OCC AI Guidance for Community Banks: What Examiners May Ask
Executive Summary
- OCC AI guidance arrived on November 4, 2026: Bulletin 2026-54 and a 22 page supervisory framework examiners may use when assessing a bank’s use of AI, including generative tools and agents. The OCC now calls AI “super intelligence” under Executive Order 14434.
- The framework is a discretionary, nonbinding examiner resource. The OCC says it sets no supervisory expectations for banks, that banks are not expected to use it to assess themselves, and that it creates no enforceable standards.
- It organizes sixteen components into six pillars and offers an illustrative four tier methodology built on seven factors such as consequences, reversibility, autonomy and data sensitivity. Banks are not required to adopt its tiers or scoring method.
- It addresses a gap in dedicated model risk guidance: OCC Bulletin 2026-13 (April 2026) put generative and agentic AI outside its scope. Comments on the interagency third party risk proposal close November 16.
- Our recommendation: a proportionate inventory, clear permitted use boundaries, vendor change controls and evidence that oversight works can help a bank explain how its existing risks are managed.
In this article
- What is the OCC AI guidance for banks?
- Does the OCC AI guidance apply to community banks?
- Why does the OCC call AI “super intelligence”?
- Is the OCC AI guidance binding on banks?
- How might OCC examiners tier your bank’s AI?
- What do the six pillars and sixteen components cover?
- How does the framework fit with model risk and third party guidance?
- What will examiners ask about vendor AI and AI agents?
- How does the OCC framework compare with state and voluntary AI frameworks?
- What should a community bank document before its next OCC exam?
1. What is the OCC AI guidance for banks?
The OCC AI guidance for banks is a 22 page examiner framework, issued with OCC Bulletin 2026-54 on November 4, 2026, that gives OCC examiners a discretionary resource for evaluating a bank’s use of artificial intelligence, including generative AI and AI agents. It organizes its considerations into six pillars, sixteen components and an illustrative four tier risk scale.
The document itself is titled Super Intelligence: Supervisory Framework, and it was announced the same day in OCC News Release 2026-97. It covers AI used in “credit decisioning, payments, fraud detection, customer service, operations, risk management and market activities.” The bulletin is signed by James M. Gallagher, Senior Deputy Comptroller and Chief National Bank Examiner.
Here is the short version for a busy bank CEO. OCC AI guidance for banks, published as OCC Bulletin 2026-54, is a nonbinding supervisory framework for reviewing how national banks and federal savings associations use AI and AI agents. It invites examiners to scale scrutiny to each use case’s consequences, autonomy and human oversight, rather than to the sophistication of the technology.
Reed Smith described it as the first AI specific examiner framework that any federal banking agency has issued. We did not find an equivalent examiner document from the FDIC or the Federal Reserve as of November 12, 2026. That makes the OCC document a useful public view of how a federal examiner may think about a bank’s chatbot, its fraud model or an agent that can move money. A bank may use it as a preparation aid, but an examiner’s concern must remain grounded in the bank’s actual risks and applicable supervisory or legal authority, not in failure to adopt this document.
One detail in a footnote is worth noticing. The OCC says the framework “was prepared with the assistance of SI” and that the content was “reviewed, edited, and approved by a human.” That describes the OCC’s own drafting process; it does not set one universal human review standard for every bank use.
2. Does the OCC AI guidance apply to community banks?
Yes, examiners may use the OCC AI guidance at community banks, but only in proportion to the bank’s size, complexity and risk profile and to the nature and consequences of its AI use cases. The bulletin’s Note for Community Banks says exactly that, and it tells examiners to tailor their work rather than run the full framework at every institution.
The bulletin is addressed to the chief executive officers of all national banks, federal savings associations, and federal branches and agencies of foreign banks. A state chartered bank is not directly in scope, because its federal supervisor is the FDIC or the Federal Reserve together with its state regulator. State banks still have reason to read it, as section 9 explains, because state examiners released their own AI framework in September.
Most insured institutions the OCC supervises are small by asset size. A November 12, 2026 query of the FDIC BankFind API identified 873 active FDIC insured institutions with the OCC as primary federal regulator. Of the 872 with June 30, 2026 asset data in our extract (reported in thousands of dollars), 609 held less than $1 billion (about 70 percent), 788 held less than $6 billion and 812 held less than $10 billion.
