AI Agent Development Company in Amsterdam

Logistics / Customs Compliance / Amsterdam

An Amsterdam-based logistics company (EUR 420M revenue, 1,800 employees, operating 6 warehouses in the Netherlands and Belgium, processing 28,000 customs declarations monthly for goods moving between the EU, UK, US, China, and 40+ other countries) needed to automate its customs compliance operations. Post-Brexit trade between the EU and UK had doubled the company's customs declaration volume — but the customs team had not doubled. (1) Declaration volume — 28,000 monthly customs declarations across: import declarations (IMA — invoer aangifte), export declarations (ECS — Export Control System), transit declarations (NCTS — New Computerised Transit System), and UK-specific declarations (CHIEF/CDS — post-Brexit requirements). Each declaration required: goods classification (HS/CN code — identifying the correct 10-digit commodity code from 15,000+ possible codes), value determination (customs value calculation per EU valuation rules — transaction value, adjustments for transport, insurance, royalties), origin determination (preferential origin for FTA benefits — EUR.1 certificates, REX self-certification, UK-EU TCA origin rules), and licence/permit verification (certain goods require import/export licences — dual-use goods, agricultural products, chemicals). (2) Error rate — the customs team (22 declarants) processed declarations manually, cross-referencing product descriptions with HS code databases, calculating duties, and verifying permits. Error rate: 4.2% of declarations contained errors (wrong HS code, incorrect value, missing permits). Customs errors resulted in: EUR 680K annual penalties from Dutch Customs (Douane), shipment delays averaging 3.2 days per error, and customer dissatisfaction (delayed deliveries, unexpected duty charges). (3) Staffing — the 22-person customs team was at capacity. Recruiting experienced customs declarants was extremely difficult (specialised profession with shrinking workforce). Training new declarants took 12-18 months to reach competency. The company was turning away new business because it lacked customs processing capacity. (4) OR requirement — the OR had been briefed on the automation concept and expressed conditional support — provided that: the system augmented rather than replaced customs declarants (no involuntary redundancies), employee data was not used for performance monitoring, and declarants received training and career development opportunities in the new AI-augmented workflow.

We built an autonomous customs compliance agent system over 14 weeks, with OR consultation completed in weeks 1-3: (1) OR engagement: OR briefing at week 1 — presenting the AI agent concept as productivity enhancement. Key commitments: no involuntary redundancies (automation enables the company to handle growth without proportional hiring, not reduce existing headcount), customs declarants transition to "customs intelligence analysts" — reviewing AI decisions, handling complex cases, and developing trade compliance strategy, individual performance metrics not generated by the AI system (aggregate team metrics only), and 40 hours of training per declarant on AI-augmented workflows. OR consent obtained at week 3 (WOR Article 27 for personal data processing policy, WOR Article 25 advisory for technology investment). (2) HS classification agent: autonomous agent classifying goods by HS/CN commodity code. Input: product descriptions (from commercial invoices, packing lists, and product databases — in Dutch, English, German, French, and Chinese). Classification process: the agent analyses product descriptions, identifies key characteristics (material composition, function, manufacturing process), and determines the correct HS code. Training: the agent was trained on 2.4M historical declarations (the company's 10-year declaration history plus Dutch Customs rulings database — BTI — Binding Tariff Information). Classification accuracy: 94.7% on the first pass (verified against a test set of 5,000 declarations annotated by senior declarants). Confidence scoring: each classification includes a confidence score. Classifications below 85% confidence automatically escalated to a human declarant for review (approximately 12% of declarations). Complex classifications: goods with classification ambiguity (could be classified under multiple HS codes) flagged with analysis showing the arguments for each classification and the duty rate implications — the human declarant makes the final decision with full context. (3) Customs value agent: autonomous agent calculating customs value per EU valuation rules. Transaction value: extracted from commercial invoices. Adjustments: the agent applies required adjustments — transport costs (Incoterms-based allocation), insurance, royalties and licence fees, assists (goods or services provided by the buyer to the seller — e.g., moulds, designs), and price adjustments for related-party transactions. Duty calculation: customs duty calculated using the correct HS code rate, anti-dumping duties (checked against EU ADD database), and preferential duty rates (when origin documentation supports FTA preference). (4) Origin determination agent: autonomous agent assessing preferential origin for FTA benefits. Origin rules: the agent applies origin rules for: EU-UK TCA (Trade and Cooperation Agreement — complex origin rules with product-specific requirements), EU-Japan EPA, EU-South Korea FTA, EU-Canada CETA, EU GSP (Generalised Scheme of Preferences for developing countries), and 30+ other FTAs. For each shipment: the agent analyses: bill of materials (if available — for manufactured goods), supplier declarations of origin, and product characteristics against the specific FTA origin rules. Origin documentation: the agent generates or verifies: EUR.1 movement certificates, REX self-certifications, statements on origin per EU-UK TCA, and supplier declarations. (5) Declaration generation agent: autonomous agent assembling complete customs declarations. The agent compiles: classified goods (HS codes from classification agent), calculated values (from value agent), origin assessment (from origin agent), licence verification (checked against Dutch licensing requirements — CDIU for dual-use, RVO for agricultural), and submits declarations to Dutch Customs (Douane) through the AGS (Aangiftesysteem) interface. Pre-submission validation: the agent validates each declaration against Dutch Customs rules before submission — catching formatting errors, missing fields, and logical inconsistencies. (6) Exception handling: when agents encounter situations outside their confidence range, they escalate with full context. Escalation package: the agent provides the human declarant with: its best classification/value/origin assessment, alternative options with analysis, relevant precedents from the BTI database, and specific questions for the declarant (focused on the ambiguity — not a general "please review"). Declarant decision: the human declarant resolves the escalation, and their decision feeds back to improve the AI agent's future handling of similar cases. (7) AP compliance: DPIA completed before deployment. The AI agents process business data (product descriptions, commercial values, shipping information) — not personal data in most cases. Where personal data appears (individual names in commercial invoices), data minimisation applied — names processed for declaration purposes only, not stored in training data. Transparency: all AI agent decisions logged with complete reasoning — available for Dutch Customs audit. GDPR Article 22: customs declarations are regulatory submissions, not decisions with legal effect on individuals — Article 22 does not apply directly. However, we implemented human oversight as a best practice.

