 2026 model risk guidance, covers traditional and non-generative AI models and leaves generative and agentic AI out of scope. The OCC's November 2026 examiner framework (Bulletin 2026-54) fills that gap: section 2.5 says model risk principles may be instructive for AI, adapted to non-determinism, constant change and vendor models, and sections 2.12 to 2.14 describe monitoring, testing (including adversarial testing and red teaming) and records that let a bank rebuild an AI decision. The framework is a resource for examiners and does not set expectations for banks.
TRACKED DOCUMENTS: OCC Bulletin 2026-54 · OCC Bulletin 2026-13
What is model risk, in plain words?
Model risk is the chance that a bank makes a bad decision because a model gave a wrong answer, or because staff used a right answer for the wrong job. Model risk management is the set of checks a bank runs so that does not happen: build the model carefully, have someone independent check it, and watch it once it is in use.
The OCC's framework defines model risk in its lexicon (Appendix A) as “the potential for adverse consequences from decisions based on incorrect or misused model outputs and reports.” It adds that model risk can build up across many models that share the same data or assumptions. A few more terms come up on this page:
- Validation means confirming, with objective evidence, that a model meets the requirements of its intended use. In practice, it is often done by people who did not build the model.
- Monitoring means watching the model after launch to see that it keeps working.
- Generative AI is AI that writes text or creates images, audio or other content. Agentic AI is AI that can make decisions and take actions with little or no human help.
One note on words. The OCC uses the term “Super Intelligence (SI)” for what everyone else calls AI. This page says AI. For the history of model risk rules in US banking, read our SR 11-7 guide and the model risk management hub.
What does OCC Bulletin 2026-13 cover, and what does it leave out?
Bulletin 2026-13 is the model risk guidance the OCC, the Federal Reserve and the FDIC issued on April 17, 2026. It covers models in the traditional sense, and it says generative AI and agentic AI models “are not within the scope of this guidance.”
What the bulletin says about its own reach:
- It replaces the 2011 guidance (OCC Bulletin 2011-12) and rescinds the “Model Risk Management” booklet of the Comptroller's Handbook. The Federal Reserve issued the same guidance as SR 26-2.
- A “model” is a “complex quantitative method, system, or approach” that turns input data into quantitative estimates. Simple spreadsheet arithmetic is excluded, and so are deterministic rule-based processes with no statistical, economic or financial theory behind them. Deterministic means the same input always gives the same output.
- It is expected to be most relevant to banks with more than $30 billion in total assets, and sometimes to smaller banks with heavy model use.
- It does not set “enforceable standards or prescriptive requirements.”
- Generative and agentic AI models are out of scope because they “are novel and rapidly evolving.” The agencies said they plan a request for information on model risk and banks' use of AI, including generative and agentic AI.
The framework in Bulletin 2026-54 picks up from there. Its footnote 1 says the Bulletin 2026-13 principles apply to non-generative and non-agentic AI models, so the framework “is more applicable to generative and agentic SI.” Put simply, a credit scoring model built with machine learning stays under Bulletin 2026-13. A chatbot that answers customers falls under the framework's lens. The framework is a resource for examiners: Bulletin 2026-54 says it “does not establish supervisory expectations for banks.”
How does the framework extend model risk practice to generative and agentic AI?
Section 2.5 says some model risk principles “may be instructive” for AI, but the scope, frequency and independence of the work may be adapted to the use case. It then names four points where AI breaks the usual assumptions.
The starting point is this sentence: some AI systems “may not conform to more traditional assumptions of model stability, reproducible outputs, a single output type, or directly traceable decision logic.” Section 2.5 lists the factors that may shape how much checking a use case gets: its potential consequences, its rate of change, its non-determinism (the same question can get different answers) and its reliance on third-party or continuously updated models. The four examples:
| Consideration (section 2.5) | What the framework says | Put simply |
|---|---|---|
| Continuous change and drift | Adaptive models, frequent retraining, prompt changes or vendor updates could alter behavior between periodic reviews and may justify more frequent monitoring or re-validation. | Drift means a model's behavior slowly moves away from what was checked. A yearly review can miss a change that happened last Tuesday. |
| Explainability limits | Explainability tools may help identify factors linked to an output, but “they do not reveal the SI's reasoning and provide a verified source.” Examiners may evaluate whether the explanation method suits the model, the use case and the bank's obligations. | Explainability means being able to say why the system gave an answer. A tool that ranks factors is a clue, not the real reason. |
| Non-determinism and reproducibility | Repeated testing, defined samples, distributional measures and consistency thresholds may give more useful evidence than a single successful run. | One good answer proves little. Ask the same thing many times and measure how much the answers vary. |
| Comparative testing | Traditional champion-challenger methods may not fit open-ended generative outputs. Alternative benchmarks, reference sets, expert review or controlled comparisons may be considered. | Champion-challenger means running a rival model beside the live one to compare scores. That is hard when the output is a paragraph, not a number. |
The framework's summary table names the risk section 2.5 is aimed at: “Point-in-time validation that does not address non-determinism, continuous change, or new output types.” Section 1.1 adds that higher-consequence use cases may warrant “more independent validation.” Our risk tiers page explains how consequence is judged.