That is an insured institution subset, not the OCC’s entire supervisory population, because the bulletin also reaches federal branches and agencies of foreign banks. Asset size is also not the same thing as a formally defined community bank, so treat the $1 billion line as a rough guide to scale.
Those numbers matter for two reasons. First, many banks that may meet this framework run with lean compliance teams. Second, the exam calendar is getting longer for many of them: the 21st Century ROAD to Housing Act raised the asset threshold for the extended 18 month exam cycle from $3 billion to $6 billion, and on September 10, 2026 the federal banking agencies issued an interim final rule implementing it for institutions that meet the other eligibility conditions. Eligibility for an 18 month cycle does not eliminate offsite monitoring or the need to address issues between examinations.
Community banks are already using the technology the framework targets. In the 2026 CSBS Annual Survey of Community Banks, which collected responses from 330 institutions in 35 states between April 15 and July 15, 2026, 53 percent of respondents said they were piloting AI or using it in limited functions, and about 9 percent said it was deployed across multiple functions or integrated into core operations. Another 18 percent reported no active use but plans within two years, 13 percent had no plans and about 8 percent were unsure.
Read those figures with care. CSBS describes the survey as a convenience sample in which respondents self selected, notes that most respondents were state chartered, says respondents did not necessarily answer every question, and says no test of how representative the banks are of the industry was performed.
The CSBS question expressly included vendor provided AI, and that is the key point for a small bank. Among community bankers already engaged with AI, 77 percent said they use it for fraud detection and security, which may include AI embedded in vendor platforms rather than something the bank built. Cornerstone Advisors’ 2026 survey of 416 bank and credit union executives found 49 percent of responding banks had already deployed generative AI. Both are respondent reports, not audited counts.
3. Why does the OCC call AI “super intelligence”?
The OCC uses “super intelligence” because Executive Order 14434, “Inaugurating the Era of Super Intelligence,” signed September 29, 2026 and published October 2, 2026 at 91 FR 63129, directs executive agencies to use “Super Intelligence” and “SI” in place of “Artificial Intelligence” and “AI” in nonstatutory documents. The change is one of vocabulary, not scope.
The order defines “Super Intelligence” as the technologies encompassed by “artificial intelligence” in 15 U.S.C. 9401(3), and the OCC repeats that definition in footnote 2 of the bulletin. In other words, the OCC framework covers the same chatbots, credit models, fraud tools and agents that everyone else calls AI. Nothing in the framework is limited to frontier systems or to anything resembling the science fiction meaning of the term.
The practical consequence is a search and records problem. Your board minutes, vendor contracts, policies and risk assessments will keep saying AI, and federal documents will now say SI. When you map your program to OCC material, write a single sentence in your AI policy stating that “AI” in bank documents has the same meaning as “SI” in OCC issuances, so an examiner never has to ask whether a policy covers the right systems.
We use “AI” in this article because that is what bankers search for and what state law, contracts and frameworks such as the NIST AI Risk Management Framework still use.
4. Is the OCC AI guidance binding on banks?
No, the OCC AI guidance is not binding: the bulletin states that it “does not establish supervisory expectations for banks,” that banks “are not expected to use this document to assess their use of SI,” and that nothing in it sets “enforceable standards or prescriptive requirements.” It is written for examiners, and nearly every sentence uses “may” or “could.”
That does not make it irrelevant. The same news release says the OCC “expects that bank activities are conducted in a safe and sound manner and consistent with applicable laws and regulations,” whether the activity is traditional or uses novel technology. The framework is a lens examiners may use when applying those existing standards to AI, so it is a useful preview of considerations they may raise.
The framework also tells examiners what not to do, which is useful to a community bank. Examiners “should not treat SI as a separate and distinct risk category,” and banks “are not required to define a separate SI risk appetite.” AI risk is meant to flow into the categories a bank already manages, such as operational, compliance and credit risk.
A separate AI program is not required by this framework. Banks may integrate AI controls into existing risk management processes or use a dedicated structure where that better fits their risks. For most community banks, our recommendation is to fold the framework’s considerations into the risk assessments, vendor reviews and board reports you already produce, which can help explain your controls without creating a parallel bureaucracy.
