Transforming Vaccine Supply Chain with Accurate Demand Forecasting

Challenges

Given the intricate nature of vaccine supply chains, an accurate demand forecast is imperative for efficiently planning complex production, inventory, and logistics processes. However, the firm’s statistical forecasting approach fell short of this requirement. The firm had to manually adjust its statistical forecasts in a time-consuming manner, with teams of demand planners modifying the forecasts using their expertise. Forecasting process is also differently performed across markets. Despite the complicated and inconsistent process, the accuracy increases from manually adjusted forecasting were not sufficient and scalable across 16 product groups, 20 markets, and various global distributors.

Suboptimal demand forecasts resulted in higher inventory holding costs, misalignment in resource allocation and millions of vaccine doses written off annually. It also hindered the company’s ability to effectively and timely serve its patients.

The company struggled to uplift its demand forecast accuracy due to multiple reasons:

• Complex underlying demand profiles (seasonal, erratic, etc.) of vaccine products for different types of diseases.
• Obfuscated demand from multiple layers of supply and distribution channels, making it difficult to forecast “sell-in” demand.
• Current forecasting approach relying only on historical data, not external data such as wholesaler inventory and sales or market insights.
• Inability to perform forecasts at SKU-market-distribution channel level to factor in different market/distributor-specific demand behaviors.
• Vaccines and medicines are at different stages of the product lifecycle (introduction, growth, etc.), further complicating the forecasting process.

Approach

The firm partnered with C3 AI to configure and deploy C3 AI Demand Forecasting on Microsoft Azure in 6 months to improve forecast accuracy for its vaccine business.

The team started by ingesting, cleansing, and unifying 10+ data sources into one image based on the C3 AI Supply Chain Digital Twin. These data sources include internal data, such as sales orders, daily finished inventory, and customer records; and external data, such as wholesaler inventory and sales and market insights. All data sources were fully virtualized from the firm’s internal data lake.

The C3 AI team then grouped the 450+ SKU-market-distribution channel combinations into 13 segments with distinct demand profiles. The C3 AI team tested the C3 AI library of 50+ ready-to-use ML models to tailor them for each segment. They ran over 100 experiments to automatically identify the best-performing model for each segment.

Incorporating all available data, auto-segmentation, and automatic identification of best-performing models helped the customer factor in industry/market-specific insights into demand patterns, product lifecycles, and seasonality at a granular level to significantly improve forecast accuracy.

To increase trust and user adoption of the C3 AI Demand Forecasting application by the demand planners and data scientists, C3 AI configured detailed evidence packages that provide feature contributions and explains key drivers behind each AI forecast. Users can review, overwrite, and accept AI forecasts via an intuitive user interface. The accepted or overwritten forecasts are written back to the company’s existing ERPs and planning systems using bilateral connectors. This process ensures a seamless user experience. Together, C3 AI and Microsoft Azure are enabling customers to achieve operational excellence, drive AI adoption at scale, and respond to complex challenges with confidence.

About the Company

• $45 billion annual revenue in 2023
• 80,000+ employees globally
• 2 billion packs of medicines and vaccine doses delivered in 5 years

Project Objectives

• Improve forecast accuracy to reduce inventory cost and vaccine write-offs due to expiration
• Drastically reduce time required by demand planners to manually adjust statistical forecasts
• Configure the C3 AI Demand Forecasting application with AI explainability for demand planners and modelers to increase their trust in AI forecasts

Project Highlights

• 26 weeks from kick-off to production-ready application
• 10+ internal and external data sources ingested
• Best-fit AI models configured and tested for 13+ segments across 450+ SKU-market-distribution channels
• Accurate forecasting for up to 12-month time horizon
• Configured the C3 AI Demand Forecasting application user interface​

Solution Architecture