How Mondelez Scales 3,000 Production AI Models Across Sales and Supply Chain with Databricks

Mondelez International makes snacks that appear in more than 160 countries, with brands like Oreo, Ritz, Toblerone, Cadbury, and Sour Patch Kids generating revenue decisions that play out at millions of retail touchpoints every day. The company set a bold ambition: to become a native AI company by 2030. To get there, it needed to fix a fragmented data foundation.

Before Databricks, each business unit worked in isolation. Data scientists built models in local environments like Anaconda, with no standardized platform, no shared governance, and no way to monitor models at scale. When teams wanted to scale a training run to multiple worker nodes, they struggled to execute the full end-to-end pipeline. Security and compliance were concerns, particularly with open source libraries. Data silos meant duplicated effort and inconsistent results across regions. “When we wanted to scale an application, that required several worker nodes. We found it challenging to run the entire end-to-end pipeline in one go,” said Sachin Kulshrestha, Global ML Engineering and Ops Lead.

Mondelez deployed the Databricks Data Intelligence Platform on Google Cloud to centralize every part of the AI lifecycle. Databricks Notebooks became the collaboration layer for data scientists, engineers, and analysts. MLflow and MLflow Model Registry gave teams a unified way to track, version, and govern all models—batch and real-time. Unity Catalog enforced fine-grained access control and data lineage. Databricks Agent Bricks is now powering SnackGPT, a domain-specific generative AI assistant for brand and category teams to ask natural language questions like “What were our top-selling products last year?” or “Why did sales drop in Q1?” using real syndicated point-of-sales data.

Across sales execution, SKU recommendation models evaluate store potential using geography, demographics, and layout to recommend smarter product placement—delivering a 2–4% increase in store-level topline sales and improving sales rep effectiveness to around 80%. Demand forecasting models now support supply chain planning from six weeks to two years out, reducing finished goods inventory waste by 2–3% and improving forecast accuracy by 3–5%. Revenue growth management models run simulations before any price change, modeling whether customers would switch brands, change pack sizes, or reduce volume. “If a business was able to sell one additional product because of a model recommendation, we can directly measure that uplift,” Kulshrestha said. “It’s a revenue growth story.”

Today, Mondelez manages close to 20,000 models on Databricks, with around 3,000 in production—giving leadership a clear view of which models are production-grade, experimental, or retired. Technical debt from managing disparate systems has dropped significantly, and Delta Sharing is in roadmap to make reusable segmentation models available across regional teams. “Databricks has been so easy for the team to operate, empowering us to meet the needs of the business,” Kulshrestha concluded.