AI Development Company in Amsterdam

Financial Services / Banking

A Dutch banking group (€48 billion in assets, 4.2 million customers, 15,000 employees) was struggling with its anti-money laundering (AML) transaction monitoring system. Following ING's €775 million fine and ABN AMRO's €480 million settlement for AML failures, the Dutch banking sector faced intense regulatory scrutiny from De Nederlandsche Bank (DNB) and the Financial Intelligence Unit (FIU-NL). The bank's rule-based transaction monitoring system generated 2.4 million alerts per year — of which 97.8% were false positives. The 52,000 true alerts (2.2%) were buried in the noise, and the investigation team of 340 analysts spent most of their time dismissing false positives rather than investigating genuine suspicious activity. The consequences: 8 confirmed cases of money laundering that passed through undetected in the previous year (the alerts were generated but not investigated in time — the queue was 6 weeks deep), the investigation team was the bank's second-largest cost centre ($42 million annually), DNB had issued a formal remediation notice requiring the bank to improve its monitoring effectiveness within 18 months, and the bank estimated €200 million in potential fines if material AML failures were identified during the next regulatory examination.

We built an AI-powered transaction monitoring system that operates alongside (not replacing) the existing rule-based system — using ML to prioritise alerts and identify suspicious patterns that rules miss. (1) Alert prioritisation model: a gradient-boosted ensemble (LightGBM) trained on 3 years of historical alerts with their investigation outcomes (true positive, false positive, escalated to FIU, case closed). Features include: transaction characteristics (amount, frequency, currency, geography, counterparty patterns), customer profile (account age, expected transaction patterns, declared income, business type), network analysis (connections between accounts — shared counterparties, similar transaction patterns, geographic clustering), and temporal patterns (velocity changes — a sudden increase in transaction volume or a shift in transaction timing). The model assigns a risk score (0-100) to every alert, enabling the investigation team to work highest-risk alerts first rather than chronologically. (2) Pattern discovery: a separate unsupervised model (Isolation Forest combined with graph-based community detection) that identifies suspicious patterns the rule-based system misses. The model detects: structuring (transactions deliberately kept below reporting thresholds — the model identifies patterns of just-below-threshold transactions even when spread across days and accounts), layering (complex transaction chains designed to obscure the source of funds — the graph analysis identifies multi-hop transaction paths where money moves through a series of accounts in rapid succession), and trade-based laundering (over/under-invoicing of trade transactions — the model compares declared trade values against market prices for similar goods and routes). New patterns discovered by the model are reviewed by financial crime analysts — confirmed patterns become new monitoring rules, creating a continuous improvement cycle. (3) GDPR and EU AI Act compliance: the system processes personal financial data at scale and produces outputs that affect individuals (suspicious activity reports can lead to account freezing). We implemented: DPIA (Data Protection Impact Assessment documenting the processing, risks, and mitigations), legitimate interest assessment (AML monitoring has a legal obligation basis under AMLD6, but the AI component requires additional legitimate interest documentation), right to explanation (when a customer is flagged, the investigation team can see exactly which factors contributed to the risk score — using SHAP values to explain the model's reasoning), human-in-the-loop (the AI prioritises and recommends — human analysts make all investigation and reporting decisions, and no customer action is taken based solely on AI output), and EU AI Act high-risk documentation (AML AI is high-risk under the Act — we prepared conformity assessment documentation including risk management, data governance, and human oversight provisions). (4) Analyst workspace: a new investigation interface where analysts see: prioritised alert queue (highest risk first, with estimated investigation complexity), alert context (the AI's risk factors, relevant transaction history, customer profile, and network visualisation showing account connections), similar historical cases (cases with similar patterns and their outcomes — helping analysts make consistent decisions), and streamlined workflow (one-click disposition for obvious false positives with mandatory documentation for true positives and escalations).

False positive reduction: from 97.8% to 89.2% (reducing false positive alerts from 2.35 million to 1.07 million annually — a 54% reduction in alert volume). This is not achieved by reducing alert generation but by prioritising: the high-risk alerts are investigated first, and the lowest-risk alerts (which are overwhelmingly false positives) are batch-reviewed. True positive detection: improved from 52,000 to 78,000 confirmed suspicious cases annually (50% improvement) — the pattern discovery model identified 26,000 additional suspicious patterns that rules missed. Investigation efficiency: average investigation time per alert decreased from 45 minutes to 22 minutes (51% reduction) — the AI-generated context and similar case suggestions accelerate analyst decision-making. Investigation team: reduced from 340 to 240 analysts while handling 50% more true positives — $12.6 million annual cost savings. The freed analysts were redeployed to complex financial crime investigation. Regulatory compliance: DNB's remediation notice was satisfied — the regulatory examination confirmed the improved monitoring effectiveness. The examiner specifically noted the "meaningful human oversight" in the AI system as a compliance strength. FIU-NL reporting: suspicious transaction reports increased 34% (reflecting better detection, not more false positives) — improving the bank's contribution to the national financial intelligence picture. Undetected money laundering: zero confirmed cases of ML that passed through monitoring undetected in the first year (from 8 in the prior year). EU AI Act readiness: the conformity assessment documentation was prepared proactively — the bank is positioned for compliance when enforcement begins. Annual value: €12.6M team cost savings + €200M+ fine risk reduction + improved regulatory standing. Development investment: €520,000 over 22 weeks.