Customs performance: declaration processing time from 45 minutes (manual) to 8 minutes (AI-generated, human-reviewed) per declaration — 82% time reduction. HS classification accuracy: 94.7% first-pass (AI) versus 95.8% (experienced human declarants) — the gap closing as the AI learns from declarant corrections. After 6 months of production operation with continuous learning: AI accuracy reached 96.1% — matching human declarant accuracy. Error rate: from 4.2% to 0.9% (79% reduction). The AI agent catches errors that human declarants miss under time pressure — particularly value calculation errors and missing FTA preference claims. Penalties: Dutch Customs penalties from EUR 680K/year to EUR 85K/year (87% reduction). Processing capacity: the 22-person customs team now processes 28,000+ declarations per month (previously at capacity) with capacity for 45,000+ declarations. The company is accepting new business that it previously turned away. Revenue impact: new customs brokerage clients generating EUR 3.2M annual revenue — business that was previously declined due to capacity constraints. Declarant role evolution: customs declarants transitioned to "customs intelligence analysts." New responsibilities: reviewing AI escalations (complex classifications requiring professional judgment), trade compliance strategy (advising clients on FTA optimization, duty minimization, and supply chain structuring), BTI applications (applying for Binding Tariff Information for complex or high-value products), and audit preparation (preparing for Dutch Customs audits with AI-generated documentation). Declarant satisfaction: team satisfaction survey showed 82% preferred the AI-augmented workflow — "I spend my time on interesting classification problems, not routine data entry." OR: zero objections during 12 months of operation. Quarterly OR reviews confirmed compliance with agreed provisions. The OR chairman: "This is a model for how technology should be introduced — the team is doing more valuable work, not less work." FTA optimisation: the origin determination agent identified EUR 2.1M in unclaimed FTA preferences that the manual process had missed (primarily EU-UK TCA and EU-Japan EPA preferences where the company's clients were paying full duties unnecessarily). These recovered preferences were a significant value-add for the company's clients — strengthening client relationships. Cost: customs department efficiency: EUR 1.4M annual savings (processing capacity without additional hiring). Penalty reduction: EUR 595K annual savings. New client revenue: EUR 3.2M annually. FTA preference recovery for clients: EUR 2.1M (generating goodwill and client retention). Development cost: EUR 205,000. Payback period: 5 weeks.