How does the OCC sort AI use cases into risk tiers?

Published Nov 21, 2026 · Last updated Nov 21, 2026

The OCC's 2026 examiner framework sorts AI use cases by what could go wrong, not by how advanced the technology is. Seven factors (decision consequence, reversibility, customer and financial impact, autonomy, data sensitivity, explainability and provenance) combine into four illustrative tiers, from Tier 4 Low (internal productivity tools, employee Q&A) to Tier 1 Critical (autonomous payment agents, fraud-hold release). Controls could get heavier as the tier rises, but the tiering is an example for examiners, and a bank may use its own.

TRACKED DOCUMENTS: OCC Bulletin 2026-54

Why does the OCC sort AI use cases by consequence and not by how advanced the AI is?

Because harm comes from what a system is allowed to do, not from how clever it is. The framework says risk classification “may be based on potential consequences, transfer of agency, and the degree of human oversight, rather than the technical sophistication of the SI used” (section 1.1).

A note on words first. The OCC uses the term “Super Intelligence (SI)” for what everyone else calls artificial intelligence. This page says “AI” and keeps “SI” only inside quotations.

The framework gives its own example in section 1.1: “a simple rules-based agent authorized to initiate transactions could present greater risk than a sophisticated model limited to internal drafting assistance.” Section 2.2 says the same thing another way: “A highly autonomous payment agent may present greater SI risk than an SI agent only used for internal drafting.”

The three ideas in the bulletin's highlights mean, in plain words:

  • Potential consequences: how bad things get if the AI is wrong or misused.
  • Transfer of agency: how much of the doing is handed from a person to the AI. A tool that suggests keeps the agency with the person. A tool that acts takes it.
  • Degree of human oversight: whether a person checks each output, watches and can step in, or is not involved in real time.

The overarching principle in section 1.1 ties this to controls. A bank's “control intensity could increase with the potential consequences of an SI agent's failure or misuse.” Put simply, the more damage a use case could do, the more checks around it. The framework is a resource for examiners. Bulletin 2026-54 says it “does not establish supervisory expectations for banks”. For the whole framework see the section-by-section reference.

What factors does the OCC use to place an AI use case in a tier?

Section 3.1 lists seven factors, and says every AI use case “may be scored” against them. The point is that “a comparable risk gets a comparable control regardless of where it originates in the bank.”

Factor (section 3.1)The question in plain wordsLower riskHigher risk
Decision consequenceDoes the AI inform a person, decide, or act on its own? The OCC says “Direct action carries materially higher risk than decision support.”Suggests to a personMoves funds, updates a customer record, trades
ReversibilityCan a mistake be caught and undone before harm? The OCC's examples of the irreversible: “a payment sent, a disclosure made, an account closed”.A draft nobody has sentA payment already out the door
Customer/financial impactHow big is the harm to a customer, the markets or the bank if it is wrong?One transaction, low dollar amountA portfolio-level decision, high dollar amount
Autonomy levelHow involved is a person? Human-in-the-loop means a person approves each action. Human-on-the-loop means a person monitors and can intervene. Fully autonomous means no real-time human checkpoint.Human approves each actionNo human checkpoint
Data sensitivityDoes it touch regulated, confidential or personally identifiable data (data that names or points to a person)? Could it create a privacy or consumer-law problem?Public or internal non-sensitive dataCustomer account and personal data
ExplainabilityCan the output be traced to a reason good enough for legal or disclosure duties? The OCC's example is adverse action notices, the letters that tell an applicant why credit was refused.No explanation owed to anyoneA reason must be given to a customer
ProvenanceWhere does the AI come from: built in-house, fine-tuned from a third-party base model, or used as an outside vendor service? “Each origin carries a different visibility and control profile.”The bank can see inside itAn opaque outside service

The “lower” and “higher” columns are our plain-word illustration, not OCC text. Section 2.2 lists a few more factors to weigh: permission scope (how much data, tools and accounts the AI can reach), interdependencies, and aggregate exposure. On aggregate exposure it warns that “shared providers, data, permissions, infrastructure, and control weaknesses could create concentration or correlated risk across otherwise separate use cases.” For example, five small tools on one vendor with one set of credentials can fail together.

What are the four OCC AI risk tiers?

Section 3.2 defines four illustrative tiers, from Tier 1 Critical to Tier 4 Low. The seven factors “combine into a ‘composite tier’” (section 3.1). The framework gives no weights or scoring formula.

