From Sidhant Tamrkar | Product & Market Analysis

SR 26-2 Left Generative AI Out of Scope: The 6 Rules Banks Still Have to Follow

On this page

On 17 April 2026 the OCC, the Federal Reserve and the FDIC replaced the 2011 model risk framework with SR 26-2, and one sentence in it removed generative and agentic AI from scope. Banks running financial services AI on large language models now have no model risk guidance written for them. They still answer to at least 6 rules that reach those systems, and supervisors have started asking agent questions anyway.

Key takeaways

  • SR 26-2 excludes generative and agentic AI by name. OCC Bulletin 2026-13 states twice that these models "are not within the scope of this guidance", and the promised request for information had not appeared 207 days later.
  • The gap sits in guidance, not in law. Third-party risk guidance, heightened standards, information security standards, the 36-hour incident rule and Regulation B all apply to an AI deployment by function, whatever the technology.
  • Supervisors are asking about controllability, not accuracy. Reuters reported in October 2026 that the Fed, OCC and FDIC asked banks about agent limits, approvals, audit trails, human involvement and how to shut systems down.
  • The cheapest defence is a written answer to each question. A permission register, a decision log, a tested shutdown runbook and a provenance inventory cover most of what the agencies have asked so far.
207 daysSince the agencies promised a request for information on AI model risk, with none published as of 10 November 2026. Source: OCC Bulletin 2026-13, Zan Digital count.
8 questionsAsked of large, mid-size and smaller US banks about AI agents, per six people cited by Reuters, 13 October 2026.
$30BAsset size above which SR 26-2 is expected to be most relevant. Source: Federal Reserve SR 26-2, April 2026.

The short answer

SR 26-2 does not cover generative or agentic AI, and no replacement guidance exists yet. Banks still follow third-party risk guidance, heightened standards above $50 billion, information security standards, the 36-hour incident notification rule and Regulation B. The practical move is to document answers to the agent questions supervisors are already asking.

What SR 26-2 changed, and the sentence that removed generative AI

SR 26-2 is the Federal Reserve's version of the revised guidance. The OCC published the same text as Bulletin 2026-13 and the FDIC as FIL-15-2026, all dated 17 April 2026. Together they replaced the 2011 guidance banks knew as SR 11-7.

The OCC bulletin rescinded four documents. They are the Model Risk Management booklet of the Comptroller's Handbook, Bulletin 1997-24 on credit scoring models, Bulletin 2011-12, and Bulletin 2021-19 on models supporting BSA/AML compliance. The Fed letter supersedes SR 11-7 and SR 21-8.

The exclusion appears twice in the OCC text, in the Highlights and again in the Background. It reads: "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance."

The definition of a model does the rest of the work. The OCC defines it as "a complex quantitative method, system, or approach that applies statistical, economic, or financial theories to process input data into quantitative estimates". A chatbot drafting a customer reply produces text, not a quantitative estimate.

The OCC says it "does not set forth enforceable standards or prescriptive requirements" and that non-compliance "will not result in supervisory criticism". It is expected to be most relevant to banks above $30 billion in assets.

My reading is that the exclusion is a deferral, not a permission. The agencies did not say generative AI carries no model risk. They said it was too new to write principles for. How the old validation method still maps onto an LLM is covered in the SR 11-7 validation file for an LLM. This post maps everything else a bank still owes.

How bank AI oversight got here, 2011 to November 2026 Events are evenly spaced, not to scale. Dark markers are guidance or rules, light markers are reporting. Apr 2011 SR 11-7 Jun 2023 Third-party guidance May 2025 CFPB AI circulars pulled Dec 2025 $700B heightened standards proposal 17 Apr 2026 SR 26-2: GenAI out of scope Jun 2026 Reuters: map AI use Sep 2026 New third-party guidance proposed Oct 2026 Reuters: agent questions Nov 2026 OCC 2026-54 examiner tool Still missing: the interagency request for information on AI model risk promised on 17 April 2026. Sources: OCC, Federal Reserve, FDIC, CFPB, Reuters. Compiled by Zan Digital, 10 November 2026.
The red marker is the only event that narrowed oversight of generative AI on purpose. Every dark marker after it either proposes, examines or reports, and none replaces it.

