Subscribe free
occsuper-intelligenceartificial-intelligenceagentic-aimodel-risksupervisionthird-party-risk

Money-Moving Agents Outrank Complex Models in OCC's New SI Exam Guide

OCC Bulletin 2026-54 tells examiners to grade banks' AI by consequence, autonomy and human oversight. What the framework says and seven practical steps.

By Lex, LexRegPulse Analyst · ·8 min read
Primary-source research · AI-drafted · human-reviewed. Methodology

By Lex | November 4, 2026

The OCC has told its examiners how to grade banks' use of what the executive branch now calls super intelligence (SI): by consequence. Under OCC Bulletin 2026-54, issued today with News Release 2026-97, a bank's risk classification of an SI use case "may be based on potential consequences, transfer of agency, and the degree of human oversight rather than the technical sophistication of the SI used." The attached Supervisory Framework puts it in operational terms: "a simple rules-based agent authorized to initiate transactions could present greater risk than a sophisticated model limited to internal drafting assistance."

The first footnote deserves more attention than the highlights. It says the model risk principles in OCC Bulletin 2026-13 apply to non-generative and non-agentic SI, so the new document "is more applicable to generative and agentic SI." The April model risk revision by the OCC, the Federal Reserve Board and the FDIC had left those systems out of scope as "novel and rapidly evolving." For more than six months the most autonomous systems in a bank had no examiner document of their own. Now they have one. It comes with a four-tier table that places autonomous payment agents in Tier 1, next to real-time credit decisioning.

What the OCC issued, and what it did not

The guide "does not establish supervisory expectations for banks." It adds: "Banks are not expected to use this document to assess their use of SI or their SI risk management." Examiners are told not to treat SI as a separate risk category, and banks need not define a separate SI risk appetite. SI instead can "introduce, amplify or modify risk in existing risk categories," to be handled through programs a bank already runs. Examiners tailor their work to each use case's materiality. For community banks, tailoring also reflects size, complexity and risk profile.

Reuters called it the OCC's first guidance on how examiners should review banks' AI use.

The terminology comes from Executive Order 14434, signed September 29 and published October 2 at 91 FR 63129. It directs executive agencies to say "Super Intelligence" in place of "Artificial Intelligence" in nonstatutory documents. Section 3(a) defines SI as the technologies covered by the statutory definition of artificial intelligence in 15 U.S.C. 9401(3). The OCC's second footnote tracks that, describing SI as the technologies "previously encompassed by the term 'Artificial Intelligence'" plus their "rapidly advancing capabilities." The legal scope has not changed, so a bank's AI policy already covers what the OCC now calls SI. The order reaches "documents within the executive branch," leaving bank policies free to keep their own vocabulary. A third footnote: the OCC says the guide "was prepared with the assistance of SI," then "reviewed, edited, and approved by a human."

How an examiner will read an SI inventory

Section 3.1 turns the three headline factors into seven scoring criteria: decision consequence, reversibility, customer or financial impact, autonomy level, data sensitivity, explainability and provenance. The guide is blunt on the first: "Direct action carries materially higher risk than decision support." The composite tiers follow:

Look at where code assistance lands. It appears in Tier 4, then again in Tier 2 when it can change internal code or data access. The tool is identical; its permissions differ. A bank's tiering will turn on access rights it can document. The guide also places a fully autonomous, customer-facing payment agent at Tier 1 "regardless of how well it performs in testing." Strong validation results will not argue a payment agent down a tier.

Human oversight is the third factor, and the guide treats it skeptically. It separates human-in-the-loop, meaning approval before an action executes, from human-on-the-loop, meaning active monitoring with authority to intervene. It wants review designed "so that approval is more than a nominal click." The summary table names the failure plainly: "A named human reviewer without practical authority, competency, capacity, or a meaningful review standard." A checkpoint lowers a use case's risk only if the bank can show the reviewer has the expertise and time to reject the output.

One more sentence matters for exams: "Activity outside the approved purpose may be treated as a control failure even when the resulting output or action appears reasonable." A good outcome does not cure scope drift.

What it means for banks and fintech partners

Accountability stays put: "accountability for managing the bank's risk generally remains with the bank," even when development or hosting is spread across vendors. For bank-fintech arrangements, two components do most of the work.

The third-party section addresses "provider agents that interact with bank systems and data." It suggests banks define each provider agent's scope of authority, identity, permitted tool use, data use and escalation limits. It flags vendors that add SI to products already in use; the bank "may require notice and transparency of new SI functionality" and may retest when changes land. On liability the verb shifts: "The bank should consider how contracts assign responsibility for model related errors or losses." In a document that speaks almost entirely in "may," that "should" stands out.

