OCC issues first AI supervisory framework, flags risks from outside AI agents
In brief
- The OCC issued its first supervisory framework for artificial intelligence in banking.
- Examiners are told to weigh AI failure consequences, not raw model capability.
- Fully autonomous payment agents sit in the highest of four illustrative risk tiers.
- The OCC said the document creates no new requirements or enforceable standards.
Consequences over capability
Under the 22-page framework, examiners can consider the consequences of failure, whether an action can be reversed, and how much control humans keep over the system, according to Decrypt's report on the document. Model size isn't the yardstick. The regulator said a simple agent authorized to initiate financial transactions may present greater risk than a more advanced model that's limited to writing internal documents.
The OCC outlined four illustrative risk tiers. Fully autonomous payment agents and real-time credit decision systems sit in the highest category, while basic internal productivity tools land in the lowest. Decrypt notes the scale places autonomous payment agents at the top regardless of how well they perform in testing. Higher-risk systems may warrant tighter limits on autonomy, independent testing and human approval before consequential actions go through. For systems that keep running while making harmful decisions, the suggested safeguards include circuit breakers, rate limits, restricted permissions and emergency shutdown controls.
Outside agents
In a section on operational resilience, the OCC warned that external agents run by customers, fintechs or other businesses can create risks for banks even when banks don't control those systems. It suggested examiners look at identity verification, delegated permissions, transaction limits, monitoring and the ability to suspend or revoke access.
The timing stands out.
The guidance arrived less than two weeks after reports that major banks were finding unauthorized third-party AI agents operating inside customer accounts, sometimes taking actions that appeared to go beyond users' instructions. According to an Axios report on October 22 relayed by Decrypt, Morgan Stanley was blocking outside agents, while JPMorgan, Bank of America and Citi were restricting access.
Bank executives described the situation as "horrifying,"
That's according to the Axios reporting. The OCC hasn't linked the framework's issuance to those incidents.
What it doesn't do
The framework isn't a new rulebook. It gives examiners a way to evaluate AI within existing operational, compliance and cybersecurity risk categories, and banks aren't required to set up a separate AI risk category. Its scope covers AI used in payments, fraud detection, customer service and other banking operations, and it extends its risk considerations to external AI agents connecting through bank channels. The OCC said it intends to update the guidance as the technology develops.
Frequently asked questions
Does the OCC AI framework create new rules for banks?
No. The OCC stressed that the document doesn't create new regulatory requirements or enforceable standards for banks. It gives examiners a way to evaluate AI within existing operational, compliance and cybersecurity risk categories, and banks aren't required to set up a separate AI risk category.
How does the OCC framework rank AI risk in banking?
The OCC outlined four illustrative risk tiers. Fully autonomous payment agents and real-time credit decision systems sit in the highest category, while basic internal productivity tools sit in the lowest. The scale places autonomous payment agents at the top regardless of how well they perform in testing. Examiners can weigh the consequences of failure, reversibility and human control.
What does the OCC say about outside AI agents?
The OCC warned that external agents operated by customers, fintechs or other businesses can create risks for banks even when banks don't control those systems. It suggested examiners consider identity verification, delegated permissions, transaction limits, monitoring and the ability to suspend or revoke access.