The generative AI model risk gap is in guidance, not in law

The gap is narrower than most commentary suggests. SR 26-2 is guidance. Removing generative AI from a guidance document does not remove it from any statute, regulation or other guidance that applies by function rather than by technology.

A bank that lends, holds customer data, buys software and runs critical systems is covered by rules about lending, customer data, vendors and outages. An LLM that touches any of those activities inherits the rule. The question is which rules, and what evidence each one expects.

The request for information has not arrived

The OCC bulletin says the agencies "plan to issue in the near future a request for information that addresses model risk management generally". The same sentence says the RFI "considers, in particular, banks' use of AI, including generative AI and agentic AI and AI-based models."

As of 10 November 2026, I found no such RFI. The FDIC's financial institution letters through 3 November and the Fed's press releases through 9 November carry nothing on it. That is 207 days since the promise. An RFI also only asks questions. Any guidance would come after the comment period closes.

What the OCC published instead

On 4 November 2026 the OCC issued Bulletin 2026-54, "Super Intelligence: Supervisory Framework". It uses the term super intelligence for the technologies previously called artificial intelligence under 15 USC 9401(3), citing Executive Order 14434 of October 2026. It is an OCC document, not an interagency one.

It is written for examiners and disclaims expectations outright: "This supervisory framework does not establish supervisory expectations for banks." It tells examiners not to treat super intelligence "as a separate and distinct risk category". It also says a bank's control intensity "could increase with the potential consequences of an SI agent's failure or misuse".

The law firm Baker Donelson describes six pillars in the framework (purpose, risk, control, oversight, resilience and assurance) and four illustrative risk tiers. I could not confirm those in the bulletin page text, so treat them as one firm's summary of the attached document.

The most useful phrase in the bulletin is quieter. It says the framework helps examiners identify considerations that "supplement existing model risk management, governance, technology, compliance, and operational-risk practices". That points straight back at the rules banks already follow, which is where the rest of this post goes.

What supervisors are asking banks about AI agents

Reuters reported on 13 October 2026 that the Fed, OCC and FDIC were questioning large, mid-size and smaller US banks about AI. The story cites six people familiar with the matter. All three agencies declined to comment.

Two of those people described the agent questions. Supervisors asked "what limits are placed on systems" and how actions are "monitored and approved". They asked whether banks can "clearly track and audit AI agent decisions" and "what safeguards exist to prevent agents from exceeding their authority".

They also asked whether "humans remain involved in key decisions" and "how banks can shut down autonomous systems". Three people said supervisors asked whether banks use Chinese-developed technology and "AI agents in cyber defence". That makes 8 distinct questions.

This was not the first round. An earlier Reuters story, dated 13 June 2026, reported that the OCC and Fed had asked banks to map AI use in higher-risk areas such as lending, KYC and sanctions screening. The October questions moved from where AI sits to whether it can be stopped.

Notice what the list leaves out. None of the 8 questions asks about hallucination rates, benchmark scores or model accuracy. Every one is about control: limits, approvals, audit, authority, human involvement, shutdown and provenance.

I would treat these 8 questions as the exam. They are information gathering, and nobody has called them requirements. But an examiner does not need an AI guidance document to find a weakness in operational risk management. A bank that cannot answer in writing has already handed one over.

Six rules that still reach financial services AI in a bank

Each row is a rule or guidance document in force on 10 November 2026 that applies to a generative AI deployment because of what the deployment does, not what it is.

Six obligations that still reach a bank AI deployment, status at 10 November 2026.
ObligationCitationWho it coversWhat it asks of an AI deploymentStatus
Third-party risk managementOCC Bulletin 2023-17, SR 23-4All banks, scaled to risk.Due diligence, contract terms, monitoring and exit plan for a vendor model.In force. Replacement proposed 11 September 2026
Heightened standards12 CFR Part 30, Appendix DOCC banks at $50B and aboveAI risk inside the risk governance framework, with independent risk management and audit.In force at $50B. Rise to $700B proposed
Information security standards12 CFR Part 30, Appendix BAll OCC banksSafeguards for customer information in prompts, retrieval and agent access.In force
Incident notification12 CFR 53.3All banksNotify the regulator within 36 hours of determining a notification incident.In force
Adverse action reasons12 CFR 1002.9 (Regulation B)CreditorsSpecific principal reasons for a denial, whatever made the decision.Rule in force. CFPB AI circulars withdrawn
Model risk guidanceSR 26-2, OCC Bulletin 2026-13Most relevant above $30BCovers the quantitative models inside an AI pipeline.In force. No criticism for non-compliance
Which existing rule reaches which part of an AI agent Zan Digital mapping from the rule texts. Dark = applies directly, light = applies in part, blank = little or no reach. Vendor model Customer data Agent actions Credit decisions Governance Third-party guidance (2023) Heightened standards ($50B+) Information security (App. B) 36-hour incident rule Regulation B reasons SR 26-2 (quantitative parts) Analytical mapping, not an agency position. Each column has at least one dark cell, so no part of an agent is unregulated.
Read down the columns. Every part of an agent deployment has at least one rule reaching it directly, and agent actions depend most heavily on a rule written for cyber incidents.

