Augmented Intelligence in Demand Planning

The Challenge:

A $3B Swiss manufacturer of specialty chemicals that had been assembled through acquisitions was experiencing challenges with demand forecast accuracy. Macro demand across products was averaging 44% accuracy and causing unsustainable growth of inventory volume and value.

The key questions to answer:

• How forecastable is the business?
  
• Should there be separate forecast models customized to product line?
  
• What are the root causes for poor demand signals (e.g., why does Sales – who has direct access to the customers – historically have a 4% accuracy rate)
  
• Is 65% accuracy achievable?
  
• What does “Good Forecast Look Like” and how can we get there? ​

Approach:

The team shifted the demand planning paradigm from a rigid SAP IBP 'black box’ features to flexible, transparent in-house AI forecasting: 'Augmented Intelligence' enabling smarter decisions. We developed and deployed a ready-for-use Demand Planning Excel-Python-AI Forecast Engine, empowering a foundational knowledgebase for continuous improvement.

A series of eight, one-week sprints:

Sprints Weeks 0–2: Discovery & Maturity​

• Discovery of data exclusions “Golden View” single source of truth definition
  
• ABC/XYZ maturity model

Sprint Week 3: Excel based ABC/XYZ Categorical Linear Regression

• Marginal increase of predicted Forecast Accuracy to 50’s%
  

Sprint Week 4: Proof of concept: AI Ensemble MPVv1

• Promising 77% predicted forecast accuracy on limited "Rapid Fail" 1% SKU sample

Sprint Week 5: Successful scaled to all SKU’s MPVv2.0

• Full 1800 SKUs deployed in time-series forecast methods
  
• At comparable parity @ 47% predicted forecast accuracy with existing SAP IBP outcomes

Sprint Week 6: Fast Sprints MVPv3-v12

• Over a dozen MVP sprint runs / rollbacks iteration
  
• ABC/XYZ Categorical Z-score weighing
  
• +3 Models ensemble forecast and predictive forecast Accuracy optimization SKU
  
• Increased predictive forecast accuracy to mid-60’s% to 70’s%

Sprint Week 7: Acceptance criteria @ +64% predicted forecast accuracy

• Module Stage 2 Prescriptive module and Stage 1 descriptive python refinement

Close-Out: Went live with Release V1

Result:

Proved through application of machine learning algorithms that the SKUs across the six disparate business units are, in fact, forecastable. Model resulted in over 3,000 basis point improvement in forecast accuracy (over 75% accuracy, up from 45%). The key driver of success was the pivot at Sprint Week 4 away from an Excel-based model to an ‘AI’ Agentic Workflow and Ensemble forecast modeling to unlock next-level accuracy through ‘Augmented Intelligence’.