Enterprise AI for Fortune 100 Food Processing Company
Challenges
A multinational food company produces and distributes a broad portfolio of primarily protein products and brands across retail and food service channels. Historically, the company conducted demand forecasting and planning through manual spreadsheets. Planners relied on tribal knowledge to prioritize the SKUs that required attention and cited the whole process as cumbersome, error-prone, non-performant, and resulting in low forecast accuracy.
The company deployed a market-leading planning solution to address these pain points. This solution created consensus demand plans with shorter 8-week forecast horizons. However, the new demand forecasts failed to meet the accuracy goals and resulted in large forecast errors. The planners still overwrote the resulting forecasts, not trusting the generated forecasts. Despite deploying market-leading software, the company still struggled with demand planning and forecasting:
- The deployed models were ineffective due to not forecasting at the right granularity level. The one-size-fits-all approach did not generate sufficiently accurate results.
- A multitude of signals drives demand forecasting. The solution did not effectively incorporate data sources with those signals outside of ERP, leaving significant possible accuracy lift on the table.
- The generated forecasts lacked explainability, leading to user distrust and low adoption.
- The solution did not track forecast analytics over time, making it hard to understand historical model performance.
- The solution struggled to scale to over 80k+ forecasting subjects the company required.
- Due to limited historical data, 40% of the trial SKUs are challenging to forecast.
Approach
C3 AI initially configured the C3 AI Demand Forecasting application for 300 SKUs and 10 distribution centers within two business divisions. This business segment averaged $3B+ in annual sales.
C3 AI Demand Forecasting improved forecast accuracy by 12% overall and demonstrated a 15% accuracy increase for the SKUs with low forecastability. After proving the value, C3 AI is scaling the solution to the entire company, encompassing ~80k forecasting subjects.
To achieve these results, the C3 AI team leveraged the Supply Chain Digital Twin data model to unify and ingest internal and external data, providing the team with an on-demand view of historical and incoming near real-time data.
The team conducted a forecast profile analysis to categorize SKUs by assessing variables influencing forecast accuracy (e.g., volatility, seasonality, intermittency). The team then used AI-based clustering to segment the SKUs for a tailored forecasting approach. Each segment and each forecast have unique drivers explained in the user-friendly evidence package.
C3 AI configured the ML models at multiple forecasting levels for SKUs and SKU-distribution center levels to mirror the existing planning processes. The team further fine-tuned the models by leveraging native C3 AI capabilities like hyperparameter optimization.
The generated forecasts mirrored the business’ extant planning processes, generating weekly forecasts for an 8-week horizon.
The joint team configured a workflow-enabled user interface empowering planners to identify forecasts with potential issues through data-driven exception management rapidly. Planners can now investigate forecasts with AI explainability and analytics and train and promote new models with scenario planning capabilities.
About the Company
- $50+ billion annual revenue in 2022
- 8+ brands
- 240 facilities worldwide
- 140,000+ employees
Project Objectives
- Integrate and unify data from 5 ERPs and 7 external data sources
- Configure high-accuracy, scalable AI models to generate weekly forecasts by SKU and distribution center for an 8-week horizon
- Configure AI-based demand segmentation to scale AI models across products with minimal effort
- Configure workflow enabled application to improve planner efficiency with exception management and AI explainability
Project Highlights
- Integrated 3 years of historical data from 5 ERP and 7 external data sources, amounting to 10 million rows.
- Configured and tested 2,000+ high-accuracy AI forecasting models at multiple granularity levels for products with different explainability profiles.
- Generated a 12% uplift overall compared to the consensus forecast.
- Generated a 15% uplift for the low-forecastability SKUs.
- Configured workflow-enabled user interface with exception-based alerting to support the users in identifying and investigating potential forecast issues.
- Configured AI explainability to explain AI forecast drivers.
Results
Solution Architecture
Benefits
- Improve short-horizon forecast accuracy by 12% over baseline for all SKUs.
- Automatically track forecasts with potential issues with data-driven exception management and alerting within the application.
- Empower demand planners to understand AI-generated forecasts with explainability.
- Provide a data foundation that enables rapid design, development, and deployment for a suite of supply chain AI applications (e.g., Inventory Optimization, Production Schedule Optimization) to further optimize customer’s supply chains.