Bank AI agents cut a report task from six hours to five minutes

A participant at a closed roundtable at the Singapore Fintech Festival 2025 runs a banking operations platform. The participant described AI agents that work on suspicious activity reports. A human inspected each report before it went to the regulator. The work time fell from six hours to five minutes. The participant said that the human check then began to feel like a formality.

What the report says

A GFTN report, made with Zühlke, gives this account. It says that approval by a person is not sufficient. It asks for audits and for steps that keep the reviewer active. It also asks for audit trails and kill switches from the start of the code.

The proof

• The source gives one account of one platform, from a person it does not name.
• It gives no data on errors or on how the human check changed.
• One banker said a writer model with a reviewer model gave 70 to 80% quality, against 20% for one prompt.

What is not known

The report is a summary of a closed roundtable. Zühlke, which helped to make it, offers services to banks. The Monetary Authority of Singapore recently announced supervisory expectations on responsible AI.

Insight: A human approval step is not a control by itself

Citigroup and Bank of America say a human must agree before agents move money. Only 14% of organizations control the agents that 88% use. In a live test, a gateway stopped a remove-all task, but prompt instructions did not. Banks use the human check as a control, but this account shows that it fades with speed. Controls must work without a person who checks. Regulators will probably test the check.