Feature AI Risk
AI Risk Tiering Just Flipped: OCC 2026-54 Ranks Consequence, Not Sophistication
OCC Bulletin 2026-54 puts a rules-based payment agent in Tier 1 and your frontier model in Tier 4. Build AI risk tiering on agency and reversibility.
Table of Contents
TL;DR
- OCC Bulletin 2026-54 (Nov. 4, 2026) tells examiners to rank AI use cases by consequence, transfer of agency, and human oversight — “rather than the technical sophistication of the SI used.” A rules-based payment agent sits in Tier 1. A frontier model that only drafts internal text sits in Tier 4.
- The framework is an examiner tool, not a rule. That does not make it optional reading. It is the document that will shape scoping, requests, and management conversations in your next OCC exam.
- Your tiering worksheet needs to score agency and reversibility first. Sophistication is the tiebreaker, not the driver.
- The two artifacts examiners will actually ask for are a testable purpose statement per use case and evidence that a human-on-the-loop reviewer has capacity to intervene — not a named reviewer on a slide.
A smarter model is not the riskier one. The OCC said so in writing on November 4, 2026, and it inverted the way most AI programs score their own inventory.
The framework’s own example is blunt: a simple rules-based agent authorized to initiate transactions “may present a greater risk than a more sophisticated model limited to internal drafting.” Examiners are told to classify by potential consequences, the transfer of agency from a person to a system, and the degree of human oversight. Technical sophistication is expressly the thing you classify away from.
Most AI risk registers built in the last two years are sorted the other way. The frontier model with the big validation file goes to the top of the board report; the Zapier-shaped payment bot that moves money sits in the same tier as a chatbot because it’s “just rules.” Nothing about it is a model, so it never got a model risk rating, a validator, or a control test.
What Bulletin 2026-54 Is, and Why It Still Matters
The document is addressed to examiners. The OCC repeats the disclaimer in the summary and again in the text: the framework “does not establish supervisory expectations for banks,” creates no new regulatory expectations, no enforceable standards, no prescriptive requirements, and banks “are not expected” to use it as a self-assessment tool.
Every vendor webinar in the next month will use those sentences to say nothing changed. Two things did.
First, the bulletin instructs examiners to tailor examination depth “to the materiality of the SI use case.” Scope is now a function of your tiering — which means a tiering decision you cannot defend becomes an examination finding about scoping, even before anyone argues about the control.
Second, it resolves a real coverage hole. The April 17, 2026 interagency model risk guidance, OCC Bulletin 2026-13, narrowed the word “model” to complex quantitative methods and pushed generative and agentic AI out of scope as “novel and rapidly evolving.” So the one class of AI that can act on its own had no home in the model guidance. The framework fills that gap on the examiner side, with a footnote confirming the division of labor: Bulletin 2026-13 for non-generative models, this document for generative and agentic use.
A control framework with no territory is how a payment agent ends up owned by a product manager and reviewed by nobody.
The Tiering Is the Deliverable
The framework carries an illustrative four-tier method and seven factors. Reproduced faithfully, the tiers look like this:
| Tier | What puts a use case here | Framework’s own examples |
|---|---|---|
| 1 — Critical | Direct action, little or no reversibility, high financial or customer impact, autonomous or partly autonomous | Autonomous payment agent; real-time credit decisioning; fraud-hold release; market access |
| 2 — High | Direct action or heavy decision support, human-on-the-loop, moderate-to-high impact | AI-assisted underwriting; agent-initiated customer communications; AML alert disposition |
| 3 — Moderate | Decision support with mandatory human-in-the-loop, limited data sensitivity, reversible | Document summarization for credit memos; agent-assisted research |
| 4 — Low | No customer or financial data exposure, no direct action, fully human-controlled | Internal productivity tools; code assistance; general employee Q&A |
Note what is missing from the top row: model class, parameter count, vendor prestige, validation rigor. A Tier 1 assignment holds “regardless of how well it performs in testing.”
The seven factors map cleanly onto a scoring sheet: decision consequence, reversibility, customer or financial impact, autonomy, data sensitivity, explainability, provenance. Score each 1–4, weight the first four, and let the last three adjust. A defensible starter weighting is 30% decision consequence, 25% reversibility, 25% autonomy, 20% customer/financial impact, with data sensitivity, explainability, and provenance as modifiers — then calibrate against three to six months of your own incident and override history before you publish it.
