Banks Learn That a Human Checkbox Is Not AI Oversight

A bank that handed suspicious activity reporting to AI agents saw review times drop from roughly six hours to about five minutes — then saw reviewers stop checking properly. The human sign-off before regulatory filing had become a formality, an executive recounted at a closed-door session during Singapore Fintech Festival 2025. The remedy: plant errors deliberately to keep people sharp.

The story emerged at the GFTN Insights Forum, where a roundtable co-produced with Zühlke gathered senior figures from lenders, regulators and tech vendors for a report on gearing up for an agentic future. KPMG figures cited there show 99% of companies want agents in production; only 11% have got that far.

Parking a person at the end of the pipeline is not oversight. Meaningful supervision has to be engineered, the report argues — periodic audits, deliberate friction, a clear-eyed call on which decisions truly need human judgement. One investment banker described pairing a writer model with a reviewer model, pushing quality to 70-80% versus 20% for a lone prompt with no second pass. Standard Chartered's Thorsten Neumann frames the goal as holding accuracy inside a defined risk appetite while weighing performance and cost.

Data is the quieter ceiling. Lenders reportedly pay big professional services bills just to pull information from software they already licence — vendor lock-in that caps what an agent can usefully do. Participants put data accessibility atop the regulatory-constraint list; one central bank's polling found 33% of regulated firms treat data-rule friction as a live obstacle. The fix, the report says, is API-native plumbing and portability standards from the start.

Governance is the structural gap. Multi-agent failures, one speaker noted, came from coordination breakdowns — communication gaps, memory drift, a small error leaping between agents — rather than weak models. That argues for compliance written into the first line of code, with audit trails and kill switches, plus the bankwide control plane Neumann expects to lean on as agent counts climb into the thousands.

Why it matters for Singapore: MAS's recently published expectations for responsible AI, which financial institutions are due to adopt on 7 October 2026, push these oversight questions onto local board agendas. Singapore's pitch as a trusted hub for enterprise AI depends on banks here proving agents can be governed, not merely deployed — fixing data portability and keeping human review genuine as rollouts scale.