The OCC's examiner framework (Bulletin 2026-54) describes three levels of human involvement in AI: a person approves before anything happens (human-in-the-loop), a person monitors and can intervene (human-on-the-loop), or the system acts without a real-time human checkpoint (fully autonomous). It does not say every AI decision needs a human. It says involvement may be calibrated to consequences, autonomy, reversibility and volume. Oversight is described as effective when the bank defines the reviewer's role and can show it works, with real authority, competency, capacity and a meaningful review standard.
TRACKED DOCUMENTS: OCC Bulletin 2026-54
Where it is
Section 2.3 of the OCC framework, plus the autonomy factor in section 3.1 and the lexicon
Three modes
Human-in-the-loop, human-on-the-loop, fully autonomous
Status
A resource for examiners. Bulletin 2026-54 says it does not set expectations for banks
What is an AI agent in the OCC's framework?
An AI agent is a system that does things, not just one that answers questions. The OCC's lexicon (Appendix A) defines an “SI Agent” as “a system that autonomously perceives its environment, decides what to do, and takes actions to achieve its goals.”
A note on words first. The OCC writes “Super Intelligence (SI)” where everyone else writes AI, so this page says AI except inside quotations. The framework came out with OCC Bulletin 2026-54, and it is written for examiners. The bulletin says it “does not establish supervisory expectations for banks”.
The lexicon also defines “Agentic SI” as “a category of SI systems capable of independently making decisions, interacting with their environment, and optimizing processes without direct human intervention.” Put simply, a chatbot that drafts a reply is generative AI. A system that drafts the reply, sends it and updates the customer record on its own is an agent.
The difference matters for oversight. A person can catch a wrong draft before it goes anywhere. An agent's wrong action may already have happened by the time anyone looks. That is why the framework ties the amount of human involvement to what the system is allowed to do.
What are human-in-the-loop, human-on-the-loop and fully autonomous AI?
They are three levels of human involvement. A person approves first (in the loop), a person watches and can step in (on the loop), or the system acts with no real-time human checkpoint (fully autonomous).
Section 2.3 says human-in-the-loop “generally refers to review and approval before an output or action is used or executed”. It says human-on-the-loop “generally refers to active monitoring with authority and practical ability to intervene.” The lexicon adds that on the loop, the system “is operating independently and a qualified human is overseeing, overriding and/or halting actions based on predefined parameters.” Section 3.1 names all three as rungs of the “Autonomy level” factor: “human-in-the-loop (approves each action), human-on-the-loop (monitors and can intervene), or fully autonomous (acts without real-time human checkpoint).”
| Mode | What the OCC text says | Plain words | Everyday bank example (ours, not the OCC's) |
|---|---|---|---|
| Human-in-the-loop | “Review and approval before an output or action is used or executed” (2.3). “Approves each action” (3.1). | Nothing happens until a person says yes. | For example, an AI drafts a credit memo summary and an analyst reads and corrects it before it goes to the credit committee. |
| Human-on-the-loop | “Active monitoring with authority and practical ability to intervene” (2.3). “Monitors and can intervene” (3.1). | The system acts. A person watches and can stop or reverse it. | For example, an AI sends routine payment-reminder messages, and a team lead watches a dashboard and can pause the campaign. |
| Fully autonomous | “Acts without real-time human checkpoint” (3.1). | The system acts alone. People look later, if at all. | For example, an AI that releases fraud holds on card payments in real time. The OCC lists “fraud-hold release” as a Tier 1 use case in section 3.2. |
The OCC's own tier table in section 3.2 links the modes to risk. Tier 3 (Moderate) describes “decision support with mandatory human-in-the-loop”. Tier 2 (High) lists “human-on-the-loop”. Tier 1 (Critical) is “autonomous or partially autonomous”. The risk tiers page explains each tier.
Does the OCC say a human must approve every AI decision?
No. The framework says the level of human involvement “may be calibrated” to the use case, and it treats fully autonomous operation as one of the levels a bank's AI may sit at.
