America's bank supervisor will judge AI by what it can do to money, not by how clever it is

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

On 4 November the Office of the Comptroller of the Currency, which supervises national banks in the US, issued a supervisory framework for its examiners on banks' use of AI and AI agents. The release uses its own terms, "Super Intelligence (SI)" and "agentic SI". It says examiners should not treat the technology as a separate risk category, that banks should not have to define a separate risk appetite for it, and that the OCC will update the framework regularly and stay technology-neutral.

Reuters, which saw the document, reported the part that will shape examinations. Examiners should focus on the consequences of a potential AI mistake rather than how sophisticated the model is. A simple AI agent that is authorised to initiate transactions could be riskier than a sophisticated model used for internal purposes. Higher-risk uses may need tighter limits on autonomy or more direct human oversight. Reuters noted that the guidance is not legally binding, but that banks usually follow such guidance closely.

Two days later BMO announced AI Agent Assist for its Virtual Connect contact centre, which takes more than 500,000 calls a year about credit-card transactions, at an average of 12 to 14 minutes each. The agent does transaction matching, policy retrieval and case preparation for employees to review. BMO is aiming for 20% to 30% less time to resolve eligible calls and automatic processing of about 40% of eligible refund requests. These are targets, not results. BMO says employees remain accountable for reviewing the AI's output and making servicing decisions.

Why it matters

The OCC has turned the usual AI inventory upside down. Most banks list their models by type and sophistication. The examiner will now ask a different question about each one: what can it do to money and to customers if it is wrong? On that test, a large model that drafts internal summaries may be lower risk than a small rules-plus-LLM agent that can start a refund.

The examiner's question is no longer how clever the model is. It is what the model can do if it is wrong.

BMO's design fits that test. The agent prepares the case, and a person makes the servicing decision. Automatic processing of eligible refunds is where the agent's reach grows, and that is where I would expect an examiner to look first: which refunds are eligible, the value limit, and who reviews the exceptions.

For any bank, the practical step is to rank AI uses by authority rather than by model: read-only, recommend, prepare for a person, and execute. Then match the controls to each tier, with tighter autonomy limits and closer oversight at the top. The same logic sits behind this issue's cover story on India's payment rails.

What I'd watch

• Whether the full framework text is published, and how it treats agents supplied by vendors rather than built in-house.
• BMO's actual handle-time and automatic-refund results against its 20% to 30% and 40% targets.
• Whether other US banking agencies adopt the same consequence-based test.