5. How might OCC examiners tier your bank’s AI?
The framework suggests examiners may classify AI use by the consequences of failure, the transfer of agency and the degree of human oversight, rather than by how sophisticated the technology is. The framework gives a blunt example: a simple rules based agent allowed to initiate transactions could present greater risk than a sophisticated model limited to internal drafting help.
Section 3.1 offers an illustrative methodology with seven factors a use case may be assessed against: decision consequence, reversibility, customer and financial impact, autonomy level, data sensitivity, explainability and provenance. The explainability factor asks whether an output can be traced to a rationale sufficient for legal or disclosure obligations, and the framework names adverse action notices as the example. Provenance asks whether the tool is built in house, fine tuned from a third party base model or consumed as a fully external vendor service.
Those factors combine into an illustrative composite tier. The framework does not require every bank to adopt the OCC’s tier labels or scoring method. The table below reproduces the OCC’s four tiers and its examples, and adds our own community bank examples showing where common tools might land depending on configuration. The right column is our interpretation, not OCC text.
| OCC illustrative tier | OCC profile | OCC examples | Possible community bank examples (our reading) |
|---|---|---|---|
| Tier 1: Critical | Direct action, low or no reversibility, high financial or customer impact, autonomous or partially autonomous | Autonomous payment agent; real time credit decisioning; fraud hold release; market access | A vendor fraud engine configured to release holds without staff review; autonomous credit approval with material consequences and limited effective human intervention |
| Tier 2: High | Direct action or high weight decision support, human on the loop, moderate to high impact | AI assisted underwriting; agent initiated customer communications; AML alert disposition; code assistance for internal code or data access changes | Potentially a website chatbot that answers fee and rate questions, depending on account access, autonomy, verification controls and consequences of error; BSA tooling that closes alerts on its own (a tool that only ranks alerts transfers less authority) |
| Tier 3: Moderate | Decision support with mandatory human in the loop, limited data sensitivity, reversible | Internal document summarization for credit memos; agent assisted research | Potentially credit memo drafting that a lender substantively rewrites and signs, where data access, confidentiality and permitted use are constrained; sensitive borrower data may justify higher controls |
| Tier 4: Low | No customer or financial data exposure, no direct action, fully human controlled | Internal productivity tools; code assistance; general employee Q&A | A policy Q&A assistant restricted to public procedures; meeting note drafting with no customer data |
Two of our placements deserve explanation. We suggest a customer facing chatbot could be Tier 2 rather than Tier 4 because the framework’s compliance component singles out hallucinations “concerning fees, rates, account status, or policies” as a source of legal and customer harm exposure. That is not a universal classification: a tightly constrained public information bot and an account servicing agent can have materially different risk profiles.
We put a fraud engine that releases holds in Tier 1 because fraud hold release is one of the OCC’s own Tier 1 examples. That function can be enabled inside a vendor platform without anyone classifying it. Assess the bank’s actual use, configuration, permissions, data, consequences and effective controls, not the vendor’s marketing label.
The framework also describes its illustrative scoring as a way to treat comparable risk comparably “regardless of where it originates in the bank.” If you adopt a tiering method, a deposit operations team and a lending team using similar tools should end up in the same tier for the same reasons. If your inventory shows two similar tools in different tiers, write down why.
GOVERNANCE INSIGHT
Scope drift can count as a control failure even when the answer looks right.
Section 2.1 says activity outside an AI tool’s approved purpose “may be treated as a control failure even when the resulting output or action appears reasonable.” For a community bank, that turns the purpose statement into a control. If the marketing team starts pasting customer complaints into a tool approved only for drafting public newsletters, the problem is the use, not the quality of the newsletter. Write purpose statements narrowly enough that drift is detectable.
6. What do the six pillars and sixteen components cover?
The six pillars of the OCC AI framework are purpose, risk, control, oversight, resilience and assurance, and they group sixteen components that run from a defined purpose for each tool through testing, documentation, internal audit and board reporting. Each component explains how AI may affect an existing control area and lists illustrative considerations examiners may use.
Purpose covers what a tool or agent is permitted to do. Risk covers what happens if its output or action is not fit for purpose. Control gathers human in the loop and human on the loop oversight, model risk management, data governance, privacy, compliance and cybersecurity. Oversight assigns accountability, monitoring and reporting.