Section 2.1 says a purpose statement “may distinguish actions the system can take autonomously from actions that call for review, approval, abstention, or escalation.” That sentence assumes some actions will be autonomous. Section 1.1 says higher-consequence uses “may warrant greater restrictions on autonomy” and “more direct human oversight”, which implies lower-consequence uses may need less. Section 3 describes the goal as “light-touch for low-risk, low-autonomy use, materially heavier for anything approaching customer-impacting, irreversible, or highly autonomous action.”
The framework also does not say that autonomy is forbidden at the top tier. It says an autonomous, customer-facing payment agent “sits at Tier 1, regardless of how well it performs in testing” (section 3.1). In practice, autonomy moves a use case up the tiers, and higher tiers draw heavier controls everywhere else in the framework.
How much human involvement does an AI use case need?
It depends on the use case. Section 2.3 says the level of involvement “may be calibrated to the potential consequences, autonomy, reversibility, and operating volume of the use case, among other factors.”
- Potential consequences. How much harm a mistake could do. Section 1.1: “control intensity could increase with the potential consequences of an SI agent's failure or misuse.”
- Autonomy. How much the system does on its own. Section 2.2 points to “the length of time a system is permitted to operate without a meaningful human checkpoint” and the breadth of its data, tool and account access.
- Reversibility. Whether a mistake can be undone. Section 3.1 gives “a payment sent, a disclosure made, an account closed” as examples of actions that may be effectively irreversible.
- Operating volume. How many outputs or actions there are. Section 2.3 warns that “high volumes or bulk approval could reduce review quality and contribute to alert fatigue.” Alert fatigue means people stop paying attention because they see too many alerts.
The phrase “among other factors” leaves the list open. Section 2.3 also says involvement can go up after something goes wrong: “A misfire, hallucination, or control breach may justify temporarily increasing human involvement.” A hallucination is a confident but wrong output. For example, if a customer-service assistant quotes the wrong overdraft fee, a bank might switch it back to having staff approve answers on fees for a while.
Section 1.1 adds a warning about what not to measure. 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.” Its example: “a simple rules-based agent authorized to initiate transactions could present greater risk than a sophisticated model limited to internal drafting assistance.”
What does "transfer of agency" mean in the OCC framework?
The framework uses the phrase but does not define it, and the lexicon in Appendix A has no entry for it. Put simply, it means how much of the power to act a bank hands from people to the AI system.
The phrase appears twice. Section 1.1 says risk classification “may be based on potential consequences, transfer of agency, and the degree of human oversight”. Section 2.2 says classification “may be based on the consequences of failure and transfer of agency.” Section 2.2 does not say what makes agency larger, but its examples point the same way (our reading): autonomy, the time a system runs “without a meaningful human checkpoint”, and “the breadth of its data, tool, and account access”. Its example: “A highly autonomous payment agent may present greater SI risk than an SI agent only used for internal drafting.”
In practice, think of a scale. An AI that suggests text to an employee transfers little agency, because the employee still decides and acts. An AI that can move money from a customer's account transfers a lot. Human oversight is one of the ways to take some of that agency back.
What makes human oversight of AI effective, and what makes it a rubber stamp?
Oversight works when the bank has defined what the person does and can show it works. Section 2.3 opens: “Human oversight is most effective when the bank defines the nature of the involvement and can demonstrate that the control operates as intended.”
