Citi's Fraser Sets ARC Controls Before AI Agents Act

Fraser's remarks at Sibos 2026 came as financial firms weigh how far to let AI agents go beyond giving information and into executing actions. She drew the line at identity, authentication, and oversight: without a way to confirm an agent is acting for a person, and for the bank and counterparties to confirm the transaction is legitimate, she is confident she won't have "wild agents without the right controls in place." Citi's answer so far is ARC, a layer every agent must pass through, though Fraser described the controls as "very early days."

The control question extends to how supervision itself might work. Fraser sketched a reversal of the usual ratio, where rather than one manager overseeing nine employees, a single agent could face nine oversight mechanisms, many run by other agents. She also tied the push to Citi's own scale, noting the bank moves roughly six trillion dollars daily, and to rising transaction volumes and client demand for round-the-clock, instant, safe service, citing Citi Token Services' digital deposit solutions as an example. The IMF's Dan Katz, speaking at the same event, put cyber risk at the top of the list, saying AI is speeding up how vulnerabilities are found and potentially exploited.

Fraser framed the whole effort against the tech axiom of moving fast and breaking things. "If you move fast and break things, we are failing in our mandate," she said, warning that breaking trust would be "a huge problem for the macro, for the markets, for everywhere." The outcome for Citi and its peers likely hinges on whether agent identity can be verified across institutions, not just inside one bank, and whether the oversight layer keeps pace as volumes rise.