From the Gold layer to a federated knowledge graph: Decathlon's journey

Discover how we structured our analytical foundations in four levels (Gold and Mart, Metric, Semantic, Context) to adapt our data architecture to agents, distribute responsibilities in a metrics mesh, and give tens of thousands of business users access to insights.

In 2025, the decentralisation of our analytics teams led us to a classic dead end: KPI chaos. Yet the technology base was healthy. Instead of looking for an isolated silver bullet for AI, we looked at what worked in the industry, Minerva at Airbnb and at Spotify, in order to build on what we had. That foundation rests on an exhaustive Gold layer made of our Core Data Products, which centralise all of the company's activities. Those products feed our Core Data Mart, built on a Kimball matrix, which flattens and aligns the data on common dimensions.

Then generative AI and prompt interfaces through Databricks Genie arrived, multiplying query volume. We had to learn, sometimes the hard way, that you cannot delegate everything to the LLM: it is a wonderful comprehension interface, but it can lead to high compute costs.

In this talk we present how we evolved our analytical base:
- Organising by taking inspiration from others: adapting the Chapter Franchise model to technologically align more than fifteen business domains and twenty country teams on the use of our Core Data Products, and the frictions overcome.
- The metric layer, the compute engine: implementation with Databricks Metric Views on top of our Core Data Mart. This is where SQL calculation rules and our strict pre-aggregation strategy are defined. Instead of scanning five hundred million rows in thirty seconds, our agents answer in one second over fifteen hundred rows, so costs go down even accounting for the rebound effect of AI.
- The semantic layer, the business translator: introducing rich business language and descriptions through YAML and Collibra on top of the computation. This is the layer that explains to the LLM what a business indicator means, avoiding hallucinations.
- Thinking ahead, context layer and knowledge graph: having a hundred and thirty metrics consumed by our executive committee is not enough. To make AI reliable we take inspiration from statistical process control to model causal KPI trees.

What the audience will learn: field feedback on organisation, the real difficulties and solutions in making a federated data model work at very large scale under transverse leadership; Decathlon's complete architecture, structured around our four foundation layers for generative AI; and a technical vision that sharpens progressively, with the conceptual basis for building metric trees and a context layer that AI agents can understand.