Resilience is the most distinctive pillar. The framework says operational resilience for AI “includes both loss of service and incorrect operation,” because “an agentic system could remain available while taking inappropriate or harmful actions.” It lists circuit breakers, rate limits, least privilege permissions and kill switches as ways to contain an agent that is running but wrong, and it suggests a separate customer remediation workstream when a misfire touches customers.
Assurance covers testing, documentation, vendor risk and internal audit. On testing, the framework accepts that a single successful run proves little for a non deterministic system and points to repeated testing, defined samples, adversarial testing, red teaming and, with caveats, using one model to grade another. On human review, it warns that approval should be “more than a nominal click” and that high volumes or bulk approval “could reduce review quality and contribute to alert fatigue.”
The definitions in Appendix A matter for scoping your inventory. The framework’s definition of an “SI system” excludes “any common commercial product within which super intelligence is embedded, such as a word processor or map navigation system.” That exclusion should not be treated as a blanket exemption from privacy, security, vendor risk or acceptable use controls. As a conservative internal scoping choice, a bank may decide to inventory material uses of embedded assistants, such as staff using one on customer data, even where the lexicon’s scope is uncertain.
7. How does the framework fit with model risk and third party guidance?
The OCC AI framework fills a deliberate gap: footnote 1 says the model risk principles in OCC Bulletin 2026-13 apply to nongenerative and nonagentic AI, so the new framework “is more applicable to generative and agentic SI.” Your credit scorecard stays under model risk guidance, while your chatbot and any AI agent are where this framework does most of its work.
OCC Bulletin 2026-13, the revised model risk management guidance issued April 17, 2026 alongside Federal Reserve SR 26-2, said generative and agentic AI models “are not within the scope of this guidance.” It also said it is expected to be most relevant to banking organizations with more than $30 billion in total assets, while it may be relevant to smaller ones, and that it does not set enforceable standards. Our Connecticut article covers how that carve out left a gap in dedicated model risk guidance for generative and agentic AI. Existing privacy, consumer compliance, cybersecurity and vendor risk expectations still applied during that interval.
The second piece is vendor oversight. On September 11, 2026 the OCC, FDIC, Federal Reserve and NCUA proposed replacing the existing interagency third party risk management guidance with a principles based version, published as OCC Bulletin 2026-46. The OCC’s proposed issuances page lists the comment period as closing November 16, 2026, and the OCC says the proposal would not set enforceable standards. Trade associations asked in late October for a 30 day extension; an extension request alone does not change the deadline, so confirm the current date before you rely on it.
Read together, the three documents overlap rather than divide cleanly. The April model risk guidance covers qualifying traditional models and excludes generative and agentic AI. The November framework supplies additional considerations for those newer uses while drawing on relevant model risk principles in its own “Model Risk Management Extensions” component. Hybrid systems may need both analyses, and anything bought from a vendor, which is most AI at a small bank, also runs through third party risk management, which the AI framework says it “may supplement.”
If you plan to comment on the third party proposal, the AI framework gives you a concrete hook. Ask the agencies to address how a bank should get notice when an existing vendor adds AI features to a product already in use, which is exactly the scenario the AI framework flags in component 2.9.
Getting your AI governance in order
Dynamic Comply helps small and medium sized businesses build a real AI governance foundation using the NIST AI RMF and ISO/IEC 42001.
8. What will examiners ask about vendor AI and AI agents?
Examiners using the framework may consider whether the bank, not the vendor, is managing the risk: the framework says that although development or hosting may be spread across vendors, “accountability for managing the bank’s risk generally remains with the bank.” For a community bank that buys almost all of its AI, that sentence is the center of the exam.
Component 2.9 lists four vendor issues. Banks may define the authority, identity, permitted tools, data use and escalation limits of provider agents that touch bank systems. They may define an AI incident vocabulary and notification timelines covering hallucination, misfire, drift and manipulation. They should consider how contracts assign responsibility for model errors and customer harm, including indemnification and caps.
The fourth issue is the one most small banks have not addressed: “Existing vendors may add SI features, functionality or change models connected to products already in use by the bank.” The framework says the bank may require notice of new AI functionality and may reassess or retest before or when changes are made. AI features can arrive at a community bank through a release note rather than a new procurement decision.
