Client Context

Our client, a global consumer goods leader operating in 130 countries, faced a critical inefficiency in its Romanian direct-to-retail (D2R) distribution network. 

The operation was complex: 

• Scale: 30+ sales offices distributing over 40 SKUs. 
• Model: ~15 agents per office would procure inventory, service 30–35 retailers daily, and return all unsold stock to their warehouse at the end of the day. 

The core problem was unpredictable demand. Without an accurate forecasting mechanism, the business was trapped in a costly cycle of two significant inefficiencies: 

1. Stock-Outs: When demand was underestimated, agents ran out of product prematurely. This directly impacted sales. 
2. Excess Returns: When demand was overestimated, agents returned large volumes of surplus stock to the warehouse, leading to high handling costs and inflated logistics expenses. 

The client sought an effective demand forecasting solution to break this cycle, reduce unnecessary stock movement, and optimize both labour and logistics costs. 

The InXiteOut Approach

Our engagement began with a comprehensive analysis of the client’s sales and distribution ecosystem. We worked in-depth with sales teams to deconstruct the business model and identify the key, demand-impacting variables — from historical sales patterns and promotions to seasonal trends, holidays, and external factors. This analysis became the blueprint for our sophisticated AI forecasting framework. 

Two-stage AI forecasting framework 

To achieve the high-precision, SKU-level forecasts the client needed, we developed a sophisticated two-stage AI framework. 

Traditional, single-step forecasts are often imprecise because they fail to distinguish between two separate business questions: 1. Will a retailer buy? and 2. How much will they buy? 

Our solution was to deconstruct the problem. The framework first predicts the purchase intent, and only then does a second model calculate the purchase quantity. This two-step method of separating intent from quantity dramatically improves forecast accuracy. 

The framework included: 

1. Purchase Intent Classifier: The first model analyzed historical patterns, promotions, and seasonal trends to predict the probability of a retailer purchasing a specific SKU on a given visit. 
2. Purchase Quantity Regressor: Once the first model confirmed a likely purchase, the second model estimated the precise quantity of each SKU. It leveraged 30+ engineered features, including outlet profiles, past purchase trends, and geographic data, to refine its forecast. 

This robust, AI-driven system was deployed to generate daily demand forecasts for over 40 SKUs across more than 15,000 retail locations.  

Technology stack used 

• Azure ETL and ML Ecosystem (Data Lake, Data Factory, Databricks)    
• Power BI 

Benefits Delivered

InXiteOut’s AI/ML-powered solution delivered immediate, measurable improvements across the entire D2R network:

• Maximized Sales Revenue: Stock shortages were virtually eliminated, maintaining optimal levels of less than 1%. This ensured products were consistently available at the point of sale, capturing previously lost revenue. 
• Slashed Operational Costs: The accurate forecasts led to a ~25% reduction in unnecessary product loading and unloading. This operational streamlining directly resulted in a ~20% reduction in warehouse manpower costs. 

Author

FullName

Role and Company