The framework's summary table names the failure. The specific risk for section 2.3 is “a named human reviewer without practical authority, competency, capacity, or a meaningful review standard”. Put simply, a name on an approval screen is not oversight by itself.
| Section 2.3 consideration | What the text says | Rubber-stamp sign (our illustration) |
|---|---|---|
| Meaningful review | Triggers, evidence and procedures “may be defined so that approval is more than a nominal click.” | Approvers click “accept” on hundreds of items an hour. |
| Span of control | A bank “may evaluate how many agents, accounts, processes, alerts, or actions one person can credibly monitor and the time available to intervene.” | One person is the named monitor for a dozen agents running overnight. |
| Override, logging, escalation | “Override authority, reasons, patterns, and escalation conditions may be logged and reviewed.” | Nobody ever overrides, and nobody asks why. |
| Independence | “Where appropriate, the reviewer may be separate from the team that developed or owns the system.” | The team that built the tool is the only one checking its output. |
An override is when a person ignores, changes, rejects or reverses what the system produced (the lexicon's definition, in plain words). The framework also lists “human-oversight” testing among the pre-deployment tests in section 2.13. In practice, that means checking that the human control works before launch, not only that the AI does.
What does a person need to really review an AI system's work?
Section 2.3 names expertise, information and authority. It also points to time, through the span-of-control consideration.
- Expertise and information. The reviewer “may need sufficient expertise, information, and authority to evaluate, reject, modify, pause, or halt the system's output or action.”
- Authority. The same sentence lists five things the reviewer may be able to do: evaluate, reject, modify, pause or halt. On the loop, the lexicon describes a “qualified human” who can oversee, override and halt.
- Time. Section 2.3 refers to “the time available to intervene” and to how much one person can “credibly monitor”.
- Evidence. Section 2.3 mentions defined “evidence” for review. Section 2.14 adds that records of inputs, outputs, tool calls, approvals and overrides may help reconstruct a decision.
The rest is our reading, not the OCC's text. For example, a reviewer approving AI-drafted adverse action letters would want to see the application data the AI used, not only the finished letter. A reviewer who may halt an agent needs a stop button that works and a clear rule that using it will not be held against them. Kill switches, meaning ways to shut a system off quickly, appear in section 2.10. See the resilience and kill switches page.
Where does human oversight fit in the rest of the OCC framework?
Human oversight is one of sixteen components, and section 1.2 places it in two of the six pillars: Control and Oversight.
It leans on others. Section 2.1 asks what the purpose statement says about actions that call for review. Section 2.4 covers who is accountable, explained on the board, management and audit page. Section 2.15 says audit may check whether a human process that serves as a backstop is “independently governed, tested, and effective”. For the full text by section, read the Bulletin 2026-54 reference guide. For all OCC AI documents, see the OCC AI regulation page and the OCC AI glossary.
What is the difference between human-in-the-loop and human-on-the-loop?
Human-in-the-loop means a person reviews and approves an output or action before it is used. Human-on-the-loop means the system acts on its own while a person actively monitors it and has the authority and practical ability to step in. The definitions are in section 2.3 of the OCC framework.
Can a bank let an AI agent act with no human approval?
The framework does not rule it out. It lists fully autonomous operation as one autonomy level and says involvement may be calibrated to the use case. Autonomy pushes a use case toward Tier 1, where the framework describes the heaviest controls.
Is a person clicking approve enough to count as oversight?
Not on its own. Section 2.3 says review procedures may be defined so that approval is more than a nominal click. The risk it names is a reviewer without practical authority, competency, capacity or a meaningful review standard.
How many AI agents can one person monitor?
The OCC gives no number. Section 2.3 says a bank may evaluate how many agents, accounts, processes, alerts or actions one person can credibly monitor, and the time available to intervene.
Should the reviewer be independent of the team that built the AI?
Section 2.3 says that, where appropriate, the reviewer may be separate from the team that developed or owns the system. It does not say this is needed in every case.
What happens to human oversight after an AI mistake?
Section 2.3 says a misfire, hallucination or control breach may justify temporarily increasing human involvement. For example, a task that ran on the loop could go back to person-approves-first for a while.
Do these OCC oversight points bind my bank?
No. Bulletin 2026-54 says the framework does not establish supervisory expectations for banks and is a resource for examiners. It describes what examiners may consider when they look at a bank's AI.
- Super Intelligence: Supervisory Framework (full text, PDF) — Office of the Comptroller of the Currency, Nov 4, 2026 (occ.gov) · our page on OCC Bulletin 2026-54
- 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
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