The privacy component adds a contract question every bank should be able to answer: whether bank data can be used to train or improve models offered to other customers. It also flags prompts, outputs, embeddings and logs as places sensitive data can accumulate, and it suggests access controls and risk based retention for them.
For AI agents specifically, the cybersecurity component treats “the breadth of agent permissions and tool access” as part of the attack surface. It names prompt injection, data poisoning, model extraction and excessive agency, and it says agent identities and credentials may be governed “in a manner comparable to privileged users.” If an agent can log in, give it a named service account, least privilege access and the same access reviews you run for administrators.
Our Illinois article shows why the compliance component matters for credit. The OCC says AI “does not change the bank’s underlying legal obligations,” so when an adverse action notice is required the bank must be able to provide accurate, legally sufficient reasons, and a vendor underwriting model remains subject to applicable federal fair lending law and any applicable state law, whatever tier you assign it.
9. How does the OCC framework compare with state and voluntary AI frameworks?
The OCC framework is the federal counterpart to two state moves in September 2026: the Conference of State Bank Supervisors released its AI Supervisory Framework for state examiners on September 16, and the Connecticut Department of Banking issued AI governance guidance on September 22. These are different supervisory instruments. The OCC and CSBS documents are examiner resources, although CSBS says its framework “doubles as a resource for industry.” Connecticut’s guidance is directed to regulated entities and expressly states governance and examination expectations. None creates a new penalty, and all emphasize proportionality.
| Document | Issued | Who it reaches | Binding? | Generative and agentic AI |
|---|---|---|---|---|
| OCC Bulletin 2026-54 framework | November 4, 2026 | OCC examiners reviewing national banks, federal savings associations and federal branches | No; examiner resource with no enforceable standards | Primary focus |
| OCC Bulletin 2026-13 model risk guidance | April 17, 2026 | Banking organizations, most relevant above $30 billion in assets | No; no enforceable standards | Expressly outside scope |
| CSBS AI Supervisory Framework | September 16, 2026 | State examiners reviewing state chartered banks and nonbank licensees | No; examiner resource that doubles as an industry resource; each state decides how far to adopt it | Covered, including agents |
| Connecticut DOB AI guidance | September 22, 2026 | Connecticut chartered institutions and Department licensees | Guidance directed to regulated entities; states governance and examination expectations; no stated penalty | Covered where AI is material |
| NIST AI RMF and ISO/IEC 42001 | January 2023 and December 2023 | Any organization that adopts them | Voluntary | Covered; technology neutral |
The overlap is good news for a small bank with limited staff. The OCC’s tier factors (consequence, reversibility, autonomy, data sensitivity, explainability, provenance) line up closely with the factors in the CSBS tiering worksheet and with the risk mapping steps of the NIST AI RMF. A common inventory and control library can support multiple frameworks, with a documented crosswalk for differences in scope, definitions, tiering and supervisory expectations.
CSBS says state regulators supervise 79 percent of U.S. banks, so a community banking market almost always contains both kinds of charter. CSBS also says each state agency will determine how far to incorporate its framework into supervision, so adoption is not automatic in every state. A holding company with a national bank subsidiary and a state chartered affiliate can usually run one program with a crosswalk rather than two separate programs.
An ISO/IEC 42001 management system can serve as that backbone, because it already requires an AI policy, defined roles, risk assessment, impact assessment and internal audit. It is not a substitute for the OCC’s examiner questions, and certification creates no safe harbor. It does give you documented, audited processes that an examiner can test, which is exactly the “assurance” pillar the OCC describes.
10. What should a community bank document before its next OCC exam?
Our recommendation is that before its next exam, a community bank should be able to show an AI inventory with a written purpose statement for each use case, a documented tier with the reasons for it, named owners and reviewers, vendor change controls, and evidence that human review and monitoring actually operate. None of that is required by the framework, but it maps to considerations the framework says examiners may raise.
The framework’s own inventory fields are a useful template. Component 2.1 suggests an inventory may capture the use case, owner, provider, permissions, deployment status, monitoring and failover arrangements. Add the tier, the date of the last review and whether the tool touches customer data, and you have a register most examiners will recognize immediately.
Here is a practical 90 day plan for a bank under $1 billion. It is our recommendation and a defensible baseline, not a safe harbor.
