Autorité Douanière de Sharjah. Système multi-agents IA pour l'automatisation complète du rapprochement douanier
Nous avons conçu un système multi-agents basé sur l'IA pour automatiser entièrement le rapprochement des documents douaniers pour une opération logistique majeure aux Émirats arabes unis. Le système a traité 30 000 documents TIFF, nettoyé les artefacts, extrait des données structurées et s'est intégré directement au système de gestion d'entrepôt (WMS). Cette transformation a résolu un écart de 4 millions de dollars qui persistait depuis longtemps, éliminé les goulots d'étranglement de la révision manuelle et réduit le temps de traitement des documents de plusieurs heures à quelques minutes.
01 —The Impact
The results, up front.
Processing time dropped from 3–4 hours per document to 3–4 minutes . $1.5M in outstanding debt was recovered through accurate reconciliation. Annual OPEX savings of $200k–$300k were achieved. Full automation eliminated manual workload and human error. The customs department now operates with real-time visibility, accuracy, and compliance.
02 — The Challenge
Where they started.
The logistics operator’s customs department relied heavily on manual document review, with each TIFF file requiring between 3 to 4 hours of processing. The volume made it impossible to catch discrepancies in real time, leading to operational delays and a $4M accounting mismatch. Our goal was to build a resilient, compliant, and fully automated reconciliation pipeline using multi-agent AI. We began by analyzing file structures, defect types, and variations in handwriting to determine the optimal OCR and cleaning approach. From there, we designed a multi-agent system where each agent independently handled specific responsibilities — image cleaning, artifact detection, OCR extraction, structured data mapping, validation, and WMS submission. The pipeline allowed agents to interact, validate each other’s outputs, and escalate exceptions through a human-in-the-loop (HITL) interface. This ensured accuracy while maintaining full auditability and compliance. Once validated, the system integrated directly with the WMS, synchronizing reconciled data automatically and maintaining a transparent digital trail.
03 —What We Did
The work.
Document Audit & Defect Mapping
We catalogued all 30,000 TIFF files, identifying artifact types, handwriting variations, and structural inconsistencies to define the precise cleaning and extraction requirements before any automation was built.
AI-Powered Document Cleaning
Dedicated agents removed ink artifacts, pen strokes, smudges, and duplicate pages, then normalized each file into a consistent OCR-ready format using AI-based layout detection for fragmented documents.
Multi-Pass OCR & Structured Extraction
A multi-pass OCR layer handled both handwritten and typed fields, assigning confidence scores at the field level and automatically mapping extracted data into standardized customs templates.
Agent-to-Agent Validation & Exception Handling
Agents cross-validated each other's outputs in real time; low-confidence or mismatched entries were escalated to a human-in-the-loop dashboard, maintaining auditability without stalling the pipeline.
Reconciliation & WMS Integration
Reconciled records were automatically compared against financial and inventory data to isolate the sources of the $4M discrepancy, then submitted directly into the WMS with a full digital audit trail.
04 — Approach & Methodology
How we got there.
Throughout the project, we evaluated extraction accuracy, error patterns, and mismatch categories to refine the pipeline. The multi-agent architecture allowed parallelization and redundancy, ensuring error handling without blocking the entire workflow. This approach revealed that the majority of previous discrepancies stemmed from manual keying errors, missing fields, and unreadable pages — all eliminated in the automated pipeline.
By shifting from manual to AI-based reconciliation, the client gained full visibility into customs flows, leading to significant financial recovery and operational stability.
05 —In Practice
Project samples.
"We had lived with this discrepancy for years because the manual process simply couldn't keep pace with volume. Within weeks of the system going live, we had recovered $1.5M and could finally see our customs flows in real time. The accuracy and speed are beyond what we thought was possible with documents in this condition."
Keep exploring Renascence
Your turn
Ready to write your own success story?
Book a discovery call and see what behavioral CX can do for your business.