Reverse-scoring reversibility is where programs get it wrong. “Can we reverse it?” is not “can we fix it later?” A wrong internal summary gets fixed by a person reading it. A wrong outbound payment needs an exception item, a funds recall, and, if the customer noticed first, a complaint record. Score reversibility on how the error surfaces and how many hours it takes, not on whether somebody eventually resolves it.
A Purpose Statement Examiners Can Test
The first framework component is defined purpose and permitted use, and it opens with the sentence that matters: “Effective SI risk management begins with a clearly defined and documented purpose.”
Purpose has to cover five things — what the system is designed to do, what it is not designed to do, who may invoke it, which data and systems it may reach, and the conditions that require human review or suspension. The failure mode is named too: an activity outside the approved purpose “may be treated as a control failure even if the outcome looks reasonable.”
Vague purpose language is the reason that happens. “Assists the servicing team with customer inquiries” covers a retrieval bot, an email drafter, and an agent that issues fee credits. Write the boundary instead:
Example control language: Scope — drafts responses to inbound servicing emails on the following 11 intent classes; may attach a payoff statement from the loan system of record; may not issue fee waivers or credits; may not commit to a modification term. Any response quoting a payoff figure over $250,000 requires servicing-lead approval before send. Invocation limited to the Servicing Queue role. Data reach: loan system read-only, no document repository write. Suspension triggers: any send failure at the SMTP gateway, any payoff mismatch flagged in reconciliation, or a compliance exception on a sampled response.
That paragraph is testable. An examiner can pull sent mail, find a fee waiver, and compare it to the scope line. It also gives the product team a review gate when they want to widen the agent, which is the practical point: scope drift is usually a feature launch, not a rogue agent.
Human-on-the-Loop Is a Capacity Question
The distinction between human-in-the-loop and human-on-the-loop is not semantic. HITL means review or approval before execution. HOTL means active monitoring with the authority and practical ability to intervene.
The word doing the work there is practical. The framework calls out a named reviewer “without practical authority, competency, capacity, or a meaningful review standard,” and names bulk approval and alert fatigue as the ways the control degrades into a signature.
So the evidence artifact is not a reviewer name. It is a span-of-control study. Take your Tier 1 and Tier 2 use cases, count the actions or alerts each generates per reviewer shift, and record the number at which the reviewer’s sampled review quality actually falls. For a payments exception queue, a defensible starting point is one reviewer per 25–40 live agent decisions per hour with a 10% sampled deep review; for an alert disposition agent, whatever your AML alert volume has already proven the team can absorb. Then post the number, and make the queue throttle when it’s exceeded.
Three questions with real answers beat a policy paragraph:
- Who holds the pause authority, and is it a named individual on shift rather than a committee?
- What is the interface action that halts the agent, and how long does it take from recognition to halt?
- When a reviewer overrides, where is the reason captured, and who reads the override pattern monthly?
The Vendor Clauses That Get Pulled
Third-party and vendor risk is a named assurance component. The vendor points the framework raises are narrow and specific: contractual limits on what a provider’s agents may do inside bank systems, agreed definitions and notice periods for AI incidents, clear allocation of liability for errors and customer harm, and advance notice when a vendor adds AI features to a product the bank already uses.
That last one is the quiet killer. Vendors ship model updates as dot releases. Without a notice clause, your Tier 2 underwriting assistant can become a different system between the approval and the first monitoring read, and nothing in your change log reflects it. Pair the clause with an internal model change assessment flow so a vendor release notification lands in the same routing decision as an internal change, and treat a missed notice as a vendor KRI breach rather than a surprise.
Accountability does not move. A provider’s agent that acts inside your core still acts for you.
Three Places This Breaks in Practice
Ownership falls through the model risk gap. Bulletin 2026-13 excluded agentic AI, so the model risk function can correctly decline to own it. In banks that have not restructured, the case lands on whoever built it. The workable split: model risk owns the Tier 1 statistical and quantitative use cases, the operational risk function owns the agent controls and override logs, and a single accountable executive signs the tiering register — because the framework says examiners should not treat AI as a separate risk category, and neither should your org chart.
