← Back to Blog

Your riskiest AI is not your smartest one. The OCC just said so.

By:
Yevgeniy Vahlis
Updated on:
November 18, 2026

SR 26-2, issued in April 2026, replaced SR 11-7 and left generative and agentic AI out of scope. OCC Bulletin 2026-54, issued on November 4, 2026, gives examiners a framework the OCC calls more applicable to exactly those systems. It ranks AI by what happens when it fails, so a simple agent that can move money outranks a sophisticated model that drafts memos. Three things carry the most weight in the exam: a complete inventory, named owners, and a tested shutdown path.

What changed in 2026

In April, regulators took agents out of model risk guidance. On November 4, the OCC put them in the exam.

The April move landed on April 17, 2026. The Federal Reserve, the OCC and the FDIC replaced SR 11-7 with SR 26-2, issued by the OCC as Bulletin 2026-13 and by the FDIC as FIL-15-2026. The guidance states that generative AI and agentic AI models are novel and rapidly evolving, and places them outside its scope.

Many banks read that carve-out as breathing room. The duty to govern those systems stayed in place. It simply lost the dedicated piece of guidance that had been attached to it.

OCC Bulletin 2026-54 is the November move, and the OCC’s first guidance telling examiners how to evaluate a bank’s use of AI, including agentic AI. King & Spalding reads it as an examiner framework rather than a new rule for banks. One quirk worth knowing: the OCC calls the technology “super intelligence” (SI). This post sticks with “AI”.

SR 26-2 itself is expected to be most relevant to banking organizations with more than $30 billion in total assets, and the SR 26-2 explainer from Fiddler AI runs through what carries over. The gap SR 26-2 left is now an exam topic.

Risk follows consequence

OCC Bulletin 2026-54 tells examiners to scale their scrutiny of a bank’s AI to three things: the consequences if it fails or is misused, how much agency it takes away from people, and how much human oversight is left. How technically sophisticated the model is does not drive the ranking.

The framework document names the factors examiners weigh: decision consequence, reversibility, customer and financial impact, autonomy, data sensitivity, permission scope and interdependencies. Its illustrative scale has four tiers:

  • Tier 1, Critical. Direct action, low or no reversibility, high financial or customer impact, autonomous or partially autonomous. Examples: an autonomous payment agent, real-time credit decisioning.
  • Tier 2, High. Direct action or high-weight decision support, human-on-the-loop, moderate-to-high impact. Examples: AI-assisted underwriting, agent-initiated customer communications.
  • Tier 3, Moderate. Decision support with mandatory human-in-the-loop, limited data sensitivity, reversible. Examples: document summarization for credit memos.
  • Tier 4, Low. No customer or financial data exposure, no direct action, fully human-controlled use. Examples: internal productivity tools.

In the framework’s words, “a simple rules-based agent authorized to initiate transactions could present greater risk than a sophisticated model limited to internal drafting assistance.”

That sentence is the argument of this post. Many tiering schemes were built around model complexity, because years of SR 11-7 work trained teams to look there first. Under this framework, the large language model that summarizes credit memos lands in Tier 3, and the plain rules-based agent with payment permissions lands in Tier 1. A scheme that ranks by complexity puts the wrong systems at the top of the list.

SR 26-2 and OCC 2026-54 side by side

Both documents are supervisory guidance. They are written for different readers and ask for different things. The table sets out what each one is and what each one asks for.

RowSR 26-2 (OCC Bulletin 2026-13, Apr 2026)OCC Bulletin 2026-54 (Nov 2026)
Issued byFederal Reserve, OCC and FDIC jointlyOCC alone
Written forBanking organizations with more than $30 billion in total assets (stated in SR 26-2)OCC examiners at national banks and federal savings associations
ScopeStatistical and non-generative models; complex quantitative methodsAll bank AI; more applicable to generative and agentic AI
Generative and agentic AIOut of scope; stated in OCC Bulletin 2026-13In scope; named throughout
What drives scrutinyModel risk, tailored to the bank’s model risk profileConsequence of failure, agency, human oversight, reversibility
InventoryModel inventory, narrower model definitionComplete AI use inventory; approval points for deployment, change and retirement
OwnershipModel owners and governanceNamed executives and use-case owners; the bank stays accountable with a vendor
TestingValidation and monitoring; depth set by materialityAdversarial testing, red teaming, hallucination measurement; decision reconstruction
Human oversightNot addressed for agentsHuman-in-the-loop against human-on-the-loop; reviewers need authority and time
ShutdownNot addressedTransaction limits, scoped permissions, kill switch, tested fallback, remediation
Binding statusInteragency supervisory guidance; risk-basedExaminer resource; sets no supervisory expectations for banks

SR 26-2 tells banks how to manage models. OCC Bulletin 2026-54 tells examiners what to look for in AI outside that definition. Neither document is a rule, and neither supersedes the other. Both shape the document request list.

State-chartered banks sit outside the OCC framework, which covers OCC-supervised institutions. State examiners have their own option: the CSBS Artificial Intelligence Supervisory Framework, released on September 16, 2026. It gives them a core guide, a document request list, a work program and an optional AI use case risk tiering worksheet. Adoption is discretionary, and each state decides.

What an examiner will ask under each

If you put the two documents next to each other and imagine the conversation in the exam room, the questions sort into two groups.

