Supply Chain Forecasting
How a Manufacturing Firm Cut €2.3M in Supply Chain Waste with AI Forecasting
The Problem
A 2,400-employee manufacturing firm was losing €4.1M annually through supply chain inefficiencies. Manual demand forecasting across 14 plants relied on spreadsheets and gut feel, resulting in €2.8M in overstock costs and €1.3M in emergency procurement from stockouts. Three previous attempts at digital forecasting tools had failed — not because of the technology, but because the data wasn't ready and the teams hadn't been aligned.
The Approach
The organization used the AI Use Case Intake playbook to formally assess and prioritize demand forecasting above 23 other candidate use cases, based on data availability, business impact, and feasibility scores. This justified executive sponsorship from the COO. A 12-week data quality audit (using the AI Data Quality Audit template) revealed that 34% of historical SKU data was unusable due to inconsistent labeling — this was addressed before any modelling began. A time-series forecasting model was built on cleaned 3-year ERP data, integrated into the existing SAP environment via API. Risk was classified as Medium (using the Risk Classification Tool) — no PII, significant operational impact — triggering a governance requirement for monthly model performance reviews and a fallback manual process. The AI Change Management playbook guided a 6-week adoption program: training for plant managers, a 90-day champions program, and weekly feedback loops. Pilot ran in Plant 3 first; results validated before rollout to remaining 13 plants.
Results
Annual cost saving
Before
€0
After
€2.3M
Overstock incidents
Before
Baseline
After
-56%
Stockout incidents
Before
Baseline
After
-41%
The Outcome
Overstock costs reduced by 56% (€2.3M annual saving). Stockout incidents dropped by 41%. The model has been live in all 14 plants for 11 months with no major incidents. Monthly model reviews are conducted by an internal team — no vendor dependency created. The project also accelerated two adjacent use cases: production scheduling optimization and supplier performance prediction, both already in the intake pipeline.
Key Lessons
- 1
Data quality is the real project. Three months of data cleaning preceded any model work. Rushing to modelling without fixing data quality is the most common reason similar projects fail at scale.
- 2
Executive sponsorship changed everything. The COO's visible support resolved cross-plant coordination issues in days that would have taken months otherwise.
- 3
Start with one pilot plant. Rolling out to all 14 plants simultaneously was considered but rejected. The pilot approach caught three significant integration issues before they became organization-wide problems.
- 4
The manual fallback process was never used — but having it documented made the plant managers willing to trust the model.
Playbooks & Tools Used