How is checking a generative AI model different from checking a traditional model?
A traditional model gives the same number every time and can be scored against what actually happened. A generative model can give different answers to the same question, can make things up, and can change when a prompt or a vendor's model changes.
The table sets the two side by side. The AI column follows the framework. The traditional-model column is our own summary of the contrast. Footnote 4 of the framework adds a caution: when an AI system gives an output that can be checked objectively, such as a default prediction, traditional testing methods and metrics may still be appropriate.
| Point | Traditional model | Generative or agentic AI | Source |
|---|---|---|---|
| Which OCC text applies | Bulletin 2026-13, including non-generative machine learning | Outside Bulletin 2026-13. The framework is “more applicable” here | Bulletin 2026-13; framework footnote 1 |
| Same input, same output | Yes, when the model is deterministic | Not always. Output variability is measured over many runs | 2.5, 2.12, lexicon |
| Kind of output | A single output type, often a number | Open-ended text or content, sometimes actions | 2.5 |
| Comparing against a rival | Champion-challenger | Benchmarks, reference sets, expert review, controlled comparisons | 2.5 |
| Checking an answer | Compare with a known result | Often no verified source. Expert review and sampling may be needed | 2.13 |
| Typical new failure | A wrong estimate (our framing) | Hallucination: fluent but fabricated output | 2.13, lexicon |
| What triggers a re-check | A new model release or a periodic review | Prompt, configuration, vendor version, fine-tuning, data, tool or permission changes, even with no release | 2.5, 2.12 |
| Explaining a decision | Decision logic can be traced | Decision paths may not be directly traceable. Records rebuild what happened | 2.5, 2.14 |
Two words in the table need a definition. A prompt is the instruction or question given to the AI. The lexicon defines a hallucination as output that is “erroneous or flawed but is still in the form of a convincing narrative or presentation.” The lexicon also says generative AI “can still produce flawed information even if underlying data is free of defects.”
How is an AI system monitored after it goes live?
Section 2.12 says testing before launch shows how a system works in a controlled setting, while monitoring shows how it behaves in real use as data, prompts, permissions, providers and models change. Metrics, thresholds and escalation may be set according to the use case's consequences, autonomy, expected variability and rate of change.
Section 2.12 also says monitoring and “in-line real time observability” may cover both average performance and rare events that could cause material harm. Its five examples:
- Accuracy and output quality. Metrics may include accuracy, precision and recall, task completion, and rates of hallucination, abstention (the system declines to answer), complaints, overrides, exceptions and remediation.
- Population and input drift. A change in who uses the system, or in what they ask, can be an early sign of decline before outcome metrics show it.
- Behavioral consistency. For non-deterministic systems, the bank may sample repeated or similar inputs to see whether variation stays within accepted ranges.
- Vendor and model changes. Monitoring may catch upstream version changes, prompt or configuration changes, fine-tuning, data changes, or new tools and permissions that change behavior “without a traditional model release.”
- Feedback loops and escalation. The bank may watch whether outputs feed into future training data, and may set thresholds that trigger review, re-validation, reduced autonomy or suspension.
For example, a branch assistant that answers staff questions about fee schedules could track how often staff override its answers. A rising override rate after a vendor update is the kind of signal section 2.12 describes. The summary table names the risk: “Production behavior drifting from the validated baseline without timely detection or escalation.” Who sees these reports is covered on our board, audit and reporting page.
How is generative AI tested, including red teaming?
Section 2.13 calls testing “central to managing risk from the use of SI” and says AI may need methods that go beyond confirming a single expected result. It names adversarial testing, red teaming, expert review, LLM-as-a-judge and hallucination testing.
The reasons given are the lack of a verified source, non-determinism, constant redevelopment and the high volume of inputs and outputs. How much testing a use case gets may depend on its materiality and the consequences of failure. Lower-materiality uses may warrant limited testing. Consequential or highly autonomous uses may warrant broader testing of both performance and controls.
Testing across the life of the system
- Design and development: requirements, data quality, architecture, permissions, known limits and whether planned controls address the risks.
- Before launch: functional, comparative, scenario, stress, adversarial, compliance, resilience, human-oversight and acceptance testing.
- In use: production sampling, monitoring, incident analysis, change testing and re-validation to find drift and new failure modes.
The AI-specific methods section 2.13 names
- Subjective reviews. When there is no verified answer to compare with, expert review can judge quality, completeness and consistency. Reviewer skills, sampling and the number of observations may be set by the rate of bad outcomes the bank is trying to detect.
- Adversarial testing. Testers may feed in misleading, malicious or unusual inputs to expose hidden weaknesses, unsafe behavior or gaps in prompt, data and tool controls.
- Red teaming. A step beyond adversarial testing: it checks whether the bank can detect, contain, respond to and recover from realistic attacks or misuse. Outside experts may help when internal skills are limited.
- LLM-as-a-judge. One model scores another model's outputs. The bank may consider the judge model's validation, independence, position or verbosity bias (put simply, favoring an answer for its place in a list or its length), shared failure modes and the need for human review.
