FDE Times: AML/KYC for FDEs — get compliance sign-off before an agent reviews alerts
A 6 November FDE Times guide argues that at a bank a working demo does not turn an AML alert agent on: compliance signs off only when validation evidence, a miss-rate trade-off and per-recommendation explainability are agreed in week one.
FDE Times walks through putting an agent into a bank AML alert queue: the agent may summarise KYC files, recommend close or escalate, and draft SAR narratives, but compliance—not engineering—owns go-live. Drawing on Wolfsberg’s responsible-innovation pillars, the piece maps validation, balancing model risk against financial-crime risk, and explainability into concrete FDE deliverables: a pilot on closed alerts with thresholds, a table of time saved versus missed cases, and cited reasons plus an audit log for every recommendation.
In the guide’s pilot framing, engineering celebrates the bulk of alerts the agent would close faster; compliance asks how many of those the investigators had escalated. Those missed cases become the meeting, not a single accuracy percentage. The agent must only recommend; a human presses the decision button and stays accountable.
Narrative templates, mandatory log fields and which model-risk policy covers generative AI are institutional choices—FinCEN SAR FAQ practice and US banking model-risk guidance do not let the FDE invent them. Week one should lock scope, miss threshold and governing policy in writing before prompts and log schemas are written. For FDEs aiming at finance roles, showing a control-function approval with opened miss cases beats claiming prompt skill alone.
Source: fdetimes.net