Third-party risk management covers the vendor model

Most bank LLM deployments run on a model the bank did not build. That makes the provider a third party under the interagency third-party guidance issued on 6 June 2023. Its life cycle runs through "planning, due diligence and third-party selection, contract negotiation, ongoing monitoring, and termination."

Contract negotiation is where the AI-specific work lands. It is where a bank secures notice of model version changes, audit rights, data use terms and an exit route. A provider that changes the model behind an API has changed a component of your control environment.

This row will move. On 11 September 2026 the OCC, Fed, FDIC and NCUA proposed new third-party guidance, and the FDIC states that "any finalized guidance would replace the 2023 Guidance". Until a final version lands, the 2023 text governs.

Heightened standards put AI inside the risk governance framework

OCC banks with average total consolidated assets of $50 billion or more must meet the heightened standards in 12 CFR Part 30, Appendix D. They require a risk governance framework, defined front line units, independent risk management and internal audit. The appendix never mentions AI, and it does not need to.

An agent that moves money or answers customers creates operational and compliance risk. Under Appendix D, that risk needs an owner in the front line, a challenger in independent risk management and coverage in the audit plan. A bank that keeps its agents in an innovation lab outside that structure has a framework gap, not an AI gap.

In Bulletin 2025-51, dated 23 December 2025, the OCC proposed raising the threshold from $50 billion to $700 billion, a 14-fold increase. Comments closed on 2 March 2026. I found no final rule, and the codified text still says $50 billion.

Information security standards follow customer data into the prompt

The Interagency Guidelines Establishing Information Security Standards, at 12 CFR Part 30, Appendix B for OCC banks, implement section 501(b) of the Gramm-Leach-Bliley Act. They require safeguards for customer information, and they remain in force.

Generative AI creates new places for customer information to sit. A prompt with account details, a retrieval index built on customer files and an agent with read access to core systems are all inside the guidelines' reach. Five controls banks already run on data, and how they transfer to agents, are set out in the bank data controls worth copying for AI agents.

The 36-hour clock applies to an agent failure

The computer-security incident notification rule sets the deadline. A bank must notify its primary regulator "as soon as possible and no later than 36 hours after the banking organization determines that a notification incident has occurred". The OCC version sits at 12 CFR 53.3, with matching Fed and FDIC rules.

The rule is technology-neutral. If an agent failure rises to a notification incident under the rule's definitions, the clock starts at the bank's determination, not when the vendor confirms a cause. That is why a shutdown plan written in advance matters, and an on-call runbook for production agents is the cheapest way to have one.

Regulation B still demands specific reasons

For credit, the oldest rule may matter most. 12 CFR 1002.9(b)(2) says adverse action reasons "must be specific and indicate the principal reason(s) for the adverse action." A model's complexity is not an exception written into the rule.

The CFPB withdrew its circulars 2022-03 and 2023-03, which applied that rule to complex algorithms and AI, effective 12 May 2025. The withdrawal notice said it "is not necessarily final." A withdrawn circular removes an interpretation. It does not remove the regulation the circular interpreted.

SR 26-2 still covers the quantitative parts

Agent workflows often contain a classical model: a fraud score, a credit score or a ranking model that decides which case an agent sees first. Those components remain models under SR 26-2, even when an LLM wraps them.

So the inventory has to split components rather than label whole systems. I would list every scoring model inside an agent pipeline separately and keep it under the model risk programme.