Tier 4 becomes a junk drawer. “Internal productivity” collects every use that nobody wants to score, including the coding assistant that touches a production repository. The framework puts code assistance in Tier 2 when it can change internal code or data access. Calibrate the bottom tier against reach, not against audience.
Compliance gets copied on the design late. For consumer-facing agent actions, the exposure is ordinary Reg E, fair lending, disclosure, servicing, and UDAAP work — the framework says using AI “does not, by itself, alter the applicability of those requirements.” If compliance first sees the agent at launch, the remediation capability (can you identify affected customers, reconstruct the error, and correct it?) is already fixed by whatever logs the build team happened to include.
So What?
Bulletin 2026-54 gives you a defensible way to spend a finite control budget for the first time: score the actions, not the architecture. Do the re-tiering this month, before exam scoping starts, and do it in a file that shows the seven factors with the reasoning next to each score. Then fix the two artifacts with the shortest shelf life — the purpose statement on your top five use cases and the reviewer capacity numbers behind them.
If the tiering exercise produces a Tier 1 list with no able owner, that is the finding. Better to write it down yourself than to read it in a supervisory letter.
Starter kit for this: the AI Risk Assessment Template & Guide ships an editable use case inventory with the consequence-based tiering factors, agent-specific scope questions, and a vendor AI clause checklist — $59.
FAQ
Does the OCC framework require a separate AI risk appetite?
No. The bulletin says examiners should not treat super intelligence as a separate and distinct risk category and that banks are not required to define a separate SI risk appetite. Practically, that means AI controls attach to operational, compliance, credit, and third-party risk categories rather than living in a standalone AI policy.
Where should a model that only summarizes documents land?
Tier 3 in the framework’s illustration, provided human-in-the-loop review is mandatory and the data is limited. Bump it if the summaries feed a credit decision nobody re-reads, because the failure then travels into the decision rather than stopping at the draft.
What if our inventory only tracks models?
Then it is missing the highest-risk use cases. The framework’s Tier 1 examples include fraud-hold release and autonomous payments, and neither is a model in the Bulletin 2026-13 sense. Build the inventory around agents and action authority, and map model records into it rather than the reverse. The examiner question set is a reasonable stress test for whether your register would survive a request, and the older agentic AI governance gap piece covers why the model framework never fit agents to begin with.
◆ Need the working template?
Start with the source guide.
These answer-first guides summarize the required fields, evidence, and implementation steps behind the templates practitioners search for.
◆ Related template
AI Risk Assessment Template & Guide
Comprehensive AI model governance and risk assessment templates for financial services teams.
◆ Immaterial Findings · Weekly
Sharp risk & compliance insights. No fluff.
◆ FAQ
Frequently asked questions.
What is OCC Bulletin 2026-54?
Why does a simple AI agent outrank a sophisticated model?
Does this framework apply to generative and agentic AI under the model risk guidance?
Does a bank need a separate AI risk appetite now?
What does meaningful human oversight mean in practice?
Author
Rebecca Leung
Rebecca Leung has 8+ years of risk and compliance experience across first and second line roles at commercial banks, asset managers, and fintechs. Former management consultant advising financial institutions on risk strategy. Founder of RiskTemplates.
◆ Related framework
AI Risk Assessment Template & Guide
Comprehensive AI model governance and risk assessment templates for financial services teams.
◆ Keep reading
Related posts.
AI Risk
Colorado's Draft ADMT Rules Put a 45-Day Clock on Human Review
Colorado's interim ADMT rules would require acknowledging human-review requests in 10 days and finishing in 45. Here's the review queue to build by Jan. 1.
Nov 19, 2026
AI Risk
Your Agent Is Logged In as a Service Account Nobody Owns
SR 26-2 pushed agentic AI out of model risk scope. AI agent access control now needs an entitlement register, JIT credentials, and a named owner.
Nov 18, 2026
AI Risk
Retiring an AI Model Is a Control, Not a Cleanup Task
AI model retirement is a control, not a cleanup task. The decommission memo, retention surface table, and inventory states that survive an exam.
Nov 15, 2026