Financial Crime Compliance Model — Retail Bank, London, UK

The Situation

A UK retail bank's compliance team processed 400 to 600 alerts daily from a generic LLM-based transaction monitoring layer. The model had been trained on general financial text. It had no knowledge of the bank's customer base, transaction patterns, or the 3 years of adjudicated alert decisions sitting in the compliance team's own case management system.

False positive rates ran above 70%. Analysts were spending the majority of their day closing alerts they already knew were noise. Genuine risk cases were getting slower attention as a result.

What Amorisoft Did

Amorisoft fine-tuned the existing model directly on the bank's adjudicated alert history. 3 years of case outcomes, analyst notes, and escalation decisions were structured into a training dataset covering 14 alert categories across retail, SME, and private banking segments.

The model learned from decisions the compliance team had already made, across enough volume to generalise reliably.

Evaluation ran against a held-out set of 6 months of historical alerts before any production deployment. The false positive rate on the held-out set came in at 31%, down from 71% on the same data using the original model.

The bank's IT security team required a full data handling audit before training could begin. Amorisoft completed the audit in week 2. Training started week 3.

Results

First-pass false positive rate dropped from 71% to 30% within 60 days of go-live. Analyst review time freed per day came in at 31% across the team. The model's outputs passed a regulator audit on explainability at 97%, which the compliance director noted was the result the bank had been most uncertain about going in.