Asset thresholds decide which bank AI rules bite

Three of the six obligations carry an asset test, and the tests do not line up. SR 26-2 is expected to be most relevant above $30 billion, though the OCC says it may apply below that where model risk is significant. Heightened standards start at $50 billion, and the pending proposal would move that line to $700 billion.

Where each rule starts, by total assets Thresholds in $ billions, linear scale. Two rules in the map have no asset threshold at all. SR 26-2 relevance $30B Heightened standards today $50B Heightened standards, prop… $700B 36-hour rule, Regulation B No asset threshold. Sources: OCC Bulletin 2026-13, 12 CFR Part 30 Appendix D, OCC Bulletin 2025-51, 12 CFR 53.3, 12 CFR 1002.9.
The red bar is a proposal, not a rule. If it is finalised, a bank between $50 billion and $700 billion keeps its AI exposure and loses the appendix that forces a framework around it.

The practical effect depends on size. A $40 billion bank sits inside SR 26-2's most relevant band but outside heightened standards. A $20 billion community bank sits outside both, yet still owes the incident rule, the information security guidelines, Regulation B and third-party risk management.

No bank escapes the questions either. Reuters reported that smaller banks received them too, and Bulletin 2026-54 says examiners "may use the framework for supervising community banks' SI use". For a small bank, I think the asset tests matter less than they look.

The evidence file for agentic AI bank regulation

The 8 Reuters questions are the closest thing to a published exam for agentic AI. The table pairs each one with the evidence that answers it and the existing rule the same evidence already serves.

Reported supervisor questions on AI agents, the evidence that answers them, and the rule it also serves.
Question, as reportedEvidence to retainExisting rule it also serves
What limits are placed on systemsPermission register per agent: tools, data scopes, transaction and spend caps.Information security standards
How actions are monitored and approvedApproval thresholds and a log of human sign-offs.Heightened standards, front line controls
Can you track and audit agent decisionsDecision log with inputs, model version, tool calls and output.Regulation B reasons, incident rule
Safeguards against exceeding authorityPre-execution checks and test results for refused actions.Information security standards
Do humans remain in key decisionsNamed list of decisions an agent may not finalise.Heightened standards, Regulation B
How you shut down autonomous systemsShutdown runbook and the date of the last tested drill.36-hour incident rule
Use of Chinese-developed technologyModel and component inventory with provenance.Third-party guidance, due diligence
AI agents in cyber defenceInventory of agents in security operations and their permissions.Information security standards

Two rows carry most of the weight. The decision log answers the audit question, supports an adverse action notice and gives incident responders a timeline. A field-level design for it is laid out in the agent audit log specification.

The permission register answers the limits and authority questions together. Most agent permissions are OAuth scopes and service accounts, so a scope review for AI agents produces most of the register as a side effect.

The shutdown row is where I would spend first. A runbook nobody has drilled is a document, not a control. Record the date of each drill, how long the agent took to stop, and what work it left half done.

Where this map is weakest

The six-rule map is my construction. No agency has said these rules are the framework for generative AI, and an examiner may weigh them differently. The matrix in particular is an analytical judgement from the rule texts, not a measured result.

The question list rests on anonymous sourcing. Reuters cites six people, the agencies declined to comment, and the wording is the reporters' paraphrase. Some secondary write-ups list the questions differently, which is a reason to treat the exact phrasing as provisional.

Two rows are moving under the map. The third-party guidance has a proposed replacement, and the heightened standards threshold may rise 14-fold.

The strongest counter-case is about direction. The CFPB withdrew its AI circulars, the OCC proposed lifting the heightened standards threshold, SR 26-2 promises no criticism for non-compliance, and Bulletin 2026-54 sets no expectations. A reasonable reader concludes supervisors will press less on AI, not more.

For a small bank with no agents in production, I think that reading is defensible. It weakens fast once an agent can act. The incident rule, Regulation B and the information security guidelines did not soften, and those are the three that bite when an agent fails.

Frequently asked questions

Does SR 26-2 apply to generative AI?

No. SR 26-2 and OCC Bulletin 2026-13, issued on 17 April 2026, state that generative AI and agentic AI models "are not within the scope of this guidance". The agencies said they plan a request for information covering AI, including generative and agentic AI. Quantitative models used inside an AI system, such as a fraud or credit score, still fall under SR 26-2 as models.

