AI Copilot Development in Amsterdam

Financial Services / Trade Finance

A major Dutch bank (EUR 280B total assets, trade finance division processing 135,000 documentary transactions annually — letters of credit, documentary collections, guarantees, and standby LCs — across 85 countries) faced capacity and compliance challenges in trade finance operations. (1) Document review bottleneck: each documentary credit transaction required manual review of trade documents against the terms of the letter of credit and UCP 600 (Uniform Customs and Practice for Documentary Credits) standards. A typical documentary credit presentation included 8-15 documents: commercial invoice, bill of lading or airway bill, insurance certificate, certificate of origin, packing list, inspection certificates, and various country-specific documents. Each document had to be checked for: compliance with the LC terms (exact matching of goods description, quantities, amounts, shipping dates, and other stipulated conditions), internal consistency (documents must not conflict with each other — e.g., the goods description on the commercial invoice must match the bill of lading), UCP 600 compliance (specific rules for each document type — e.g., bills of lading must indicate "on board," insurance must cover at least 110% of CIF value), and sanctions screening (verifying that parties, vessels, ports, and routes do not involve sanctioned entities or jurisdictions). Expert document checkers reviewed each presentation — average review time: 45 minutes per presentation for straightforward transactions, 90-120 minutes for complex or discrepant presentations. With 135,000 annual transactions and 85 document checkers, capacity was stretched. Backlogs during peak periods (quarter-end, seasonal trade flows) caused SLA breaches — the bank was committed to 5-business-day turnaround for document examination, and was achieving only 72% on-time during peak periods. (2) Discrepancy identification inconsistency: UCP 600 requires strict compliance — documents must conform exactly to LC terms. However, discrepancy identification varied between checkers. Some were overly strict (raising discrepancies that were not material — causing unnecessary delays and client friction), while others missed genuine discrepancies (accepting non-compliant documents — exposing the bank to financial risk). The bank's quality assurance sampling found a 12% disagreement rate between checkers reviewing the same presentations — meaning different checkers would reach different conclusions on whether the same documents were compliant. (3) Sanctions and trade compliance complexity: the sanctions landscape had become increasingly complex — EU sanctions (multiple packages targeting Russia, requiring ongoing monitoring of vessel movements, goods classifications, and beneficial ownership), US secondary sanctions (affecting transactions with certain countries and entities — relevant because many transactions involved USD), and Dutch national sanctions regulations. Each transaction required screening against multiple sanctions lists — and the regulatory consequences of a sanctions breach were severe. (4) Trade-Based Money Laundering (TBML) detection: DNB and European regulatory authorities had increased focus on TBML — the use of trade transactions to launder money. Red flags included: over-invoicing or under-invoicing, unusual goods descriptions, mismatches between trade flows and business profiles, and complex transaction structures without apparent economic rationale. The bank's existing TBML screening was largely rule-based — generating a high false-positive rate (92%) that consumed compliance analyst time reviewing alerts that were overwhelmingly benign. (5) Multilingual document processing: trade documents arrived in multiple languages — Dutch, English, German, French, Spanish, Arabic, Chinese, and others — reflecting the bank's global trade finance activities. Non-English documents required either translation or review by language-specialist checkers, creating bottlenecks for certain language pairs.