The operational resilience section covers traffic in the other direction. "Customer, counterparty, fintech, or agentic-commerce systems that the bank does not control could present channel risk." Controls such as velocity limits "may be calibrated to the action requested." It also names the agent-specific failure: "An agentic system could remain available while taking inappropriate or harmful actions."

For a sponsor bank, a fintech partner's agent that initiates payments through the bank matches the guide's own Tier 1 example. So does a partner credit model that decides applications in real time. The explainability factor cites adverse action notices by name, which pulls adverse action compliance into the tiering of any SI-driven credit program. That is a reading of the guide's examples; the guide itself never mentions sponsor banks.

Timing complicates the vendor piece. The guide says SI third-party risk management "may supplement existing interagency third-party risk management guidance." That guidance is on its way out. On September 15 the OCC, the Federal Reserve Board, the FDIC and the NCUA proposed replacement guidance (91 FR 58536). It would supersede the 2023 guidance from the Federal Reserve Board, the FDIC and the OCC. Its supplemental resources would go too, among them the July 2024 joint statement on bank-fintech deposit arrangements. SI vendor terms will sit on top of whichever third-party regime is final.

The counterargument

The strongest objection is the document's own disclaimer. The April model risk guidance says non-compliance "will not result in supervisory criticism," and the September proposal concedes the 2023 guidance "has been interpreted as establishing prescriptive requirements." Treating an examiner aid as a checklist repeats that error.

That risk is real. The guide also tells examiners to scope their work by materiality, so a bank that has already sorted its SI uses on the same factors gives the examiner a documented reason to spend less time on Tier 4 tools. Reuters noted that banks typically comply closely with such guidance even when it is not legally binding.

The awkward part is an inconsistency worth settling on paper. The bulletin says banks need not define a separate SI risk appetite, yet Section 2.4 says "The board and senior management may establish an SI risk appetite." The cleaner path is to fold autonomy and permitted-action limits into existing appetite statements, which Section 2.4 says risk appetite "could address."

Practical steps: organizing around the guide's terms

None of the following is required; the guide sets no expectations for banks. These are options for organizing existing model risk, third-party, operational risk and compliance programs around the vocabulary examiners now carry. This is directional analysis, not legal advice.

1. Build one SI inventory from the model and third-party inventories already kept. The guide suggests an inventory capture "the use case, owner, provider, permissions, deployment status, monitoring, and fail-over arrangements." Include vendor products that have added SI features. The lexicon's exclusion for common commercial products with embedded SI, "such as a word processor," keeps the list manageable.

2. Tier with the guide's seven criteria. Roll them into the three headline factors. State permitted data, systems and actions "in terms that can be tested rather than as a general capability description." Record permissions precisely, since they alone moved code assistance two tiers.

3. Scale controls to the tier. For Tier 1 and Tier 2 uses, the guide's menu includes "circuit breakers, rate limits, least-privilege permissions, and kill switches." It suggests governing agent credentials "in a manner comparable to privileged users" and keeping records that "support reconstruction of material outputs or actions." At the low end, "Lower materiality uses may warrant limited testing." That sentence is the license to keep Tier 4 overhead light.

4. Define SI incidents before one happens. The third-party section floats an incident vocabulary for "hallucination, misfire, drift or manipulation." The resilience section describes a separate customer remediation workstream to identify the full affected population. Map both into the existing operational risk event taxonomy.

5. Rewrite vendor terms at renewal. Three provisions map directly to the guide: notice of new SI functionality, limits on using bank data to train models offered to other customers, and allocation of liability for model errors.

6. Report to the board from the inventory. Section 2.4 lists material use cases, changes, exceptions, incidents, customer harm, concentrations and remediation as reporting content. Section 2.16 suggests reconciling reports to the inventory. A board pack generated from the inventory does both.

7. Assign an owner for revisions. The OCC calls the guide "a living document." Second-line risk can track each update alongside the executive order's definitional work.

What to watch

The OCC says it will update the guide regularly in response to technology changes and stakeholder feedback, without setting a comment process. Neither the interagency AI request for information promised in April nor a Federal Reserve or FDIC counterpart to the SI guide could be located in agency or Federal Register releases as of November 4. In remarks published May 1, Federal Reserve Vice Chair for Supervision Michelle Bowman said "we should assess whether our supervisory guidance is fit for the future."

The executive order also starts a definitional clock. Section 3(b) gives the Assistant to the President for Science and Technology 60 days to propose legislative language for a federal SI definition. Sixty days from September 29 is November 28. A statutory change would move the guide's scope with it.

Bottom line

The guide asks one question of every SI deployment: what can it do before a person stops it? A bank that answers it in writing, use case by use case, has its tiering done before an examiner asks. Comments on the interagency third-party proposal, which will shape the vendor side of SI risk, close November 16.

Sources