What replaced SR 11-7 for model risk management?

SR 26-2 from the Federal Reserve, published by the OCC as Bulletin 2026-13 and by the FDIC as FIL-15-2026, all dated 17 April 2026. It supersedes SR 11-7 and SR 21-8, and the OCC rescinded its Model Risk Management handbook booklet. The guidance is expected to be most relevant to banks above $30 billion in assets.

What is OCC Bulletin 2026-13?

OCC Bulletin 2026-13 is the OCC's revised model risk management guidance, issued jointly with the Federal Reserve and FDIC on 17 April 2026. It rescinds Bulletin 2011-12, Bulletin 1997-24 and Bulletin 2021-19. It says it sets no enforceable standards, that non-compliance will not result in supervisory criticism, and that generative and agentic AI are out of scope.

Which regulations apply to AI agents at US banks?

No AI-specific bank regulation exists yet, but several apply by function. The 2023 third-party risk guidance covers vendor models. Heightened standards cover OCC banks above $50 billion. Information security standards cover customer data, the 36-hour incident rule covers disruptions, and Regulation B requires specific reasons for credit denials. SR 26-2 still covers quantitative models inside an agent.

What are regulators asking banks about AI agents?

Reuters reported on 13 October 2026, citing six people, that the Fed, OCC and FDIC asked banks about limits on AI systems and how actions are monitored and approved. They also asked whether agent decisions can be tracked and audited, and about safeguards against agents exceeding authority, human involvement, shutdown plans, Chinese-developed technology and AI agents in cyber defence.

Has the AI request for information from bank regulators been published?

Not as of 10 November 2026, based on FDIC letters through 3 November and Federal Reserve press releases through 9 November. The agencies promised it on 17 April 2026 "in the near future". The OCC issued a separate examiner framework, Bulletin 2026-54, on 4 November 2026, which states that it does not establish supervisory expectations for banks.

Where to start this week

Run the 8 questions as a tabletop exercise with your model risk, information security and operations leads in one room. For each agent in production or pilot, write the answer and name the artefact that proves it. Every blank cell is your first work item, and it is the same blank an examiner would find.

Then split your AI inventory by component. List each scoring model inside an agent pipeline under SR 26-2, and list the LLM layer against the other five rules. If one system appears under no rule at all, the map has found the gap you actually have.

Still open

This map will change when the interagency request for information appears, or when either the third-party proposal or the heightened standards proposal is finalised. Re-check the status column of the six-rule table before relying on it.

References

  1. Office of the Comptroller of the Currency, Bulletin 2026-13, Model Risk Management: Revised Guidance, 17 April 2026. Used for the scope exclusion, the RFI promise, the model definition, the $30 billion statement and the rescissions.
  2. Board of Governors of the Federal Reserve System, SR 26-2, Revised Guidance on Model Risk Management, 17 April 2026. Used for the superseded SR letters and applicability.
  3. Reuters via Yahoo Finance, Exclusive: Banking watchdogs ramp up AI scrutiny as stability fears mount, sources say, 13 October 2026. Used for the agent questions and the source count.
  4. Office of the Comptroller of the Currency, Bulletin 2026-54, Super Intelligence: Supervisory Framework, 4 November 2026. Used for the examiner framework quotes.
  5. Office of the Comptroller of the Currency, Bulletin 2023-17, Third-Party Relationships: Interagency Guidance on Risk Management, 6 June 2023. Used for the life cycle stages.
  6. Office of the Comptroller of the Currency, Bulletin 2025-51, proposed heightened standards threshold, 23 December 2025. Used for the $700 billion proposal.
  7. Code of Federal Regulations, 12 CFR 53.3, Notification, in force since 2022. Used for the 36-hour requirement.
  8. Code of Federal Regulations, 12 CFR 1002.9, Notifications (Regulation B). Used for the specific reasons requirement.

Weakest point in the source base: the 8 agent questions come from anonymous sources paraphrased by Reuters, with no agency confirmation. The six pillars attributed to Bulletin 2026-54 come from a law firm summary and were not confirmed in the bulletin page text. Last verified 10 November 2026.

ST
Sidhant Tamrkar
Contributing Analyst, Zan Digital. Writes about AI regulation, enterprise controls and what primary filings and rule texts actually require.

Related reading