TierProfile (section 3.2)The OCC's own examples
Tier 1 · CriticalDirect action, low/no reversibility, high financial or customer impact, autonomous or partially autonomousAutonomous payment agent; real-time credit decisioning; fraud-hold release; market access
Tier 2 · HighDirect action or high-weight decision support, human-on-the-loop, moderate-to-high impactAI-assisted underwriting; agent-initiated customer communications; AML alert disposition; code assistance for internal code or data access changes
Tier 3 · ModerateDecision support with mandatory human-in-the-loop, limited data sensitivity, reversibleInternal document summarization for credit memos; agent-assisted research
Tier 4 · LowNo customer/financial data exposure, no direct action, fully human-controlled useInternal productivity tools; code assistance; general employee Q&A

The framework does not attach a fixed list of controls to each tier. What it gives is a direction. Section 3 calls the goal “a graduated control model: light-touch for low-risk, low-autonomy use, materially heavier for anything approaching customer-impacting, irreversible, or highly autonomous action.” Section 1.1 names what grows with consequence: “greater restrictions on autonomy, stronger controls, more direct human oversight, more independent validation, and more robust resilience measures.” The sections below apply that direction tier by tier. Which framework points go with which tier is our reading, not the OCC's.

What is a Tier 1 Critical AI use case?

Tier 1 is AI that acts directly, where the action is hard or impossible to undo and the harm could be large. It runs autonomously or partly autonomously (section 3.2).

The OCC's examples are an autonomous payment agent, real-time credit decisioning, fraud-hold release and market access. Section 3.1 says a fully autonomous, customer-facing payment agent sits at Tier 1 “regardless of how well it performs in testing.” Put simply, good test results do not lower the tier.

This is the heaviest end of the gradient. In our reading, these framework points bear most on uses like these:

  • Section 2.10: for more consequential or autonomous uses, a bank may document and test detection, containment, suspension, recovery and customer remediation. Tools named include circuit breakers, rate limits and kill switches (a way to stop the AI at once).
  • Section 2.13: consequential or highly autonomous uses “may warrant broader testing” of both performance and controls.
  • Section 2.15: these uses “may warrant more frequent or substantive” internal audit coverage.
  • Section 2.3: the level of human involvement may be calibrated to consequences, autonomy, reversibility and volume.

More on stopping an agent is in resilience and kill switches.

What is a Tier 2 High AI use case?

Tier 2 is AI that acts directly or carries heavy weight in a decision, with a person monitoring and able to intervene (human-on-the-loop). The impact is moderate to high (section 3.2).

The OCC's examples are AI-assisted underwriting, agent-initiated customer communications, AML alert disposition and code assistance for internal code or data access changes. AML means anti-money laundering. Alert disposition is deciding whether a suspicious-activity alert is closed or escalated.

In our reading, the person on the loop is the key control here. Section 2.3 says the reviewer may need “sufficient expertise, information, and authority to evaluate, reject, modify, pause, or halt” the AI. It also warns that “high volumes or bulk approval could reduce review quality.” Section 2.8 flags hallucinations, meaning confident but false output, “concerning fees, rates, account status, or policies”, which matters for AI that writes to customers. See human oversight for what makes oversight real.

What is a Tier 3 Moderate AI use case?

Tier 3 is AI that supports a decision, where a person must check the output before it is used (mandatory human-in-the-loop). Data sensitivity is limited and mistakes can be reversed (section 3.2).

The OCC's examples are internal document summarization for credit memos and agent-assisted research. In practice, a credit officer reads the AI summary, checks it against the file and owns the memo. In our reading, the control weight sits on that review. Section 2.3 says review procedures may be defined “so that approval is more than a nominal click.” Section 2.13 adds a testing method for outputs with no single right answer: expert review of samples.

What is a Tier 4 Low AI use case?

Tier 4 is AI with no customer or financial data exposure, no direct action, and fully human-controlled use (section 3.2). Section 3 calls the control approach for low-risk, low-autonomy use “light-touch”.

The OCC's examples are internal productivity tools, code assistance and general employee Q&A. Section 2.13 says “lower materiality uses may warrant limited testing.” Light-touch is not no-touch. Section 2.1 still describes a written purpose and an inventory entry, and treats use outside the approved purpose as a control failure, called scope drift. For example, an employee Q&A tool that staff start using to answer customer questions has left Tier 4 territory.

How can a bank employee work out which tier an AI tool falls into?

Ask what the tool does when nobody is watching, and what happens if it is wrong. The walk-through below is our own illustration of applying the section 3.1 factors. It is not OCC text, and the OCC does not place these exact tools.

Put simply, five questions get most of the way:

  1. Does it only suggest, or can it act (send, pay, release, change a record)?
  2. If it is wrong, can someone undo it before a customer or the bank is hurt?
  3. Does a person approve each output, watch and step in, or neither?
  4. Does it see customer or confidential data?
  5. Would the bank need to explain its result to a customer or a regulator?
Tool (our example)How the factors readLikely tier
Internal Q&A assistant answering staff questions about HR policy and the expense systemOnly answers, takes no action. A wrong answer is caught by the employee or HR. No customer data. A person decides what to do.Tier 4. This matches the OCC's “general employee Q&A”.
AI drafting replies to customer emails, each read and approved by an agent before sendingSupports a decision. A person approves each draft, so a bad draft can be fixed. It does read customer data.Tier 3, by our reading. If the AI sent emails on its own, it would match “agent-initiated customer communications” at Tier 2.
Agent that releases fraud holds on card payments without a person approving each releaseActs directly. A released payment may be hard to recover. Customer and bank money at stake. Autonomous.Tier 1. “Fraud-hold release” is one of the OCC's own Tier 1 examples.