We built a trade finance copilot over 16 weeks, integrated into the bank's trade finance operations workflow. (1) Automated Document Extraction and Checking: intelligent document analysis. Document ingestion: trade documents (arriving primarily as PDF via SWIFT and email) automatically ingested, classified by document type, and parsed. The copilot used specialised document extraction models trained on trade documents — recognising the specific formats and fields of commercial invoices, bills of lading, certificates of origin, and other standard trade documents. Field extraction: the copilot extracted structured data from each document — goods descriptions, quantities, amounts, currencies, parties, dates, vessel names, ports, and all other fields relevant to LC compliance checking. Cross-document validation: the copilot automatically checked: LC terms compliance (comparing extracted document data against the LC terms — identifying exact discrepancies such as "LC requires shipment by 15 March; bill of lading shows shipped on board 18 March"), inter-document consistency (checking that goods descriptions, quantities, and amounts are consistent across all documents in the presentation), and UCP 600 compliance (applying the specific rules for each document type — checking bill of lading requirements under UCP 600 Articles 19-25, insurance requirements under Article 28, invoice requirements under Article 18). Discrepancy report: for each presentation, the copilot generated a structured discrepancy report — listing each identified discrepancy with: the relevant LC term, the specific non-compliance found, the UCP 600 article applicable, and a severity classification (material discrepancy vs. minor inconsistency). The document checker reviewed the copilot's report rather than reviewing the documents from scratch — confirming or adjusting the copilot's findings. (2) Sanctions Screening Enhancement: intelligent sanctions analysis. Multi-list screening: the copilot screened all parties, vessels, ports, and routes against: EU consolidated sanctions list, Dutch national sanctions regulations, US OFAC SDN list (relevant for USD-denominated transactions), and other applicable sanctions lists. Contextual analysis: beyond simple name matching, the copilot performed contextual analysis — evaluating whether identified matches were true positives (the sanctioned entity) or false positives (a different entity with a similar name). The copilot considered: entity type (individual vs. company), jurisdiction, industry, and transaction context. For vessel screening: the copilot checked vessel names against sanctioned vessel lists and analysed vessel movement patterns (AIS data integration) to identify potential sanctions evasion through vessel-to-vessel transfers or unusual routing patterns. Alert quality: sanctions alert false-positive rate reduced from 87% to 31% — because the copilot's contextual analysis eliminated obvious false positives before human review. (3) TBML Detection Enhancement: intelligent trade compliance analysis. Pattern analysis: the copilot analysed each transaction for TBML red flags — comparing invoice prices against commodity price benchmarks, analysing trade flows against the client's known business profile, identifying unusual transaction patterns (e.g., a company suddenly trading in goods outside its normal business), and evaluating the economic rationale of complex transaction structures. Anomaly scoring: each transaction received a TBML risk score based on multiple factors — pricing anomalies, counterparty risk, geographic risk, goods description analysis, and transaction pattern analysis. Analyst triage: compliance analysts reviewed high-scored transactions — the copilot provided a pre-analysis that explained why the transaction was flagged and what specific anomalies were identified. TBML alert false-positive rate: reduced from 92% to 38% — because the copilot's contextual analysis identified transactions with legitimate explanations for apparent anomalies. (4) Multilingual Processing: cross-language document analysis. Language detection and processing: the copilot automatically detected document language and processed documents in their original language — no translation required. The copilot could compare a Chinese commercial invoice against English-language LC terms, extracting the relevant fields from each and performing the compliance check across languages. Accuracy: multilingual processing accuracy equivalent to single-language processing — the copilot's multilingual models understood trade document terminology in all major trading languages. (5) Knowledge Base and Learning: continuous improvement. Decision database: every document checking decision (copilot analysis plus checker confirmation/correction) fed into the copilot's knowledge base. Over time, the copilot learned the bank's specific interpretation of UCP 600 requirements — for example, how the bank handled certain edge cases in goods description matching or date computation. Best practice sharing: the copilot's consistent application of checking standards effectively transferred the expertise of the bank's best document checkers to the entire team — new checkers could achieve experienced-checker quality faster because the copilot provided expert-level pre-analysis.

Document review throughput: checkers processed 40% more presentations per day — reviewing the copilot's pre-analysis (15-20 minutes) rather than conducting the full document review from scratch (45-90 minutes). SLA performance: on-time document examination improved from 72% to 96% during peak periods — eliminating the backlog that had caused client friction and reputational risk. Discrepancy consistency: checker disagreement rate dropped from 12% to 3% — the copilot's structured analysis provided a consistent baseline that reduced subjective variation between checkers. Sanctions screening: false-positive rate reduced from 87% to 31% — freeing compliance analyst capacity that had been consumed reviewing benign alerts. Zero sanctions breaches in the first year of operation. TBML detection: alert false-positive rate reduced from 92% to 38%, while true-positive detection rate increased by 24% — the copilot identified TBML patterns that the rule-based system had missed. DNB supervisory feedback acknowledged "material improvement" in trade compliance monitoring. Multilingual processing: eliminated the language-specialist bottleneck — all documents processed without translation delays regardless of language. Processing time for non-English documents reduced by 60%. Staff reallocation: 15 document checker positions (natural attrition, not layoffs) redirected to higher-value roles — complex advisory, client relationship management, and trade structuring. Client satisfaction: trade finance NPS improved from +18 to +41 — driven by faster turnaround and fewer unnecessary discrepancy raises. Annual value: EUR 8.2M (throughput improvement — processing 40% more transactions without additional headcount), EUR 3.4M (sanctions and TBML compliance efficiency — reduced false-positive investigation costs), EUR 2.1M (eliminated language-specialist bottleneck and translation costs), EUR 1.4M (SLA improvement — retained clients who had been considering competitors due to slow turnaround) = EUR 15.1M. Development cost: EUR 340,000. Payback: under 9 days.