NLP Development in Amsterdam, Netherlands

Logistics / International Trade

A Dutch logistics and customs brokerage company (processing 420,000 import and export declarations annually, 850 employees, handling goods worth €18B for 2,400 clients across 65 countries) managed trade document processing through a team of 180 customs declarants and trade compliance specialists. The document processing workflow: (1) Document receipt: for each shipment, the company received a document package from the shipper — commercial invoice, packing list, bill of lading, and (depending on the goods and origin) certificate of origin, phytosanitary certificate, inspection certificate, conformity certificate, and other trade documents. Average: 6-8 documents per shipment. Documents arrived in 14 languages (primarily English, German, French, Chinese, Spanish, Italian, Turkish, and Dutch). (2) Data extraction: customs declarants manually read each document, identified the relevant data (goods description, quantity, weight, value, HS code, country of origin, incoterms, parties), and entered this data into the customs declaration system (Douane ATLAS interface). For each shipment, this involved extracting 25-40 data fields from 6-8 documents. Average extraction time: 35 minutes per shipment for standard goods, 60-90 minutes for complex goods (chemicals, dual-use items, goods requiring special permits). (3) Tariff classification: the declarant determined the correct HS/CN code for each product line — mapping the natural language goods description to the appropriate classification code. This required expertise: classifying "polyester fabric, woven, dyed, weight 180g/m²" requires knowing that this falls under CN 5407.52 (not 5407.51 for unbleached, not 5407.53 for yarn-dyed, not 5407.54 for printed). Classification errors directly affect duty rates — and the Netherlands customs authority (Douane) conducts post-clearance audits with penalties for misclassification. (4) Trade agreement assessment: for goods from countries with EU trade agreements, the declarant assessed whether preferential duty rates applied — checking rules of origin against the certificate of origin and goods description. Incorrect preferential claims result in duty recovery with interest. (5) Compliance screening: every shipment required sanctions screening (EU, UN, OFAC), dual-use goods assessment, and embargo checking. The problems: (1) Processing time: average 45 minutes per declaration (combining extraction, classification, and compliance). Volume: 420,000 declarations × 45 minutes = 315,000 hours of declarant work annually. At current capacity, the company could not scale to handle client growth without proportional hiring. (2) Multilingual challenge: declarants needed to read documents in 14 languages. Most declarants were fluent in Dutch, English, and German — documents in other languages (Chinese, Turkish, Arabic) required translation support, adding delays and costs. (3) Classification accuracy: tariff classification error rate: 4.8% (measured by Douane post-clearance audits). Each classification error potentially incurred: duty underpayment recovery (with interest), administrative penalties (ranging from €150 to €50,000 depending on severity and frequency), and client relationship damage. (4) Preferential origin errors: 6.2% of preferential claims were incorrectly applied — either claiming preference when rules of origin were not met, or failing to claim preference when entitlement existed. Incorrect claims: customs penalties. Missed claims: duty overpayment averaging 4-8% of goods value. (5) Declarant shortage: experienced customs declarants require 2-3 years of training. The Netherlands (and the EU generally) faces a shortage of qualified customs professionals — the company had 12 unfilled declarant positions.

We built a multilingual trade document NLP platform over 14 weeks. (1) Multilingual Document Intelligence Engine: processing trade documents in 14 languages. The engine: classifying document types (commercial invoice, bill of lading, certificate of origin, packing list — across all language variants), extracting 25-40 data fields from each document (goods description, quantity, weight, value, country of origin, parties, HS code if stated, incoterms, shipping marks, container numbers), processing Chinese-language documents (commercial invoices from Chinese manufacturers — extracting 品名/product name, 数量/quantity, 金额/amount, and other fields from Simplified Chinese text), handling mixed-language documents (a German commercial invoice with English product descriptions and French payment terms), and cross-referencing between documents (verifying consistency of quantities, values, and descriptions across invoice, packing list, and bill of lading). Extraction accuracy: 96.8% across all languages and document types. Processing time: 4 minutes per shipment document package (from 35-90 minutes manual). (2) Tariff Classification Engine: automating HS/CN code determination from goods descriptions. The engine: processing natural language goods descriptions in any language (translating to English for classification if needed, then mapping to the CN nomenclature), applying classification rules (GRI — General Rules of Interpretation — the systematic rules for tariff classification), suggesting CN codes with confidence scores (presenting the top 3 most likely classifications with reasoning — "classified under 5407.52 because: woven fabric (Chapter 54), polyester (heading 5407), dyed (subheading .52)"), flagging ambiguous classifications for human review (goods that could reasonably fall under multiple headings), and learning from declarant corrections (when a declarant overrides the AI suggestion, the system learns from the correction). Classification accuracy: 92.4% exact match at 8-digit CN level (from 95.2% manual — but the AI catches its own uncertainty, flagging 15% of classifications for human review; within the 85% auto-classified, accuracy is 97.1%). (3) Origin and Preference Engine: automating trade agreement assessment. The engine: identifying the applicable trade agreement (if any) for the origin/destination pair, extracting rules of origin from the trade agreement text (each EU trade agreement has different rules of origin — some value-based, some process-based, some with product-specific rules), assessing whether the goods meet the applicable rules based on certificate of origin data and goods description, and recommending preferential or MFN (Most Favoured Nation) duty rate application. Preference assessment accuracy: 96.3% (from 93.8% manual). Additional preference claims identified: 2,800 annually (shipments where preference was available but not claimed under the manual process — saving clients an average of €420 per shipment in duty). (4) Compliance Screening Engine: automating sanctions and trade compliance checks. Screening time: 2 minutes per shipment (from 8-12 minutes manual). Entity screening accuracy: false positive rate reduced by 65% through multilingual entity resolution. (5) Declaration Generation Engine: producing Douane ATLAS-compatible declaration data from NLP-extracted information. The engine: populating all required declaration fields from extracted data, generating declaration in ATLAS format for electronic submission, flagging fields requiring declarant verification (values below extraction confidence threshold), and producing supporting documentation packages for potential audit.

Declaration processing time: average 12 minutes per declaration (from 45 minutes) — including human review of AI-generated declarations. Straight-through processing (declarations requiring minimal human review): 42% of standard goods declarations. Multilingual processing: all 14 languages processed with equal accuracy — eliminating translation delays and costs. Tariff classification accuracy: 97.1% for auto-classified items (85% of total) — above the manual rate of 95.2%. Overall classification errors detected by Douane: reduced from 4.8% to 1.2%. Preferential origin: 2,800 additional preference claims identified annually — saving clients €1.18M in duties. Incorrect preference claims: reduced from 6.2% to 1.8%. Declarant capacity: effective capacity increased by 65% — the 168 active declarants (12 positions still unfilled) now process the equivalent of 277 declarants' manual workload. Growth: the company onboarded 340 new clients in the year following deployment without additional declarant hiring. Annual value: €4.2M (declarant efficiency — avoided hiring and capacity increase), €1.18M (client duty savings from preference identification), €820K (reduced classification error penalties and audit costs), €450K (compliance screening efficiency) = €6.65M. Development cost: €295,000. Payback: under 17 days.