In practice, when factors point to different tiers, the higher-risk factors tend to drive the answer, since direct action and irreversibility are what the framework stresses most. That is our reading. The framework itself only says the factors combine.

Does a bank have to use the OCC's four tiers?

No. The tiering is labelled “illustrative” throughout section 3, and the bulletin says banks “are not expected to use this document to assess their use of SI or their SI risk management.”

Illustrative means an example of one way to do it. A bank may use its own tiering, with three tiers or five, its own names and its own scoring. Bulletin 2026-54 says nothing in the guide is meant “to establish new regulatory expectations or set forth enforceable standards or prescriptive requirements.” It also says examiners should tailor their work “based on the materiality of the SI use case”. In practice, a bank whose own scheme sorts by consequence, action and oversight gives an examiner something easy to compare with the OCC's picture.

Two more points from the framework. Examiners “should not treat SI as a separate and distinct risk category, and banks are not required to define a separate SI risk appetite” (section 1). And footnote 1 says the model risk guidance in OCC Bulletin 2026-13 applies to non-generative, non-agentic AI models, so the framework is “more applicable to generative and agentic SI.” See model risk management for generative AI and the OCC tracker.

What moves an AI use case up or down a tier?

Changing what the AI can do, who checks it, or what it touches. Section 2.2 says a bank may assess each use case “before deployment and when material changes occur.”

Changes that push a use case up, each tied to the framework:

  • Going from internal to customer-facing or decision-relevant. Section 2.1: a use approved for internal research “may warrant additional review” before that step.
  • Gaining the power to act: send, pay, release, update records (section 3.1, decision consequence).
  • Removing the human checkpoint or lengthening the time without one (section 2.2, autonomy and permissions).
  • Broader data, tool or account access (section 2.2, permission scope).
  • A vendor adding AI features or changing the model under a product already in use (section 2.9).
  • Sharing a provider, data or credentials with other use cases (section 2.2, aggregate exposure).

Changes that bring it down are the reverse: narrower permissions, a person approving each action, actions that can be undone, less sensitive data. Section 2.3 also allows a temporary move the other way: a misfire, hallucination or control breach “may justify temporarily increasing human involvement.” What does not lower a tier is test performance alone (section 3.1). The OCC's own placements show this. “Code assistance” is Tier 4, but code assistance for “internal code or data access changes” is Tier 2. “AI-assisted underwriting” is Tier 2, while “real-time credit decisioning” is Tier 1.

Is a customer chatbot Tier 1 or Tier 2?

The framework does not place a customer chatbot by name. Its closest examples are general employee Q&A at Tier 4 and agent-initiated customer communications at Tier 2. Where a specific chatbot lands depends on the factors: whether it can take actions, what data it sees, and whether a person checks what it says.

Can a very accurate AI tool be put in a lower tier?

Not on accuracy alone. Section 3.1 says a fully autonomous, customer-facing payment agent sits at Tier 1 regardless of how well it performs in testing. The tier follows from what the tool can do and how hard a mistake is to undo.

Does a simple rules-based system count as low risk?

Not necessarily. Section 1.1 says a simple rules-based agent authorized to initiate transactions could present greater risk than a sophisticated model limited to internal drafting. Simplicity of the technology is not one of the tiering factors.

Who decides which tier a use case is in?

The bank does, under its own risk management. The framework is a resource for examiners, who may look at how a bank classified its use cases and tailor their work to the consequences. Section 2.4 lists risk classification among the decisions for which a bank may define roles and decision rights.

Is there an official scoring formula for the tiers?

No. Section 3.1 says the factors combine into a composite tier but gives no weights or formula. The four tiers in section 3.2 are labelled illustrative.

How often should a use case's tier be checked again?

Section 2.2 says a bank may assess each use case before deployment and when material changes occur. The framework sets no calendar for it. A new permission, a new data source or a vendor model change are the kind of events that could prompt a fresh look.

Do the tiers apply to traditional credit models?

Footnote 1 of the framework says the model risk guidance in OCC Bulletin 2026-13 applies to non-generative and non-agentic AI models, so the framework is more applicable to generative and agentic AI. Traditional models are covered by that model risk guidance.

  1. 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
  2. Super Intelligence Supervisory Framework (PDF) — Office of the Comptroller of the Currency, Nov 4, 2026 (occ.gov)
  3. 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

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