US bank regulator rates AI by consequence: a test NZ firms can use

What the OCC published

On 4 November the US Office of the Comptroller of the Currency issued Bulletin 2026-54, a supervisory framework telling its bank examiners how to assess AI in everyday banking. The OCC calls the technology "Super Intelligence", yet the bulletin's footnote defines the label as the technologies previously covered by the US statutory term for artificial intelligence. Reuters reported the OCC's description of the document as its first guidance on how examiners should probe AI use.

The timing closes a gap. The OCC's April model risk bulletin left generative and agentic AI outside its scope, so examiners had no dedicated reference for agents. The new document runs to 22 pages and organises 16 components under six pillars, which Retail Banker International lists as purpose, risk, control, oversight, resilience and assurance.

The bulletin says it sets no enforceable standards and that banks aren't expected to use it for self-assessment. Reuters still expects it to carry weight, noting that banks typically comply closely with such guidance.

Consequence outranks sophistication

The OCC tells examiners to keep AI inside existing risk categories, and banks don't need a separate AI risk appetite. Classification instead depends on the potential consequences of failure, the authority handed to the system and the degree of human oversight. Reuters gives the clearest example: a simple agent authorised to initiate transactions could be riskier than a sophisticated model kept to internal work.

Published summaries describe four illustrative tiers. Elevate Consult's breakdown puts autonomous payment agents and real-time credit decisions in Tier 1, and internal productivity tools in Tier 4. The same breakdown says strong test results don't move a fully autonomous payment agent out of Tier 1, because testing lowers the odds of failure and leaves the damage from a failure unchanged. Elevate also stresses that the tiering is an example for examiners and carries no mandated status.

We think the tier logic is the most useful part for businesses outside banking. Vendor demos invite buyers to compare how capable a model looks. The OCC's approach asks what the system can touch and whether a mistake can be undone before anyone's harmed. An agent that pays supplier invoices in an ERP, issues refunds or edits bank account details sits much closer to the OCC's Tier 1 examples than a chatbot summarising policy documents, even if the chatbot runs on the more advanced model.

Why a US bank document matters in New Zealand

The OCC has no authority here, and no NZ business has to do anything because of Bulletin 2026-54. New Zealand's own Responsible AI Guidance for businesses is voluntary. Regulators on both sides of the Tasman are pushing harder on the same issues, though.

At an Auckland roundtable in October, the Reserve Bank and the FMA set out expectations on AI-linked cyber risk. The FMA's Louise Unger said: "Boards and senior management need to know where AI is being used in their business." The same expectations make clear that outsourcing to AI or cloud providers leaves operational risk with the regulated entity. In Australia, APRA's April letter found that identity and access management hasn't adjusted to non-human actors such as AI agents, and it requires credible fallback processes where AI supports critical operations. The FMA's review of AI in financial advice is still at the fact-finding stage, with no findings on its page yet.

The trans-Tasman material tells boards to find their AI and name an owner. The OCC's framework adds a method for deciding how much control each use deserves, which is the question operations and IT leaders get stuck on once the inventory exists. A tiered register also gives a firm something concrete to show a regulator, auditor or insurer who asks how AI risk is being managed, even where no NZ rule demands one.

Australia adds a harder deadline. From 10 December, entities covered by the Australian Privacy Principles must describe in their privacy policies the kinds of decisions computer programs make or help make, where those decisions could significantly affect individuals and personal information is used. Allens' reading of the guidance is that human review doesn't take a tool out of scope. NZ firms with Australian operations should get advice on whether they're covered, and a tiered register would supply most of the raw material for that disclosure.

Agents your business doesn't control

Elevate's summary says the framework also reaches outside agents acting for customers through bank channels. The concern is live in the US, where large banks including JPMorgan and Bank of America have restricted outside agents' access to customer accounts, and liability for failed automated payments is still being debated. A New Zealand business with a customer portal could meet the same problem when a customer's agent logs in with the customer's credentials, and that firm's terms of use probably don't address the scenario.

The framework also separates availability from safe behaviour. An agent can stay online while taking harmful actions, so a conventional uptime alert won't catch the failure. Elevate lists circuit breakers, rate limits, least-privilege permissions and kill switches as the containment tools.

Plenty remains unknown. The OCC says it will update the framework over time, and examiners' use of it in practice hasn't been tested. Neither the Reserve Bank nor APRA has said whether it'll adopt a comparable tiering model.

Next steps

• List every AI tool and agent in use, including AI features vendors have switched on inside existing products, with a named owner and a note of the systems each one can change.
• Score each entry on consequence of failure, reversibility and level of human control, then sort entries into four tiers. Treat payments, credit, refunds and master data changes as top tier, whatever the vendor's test results show.
• Cap authority in the top tiers with dedicated service accounts, transaction limits, rate limits, human approval before money moves and a stop switch you've tested.
• Check that reviewers can really review. Elevate's summary says examiners look at how many agents or alerts one person can credibly monitor, and a reviewer needs the authority to reject or halt an action.
• Update vendor contracts to require notice of model changes and agreed notification times for AI incidents, then retest when a model changes.
• Decide your position on customers' agents and write it into your terms of use and access controls.
• If you trade in Australia, use the register to test your privacy policy against the 10 December obligation.

Firms that want help tiering their AI use can talk to our AI consultancy team. Those already building agents can read how we approach AI agents.