OCC Issues First AI Supervisory Framework, Ranking Autonomous Payment Agents as Highest Risk

The US Office of the Comptroller of the Currency has published its first supervisory framework for artificial intelligence in banking, directing examiners to weigh the consequences of an AI system going wrong rather than how advanced the technology is.

The guidance lands shortly after reports that several of the largest US banks had found unauthorised third-party AI agents operating inside customer accounts, although the regulator has not tied the document to those incidents.

What the framework says

The 22-page document, as described by Decrypt, draws a clear line between an AI system's technical capability and the damage it could do. In the OCC's framing, a relatively simple agent that is permitted to initiate financial transactions may pose more risk than a far more sophisticated model whose role is confined to drafting internal documents.

To make that judgement, examiners are invited to consider three broad questions: what happens if the system fails, whether its actions can be undone, and how much control humans keep over it. Those factors shift attention away from model size or benchmark performance and towards the practical exposure a bank takes on when software acts on its behalf.

The regulator sets out four illustrative risk tiers. At the top sit fully autonomous payment agents and systems that make credit decisions in real time. At the bottom are basic internal productivity tools. Systems in the higher tiers may justify tighter limits on autonomy, independent testing and a requirement for human sign-off before any consequential action goes ahead.

The framework also tackles a specific failure mode: AI that continues to operate even as it takes damaging decisions. Among the proposed protections are circuit breakers, caps on the pace of activity, narrowed permissions and controls to shut a system down in an emergency, mechanisms designed to stop a malfunctioning system before errors compound.

Outside agents and the account-access problem

One of the more notable sections deals with operational resilience and, in particular, AI agents that banks do not control. The OCC warns that external agents run by customers, fintech companies or other businesses can create risk for a bank even though the bank has no say over how those systems are built or operated.

For those cases, the guidance suggests examiners look at measures such as identity verification, delegated permissions, transaction limits, ongoing monitoring and the ability to suspend or revoke an agent's access. Its risk considerations explicitly extend to external agents connecting through bank channels.

This goes to the heart of a dilemma lenders are already grappling with: how to let customers use AI assistants to manage their money without handing autonomous software unlimited access to their accounts.

That tension came into public view less than two weeks before the framework appeared. According to an Axios report on 22 October, Morgan Stanley had shut out third-party agents altogether, and JPMorgan, Bank of America and Citi had placed limits on what such agents could reach. Bank executives quoted in that reporting called the situation "horrifying", and the episode raised a basic question about who is in charge of an autonomous agent once it can move someone's money.

Banks in those cases reportedly struggled to tell legitimate, customer-authorised software apart from systems acting beyond the permissions they had been given, sometimes taking steps that appeared to exceed users' instructions. The safeguards listed in the OCC document, from delegated permissions to revocable access, map closely onto that problem.

Why it matters

The framework is an early attempt by a major US banking supervisor to give examiners a consistent way of thinking about AI across a wide range of uses. Decrypt reports that it covers AI in payments, fraud detection, customer service and other banking operations.

Its central idea, that risk should be measured by consequence rather than capability, is significant for how banks deploy the technology. A consequence-based approach means a bank cannot rely on a system's strong performance in testing to escape scrutiny if that system can move funds irreversibly. Indeed, the illustrative scale places autonomous payment agents in the highest category regardless of how well they perform under test.

For institutions, that creates a practical incentive to design AI deployments with reversibility and human oversight in mind. Keeping a person in the loop for high-stakes decisions, capping transaction sizes or limiting what an agent can do may all help move a system into a lower tier of supervisory concern.

The approach also slots AI into existing supervisory structures rather than creating a parallel regime. Examiners are told to assess AI within established operational, compliance and cybersecurity risk categories, and banks are not required to set up a separate AI risk category. That should limit the administrative burden while still drawing AI into routine examination.

Limits and open questions

The OCC was explicit that the document does not impose new regulatory requirements or create enforceable standards. It is a tool for examiners, not a rulebook for banks. That leaves considerable room for interpretation, and the practical effect will depend on how individual examiners apply it and how banks respond to supervisory feedback.

The treatment of outside agents is likely to be the hardest area to manage. By definition, banks do not control those systems, so their options are largely limited to the edges: verifying who or what is connecting, constraining what it can do and cutting it off when necessary. Striking a balance between protecting accounts and allowing customers to use tools they have chosen is not a problem the framework resolves on its own.

There is also a terminology wrinkle. According to Decrypt, the document uses the term "Super Intelligence", which the Donald Trump administration insists on in place of "artificial intelligence". Despite the label, its scope covers conventional AI and autonomous agents, not only hypothetical systems that would exceed human intelligence. The choice of language may cause some initial confusion for readers expecting guidance aimed at far more speculative technology.

Finally, the four tiers are described as illustrative. That signals flexibility, but it also means banks may face uncertainty about precisely where a given system sits, particularly for tools that blend low-risk and high-risk functions.

What to watch next

The OCC has said it plans to update the guidance as the technology evolves, suggesting this is a starting point rather than a settled position. Several developments will indicate how much weight it carries in practice:

• Whether banks that have blocked or restricted outside agents begin to reopen access under controls resembling those in the framework.
• How examiners apply the risk tiers in actual supervisory reviews, especially for payment and credit systems.
• Whether later revisions move any elements from guidance towards firmer expectations.
• How fintechs and other operators of customer-facing agents adapt to tighter identity checks, permissions and transaction limits.

For now, the message from the regulator is that the critical question about any AI system in a bank is not how clever it is, but what it can break and whether anyone can stop it.