- Hallucination testing. Testing could measure hallucination under normal and adversarial prompts, check sources, assess when the system declines to answer, define which outputs need human checking, and sample live outputs. The bank may set responses for outputs that cannot be verified and revisit them when prompts, context windows, data or model versions change.
In practice, a test of a customer chatbot might ask the same fee question 200 ways, including tricky phrasings, and count wrong answers. Attacks such as prompt injection, where hidden instructions in an input hijack the system, are covered on our resilience and cybersecurity page.
What records make an AI decision traceable?
Section 2.14 says records may let a bank rebuild a material output or action even when the model cannot give a rule-based reason. The records it names are inputs, outputs, context, retrieved sources, tool calls, approvals, overrides and actions.
Section 2.14 explains why this is harder for AI: outputs are probabilistic, decision paths may not be directly traceable, and small changes may materially alter behavior. Documentation, it says, “enables individuals other than the system's developers to understand, evaluate, and audit its design and operation.” Its four examples:
- Decision traceability: the inputs, outputs, context, retrieved sources, tool calls, approvals, overrides and actions behind a material decision.
- Version and change control: changes to the model, prompt, fine-tuning, data, configuration, provider, tool and permissions may be logged and linked to testing and approval evidence.
- Retention and access: records may be kept as long as legal, supervisory, audit, customer-remediation and operational needs require, while protecting sensitive information.
- Proportionality: documentation may scale with complexity and consequence, without pushing a bank toward a less suitable system just because it is easier to document.
The documentation itself may cover purpose, ownership, design, data, providers, versions, prompts, permissions, assumptions, testing, validation, monitoring, decisions, approvals, incidents and changes. For example, if an AI agent refunded a fee in error, these records would show what the customer wrote, which policy page the agent retrieved, which tool it called, and whether anyone approved it. The summary table names the risk this section targets: “Inability to reconstruct material decisions, changes, tool calls, approvals, or system behavior.” Section 2.7 notes that prompts and logs can hold sensitive data, which our data and privacy page covers.
Why does it matter when the AI model belongs to a vendor?
Many generative AI tools run on a foundation model, a large model trained by a provider and adapted to many tasks. Sections 2.5 and 2.12 treat vendor updates as one of the ways behavior can change between reviews.
Section 2.5 lists “reliance on third-party or continuously updated models” as a factor that may shape the scope, frequency and independence of model risk work. Section 2.12 says monitoring may identify upstream version changes. Put simply, a bank can test a chatbot in March and find it behaves differently in May because the provider changed the model underneath. The full treatment of providers, contracts and provider agents is on our vendor and third-party AI page. For every section of the framework, read the full Bulletin 2026-54 reference or the OCC AI regulation tracker.
Does Bulletin 2026-13 apply to ChatGPT-style tools at my bank?
No. Bulletin 2026-13 says generative AI and agentic AI models are not within its scope. The OCC's framework in Bulletin 2026-54 says it is more applicable to those systems. Machine-learning models that are not generative or agentic stay under Bulletin 2026-13.
Does the OCC framework require banks to validate generative AI?
No. Bulletin 2026-54 says the framework does not establish supervisory expectations for banks. Section 2.5 says model risk principles may be instructive for AI and that the scope, frequency and independence of validation may be adapted to the use case. It describes what examiners may look at, not rules for banks.
Why is one successful test not enough for a generative AI system?
Generative AI can give different answers to the same question. Section 2.5 says repeated testing, defined samples, distributional measures and consistency thresholds may give more useful evidence than a single successful run. Put simply, you measure how often it gets things right, not whether it got it right once.
What is the difference between adversarial testing and red teaming?
Section 2.13 describes adversarial testing as using misleading, malicious or edge-case inputs to expose weaknesses. Red teaming goes further: it checks whether the bank can detect, contain, respond to and recover from realistic attacks or misuse. The framework says outside expertise may help when internal skills are limited.
What should be logged so an AI decision can be explained later?
Section 2.14 names inputs, outputs, context, retrieved sources, tool calls, approvals, overrides and actions as records that may help rebuild a material decision. It also says changes to the model, prompt, fine-tuning, data, configuration, provider, tools and permissions may be logged and linked to testing and approval evidence.
What does the OCC suggest monitoring for once an AI tool is live?
Section 2.12 lists accuracy and output quality, including hallucination, abstention, complaint and override rates, plus drift in who uses the system and what they ask. It also lists consistency of answers, vendor and prompt changes, and feedback loops. Thresholds may trigger review, re-validation, reduced autonomy or suspension.
Can one AI model be used to check another?
Yes, section 2.13 describes this as LLM-as-a-judge. It says the bank may consider the judge model's own validation and independence, its biases such as position or verbosity bias, failures it shares with the model being checked, and the need for human review.
- 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
- SR 26-2: Revised Guidance on Model Risk Management — Federal Reserve Board, Apr 17, 2026 (federalreserve.gov) · our page on SR 26-2
- OCC Bulletin 2011-12: Sound Practices for Model Risk Management (rescinded archive) — Office of the Comptroller of the Currency, Apr 4, 2011 (occ.gov) · our page on OCC Bulletin 2011-12
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