Days 1 to 30: find everything. Ask every department head and your top ten vendors which products include AI features, including features that were switched on by an update. Record each use case with its owner, provider and data. Apply the word processor exclusion carefully and write down what you left out and why.
Days 31 to 60: tier and fix the top. Score each use case on the seven OCC factors and assign a tier. For anything at Tier 1 or Tier 2, confirm who can override or halt it, test what happens when it is wrong, and add a kill switch or fallback process where none exists. Review customer facing chatbots for any ability to quote fees, rates or account status.
Days 61 to 90: contracts, reporting and audit. Ask key vendors for written notice of new AI features and model changes, confirm whether your data trains shared models, and agree an AI incident definition. Send the board a one page report listing material use cases, tiers, incidents and open remediation. Ask internal audit or an outside reviewer to test one Tier 1 or Tier 2 control end to end.
Keep the proportionality message in mind. The framework says documentation “may be scaled to complexity and consequence without creating incentives to select a less suitable system solely because it is easier to document.” A bank whose AI is limited to Tier 3 and Tier 4 tools needs a short register and a sensible policy, not a binder. Our RIA policy article shows how a short written AI policy can work when use is modest, although a policy written for SEC examiners does not necessarily satisfy an OCC examiner.
Talk to Ross about your AI governance needs
Every business has different AI governance requirements. Let us talk about yours.
Frequently Asked Questions
What is the OCC AI guidance for banks?
It is a 22 page supervisory framework, sometimes called the OCC super intelligence framework, issued with OCC Bulletin 2026-54 on November 4, 2026 that gives OCC examiners a discretionary resource for evaluating a bank’s use of AI, which the OCC now calls super intelligence. It groups sixteen components into six pillars and offers an illustrative four tier risk scale based on consequences, autonomy and human oversight.
Does the OCC AI guidance apply to community banks?
Examiners may use it at community banks, but the bulletin tells them to tailor the review to the bank’s size, complexity and risk profile and to the consequences of each AI use case. It reaches national banks, federal savings associations, and federal branches and agencies; a November 12, 2026 FDIC BankFind query found 873 active FDIC insured institutions with the OCC as primary regulator, and 609 of the 872 with June 30, 2026 asset data held less than $1 billion.
Is the OCC AI guidance binding on banks?
No. The OCC says the framework does not establish supervisory expectations for banks, that banks are not expected to use it to assess their own AI, and that it does not set enforceable standards or prescriptive requirements. Banks remain subject to existing safety and soundness standards and applicable law, and the framework shows how examiners may apply those standards to AI.
How does the OCC tier AI use cases?
The framework offers an illustrative methodology under which a use case may be scored on seven factors: decision consequence, reversibility, customer and financial impact, autonomy, data sensitivity, explainability and provenance. Those roll up into four illustrative tiers, from Tier 1 critical, such as an autonomous payment agent or fraud hold release, to Tier 4 low, such as internal productivity tools with no customer data. Banks are not required to adopt the OCC’s tier labels or scoring method.
Why does the OCC say super intelligence instead of AI?
Executive Order 14434, signed September 29, 2026 and published at 91 FR 63129 on October 2, 2026, directs agencies to use “Super Intelligence” and “SI” in place of “Artificial Intelligence” and “AI” in nonstatutory documents. The term keeps the same meaning as artificial intelligence in 15 U.S.C. 9401(3), so the framework covers ordinary AI tools.
What should a community bank do before its next OCC exam?
Our recommendation is to build an AI inventory that records each use case’s purpose, owner, provider, permissions, monitoring and fallback, then assign a risk tier using the seven OCC factors or a comparable method. Add vendor notice terms for new AI features, confirm human override for higher risk uses, and report material use cases to the board. The framework does not require these steps, but a 90 day effort can produce a credible baseline.
About the author
Ross J. is the founder of Dynamic Comply, an AI governance, compliance, and cybersecurity consulting firm based in Leesburg, Virginia. He brings more than 15 years of federal cybersecurity experience across the Department of State, the Department of Defense, and the Department of Homeland Security, and holds the CGRC certification along with credentials as a GSDC AI Compliance Lead Implementer and Auditor and Certified Ethical Hacker.
Related articles
This article is provided for general informational purposes and reflects the state of the law as of November 2026. It is not legal advice. Regulations in this area are changing quickly. Confirm current requirements and consult qualified counsel before making decisions for your organization.