One post tagged with "EUDR"

The Challenge of EUDR Compliance for Product Catalogs​

The EU Regulation on Deforestation-free Products (EUDR) represents a significant compliance challenge for retailers and manufacturers. This regulation, which came into force on June 29, 2023, and becomes enforceable on December 30, 2025, aims to minimize the EU's contribution to global deforestation and forest degradation. It prohibits placing products linked to deforestation on the EU market and requires operators to implement robust due diligence systems. The regulation covers cattle, cocoa, coffee, oil palm, rubber, soya, and wood, along with derived products like leather, chocolate, and furniture. For retailers with extensive product catalogs containing thousands or even millions of items, manually classifying which products fall under EUDR jurisdiction is an overwhelming task.

The Need for an Automated Solution​

We helped one of our clients, a major Dutch online retailer with over 15 million products in their catalog, with this exact problem. They needed to identify which products in their extensive catalog needed to be EUDR compliant. EUDR compliance is determined by specific Harmonized System (HS) codes that classify products containing the regulated commodities, making accurate product classification essential for compliance.

They needed a solution that could:

• Automatically analyze product descriptions and specifications
• Accurately identify the HS-code for each product in their catalog
• Scale to handle their entire product catalog
• Generate accurate LLM token cost estimates
• Provide consistent classification results

Building an AI Solution with Aigentic: The POC Approach​

To address this challenge, we developed a solution using our Aigentic library. Aigentic provided the foundation for this POC, offering a powerful Kotlin Multiplatform DSL for building and integrating AI agents tailored to specific business needs.

For this POC, our approach involved creating a specialized classification agent that could analyze product data and determine the HS-code with high accuracy. The solution leveraged several key components of the Aigentic ecosystem to demonstrate the potential of this approach.

The Data Validation Challenge in the POC​

One of the most significant challenges in the POC was validating the data and ensuring the accuracy of the AI classifications. This is where Aigentic proved particularly valuable for our solution. By manually validating the data, we were able to identify and correct errors that were introduced during the classification process. Aigentic allowed us to review and correct the output of the AI agent, providing a human-in-the-loop validation process that was essential for the POC.

We implemented a multi-stage validation process:

• Initial exploration: We started with a small set of products to explore feasibility
• Iterative Labeling: Using Aigentic, we created an interface through Excel to review and correct AI classifications
• Golden Dataset Creation: Through this process, we gradually built a "golden dataset" of correctly classified products
• Continuous Improvement: The agent was continuously adapted based on feedback, improving its accuracy over time

Aigentic's ability to combine human-in-the-loop validation with AI automation was crucial for building trust in the system. During the solution, stakeholders could review classifications, provide feedback, and see how the system improved over time, which was essential for validating the approach.

Operational Observability: Monitoring Performance and Costs in the POC​

A critical aspect of any AI-powered solution is understanding its operational characteristics. Aigentic provided comprehensive observability features that were essential for monitoring and optimizing our EUDR classification POC. Through Aigentic Platform's built-in monitoring, we gained visibility into key metrics:

• Cost Tracking: Allowing us to provide accurate cost projections for scaling to the full product catalog
• Latency Monitoring: Ensuring the solution could meet real-world requirements
• Run Logging: Capturing every classification decision and reasoning, creating a complete trail essential for quality assurance

This observability enabled us to fine-tune performance and build confidence in the solution's reliability—critical factors for successful implementation.

Results​

The solution demonstrated promising results. The agent achieved 98% accuracy on the golden dataset. Statistical analysis with a 95% confidence interval suggests that when applied to the complete product catalog, the real accuracy would likely fall between 96.77% and 99.23%, indicating robust performance suitable for large-scale EUDR compliance operations.

Looking Forward: From POC to Complete EUDR Classification​

With EUDR compliance deadlines approaching, organizations need to act quickly to prepare their due diligence systems. The POC has demonstrated that our approach can be developed into a solution for ongoing compliance that helps companies meet these critical deadlines.

As regulations evolve and product catalogs grow, the POC suggests that AI-powered classification using Aigentic could offer a scalable, accurate, and efficient solution. The POC demonstrated how the combination of powerful language models with human expertise can create a system that continuously improves while maintaining the critical human oversight necessary for regulatory compliance.