Under SR 26-2 (models)

  • Is this a model under the narrower definition? Evidence: a re-baselined inventory with scoping rationale.
  • How material is it, and does validation depth match? Evidence: tiering and validation reports.
  • Who owns it, and who challenged it? Evidence: the owner, the validator and the challenge record.

Under OCC 2026-54 (AI, including agents)

  • What AI is in use, including vendor features and agents? Evidence: one inventory covering both, the subject of Your AI Inventory Is a List of What People Remembered to Tell You.
  • What may each tool do, reach and decide, and when must a person step in? Evidence: a testable purpose statement.
  • Who is accountable by name? Evidence: an executive and a use-case owner per tool.
  • How would the bank stop it, and what happens to harmed customers? Evidence: a tested shutdown, fallback and remediation workstream, set out in An AI Kill Switch Isn’t a Button.
  • Can the bank reconstruct a material decision? Evidence: an audit trail of inputs, tool calls, approvals and overrides.

A mature SR 26-2 program handles the first list well. Three questions on the second list, inventory, accountability and shutdown, are ones that program alone cannot answer. That is where the next month should go.

Three things to have ready

A single inventory

Models under SR 26-2 and AI uses under OCC Bulletin 2026-54 belong in one register, with the tier, purpose, data reached and approval history for each entry. When agents live in a separate spreadsheet, the exam finds that gap before you do. The AI inventory post covers the fields examiners ask for.

Named owners

Every Tier 1 and Tier 2 use needs an executive and a use-case owner, including vendor-hosted tools. The framework is clear that the bank stays accountable even when a vendor runs the system. Write down who may approve, pause and retire each one, too.

A tested shutdown path

For each consequential use, you want limits and scoped permissions, a kill switch someone has actually pulled in a drill, a fallback process, a way to find affected customers, and dated drill results to show for it. The kill switch post walks through the drill end to end.

A 30-day plan. Week 1: re-baseline the inventory. Week 2: tier every use by consequence. Weeks 3 and 4: run one shutdown drill on the highest-tier agent. The drill will not be perfect, and that is the point. Find the gaps before an examiner does.

Why this is easier when the AI runs in one place

Inventory, ownership and shutdown all get hard when agents are spread across vendor consoles, features bundled into software you already bought, and people’s personal logins. There is no single list to hand over, and no single switch to pull.

Shakudo gives teams a single control plane to govern every AI model, agent and workflow, deployed entirely in their own infrastructure. Policies apply per agent, per team or platform-wide, and tamper-proof logs capture every prompt, tool call and model response, as set out on the AI governance platform page. Inventory, owner and shutdown evidence come from one place.

A platform will not make your tiering decisions or name your owners. That work stays with the bank. What it does is make the evidence much easier to produce when someone asks for it.

Start with the Tier 1 list

The systems an examiner probes first carry the most consequence. If an examiner asked for your Tier 1 AI list tomorrow, would your agents be on it?

Start with that list. Tier every use by consequence, name an owner for each Tier 1 and Tier 2 use, and run one shutdown drill this month. For help building the inventory and closing the gaps, book a free AI inventory and gap review.

Frequently asked questions

What is SR 26-2?

Revised interagency model risk guidance from the Federal Reserve, the OCC and the FDIC, issued on April 17, 2026 and published by the Federal Reserve. It replaced SR 11-7 and SR 21-8, and it is most relevant to banking organizations with more than $30 billion in total assets.

Does SR 26-2 cover generative AI or AI agents?

No. OCC Bulletin 2026-13 calls generative AI and agentic AI models novel and rapidly evolving, and places them outside its scope. Banks still govern those systems under broader risk management, and coverage of OCC Bulletin 2026-54 explains how examiners apply the framework to them.

What is OCC Bulletin 2026-54?

The OCC’s supervisory framework on bank use of AI, issued on November 4, 2026. The OCC calls the technology “super intelligence”. It sets no supervisory expectations for banks and is not intended for bank self-assessment.

How is SR 26-2 different from SR 11-7?

SR 26-2 is shorter and explicitly risk-based. It narrows the definition of a model, focuses on larger banks, and places generative and agentic AI outside its scope. The SR 11-7 structure of development, validation and governance survives. Fiddler AI’s explainer runs through what carries over and what changes.

How does OCC 2026-54 tier AI risk?

By consequence. The framework weighs whether a tool acts or advises, whether a mistake can be undone, potential harm, autonomy and permission scope. Its illustrative scale runs from Tier 1, Critical, to Tier 4, Low.

Does OCC 2026-54 apply to state-chartered banks?

No. It is written for OCC examiners. State examiners may use the CSBS Artificial Intelligence Supervisory Framework, released on September 16, 2026, which offers an optional use case risk tiering worksheet. Each state decides whether to adopt it.

What should a bank have ready for an AI exam?

One inventory covering models and AI uses, a named owner for every consequential use, and a tested shutdown path with dated drill results and a customer remediation process.

Use 225+ Best AI Tools in One Place.
Get Started
trusted by leaders
QuadReal
Loblaw Digital
CentralReach
Huntington Bank
Whitecap Resources
Gallo
Shakudo powers AI infrastructure for the these companies
QuadReal
Loblaw Digital
CentralReach
Huntington Bank
Whitecap Resources
Gallo
CloudHQ
Flexivan
BWX Technologies
Ready for Enterprise AI?
Neal Gilmore
Request a Demo