11 / Manufacturing
SAP-Integrated AI Demand Forecasting Platform
We gave a manufacturer one trustworthy data estate across 3 companies and SKU-level forecasts, where the AI only picks which reviewed number to run and never writes back to SAP.
- Client
- Undisclosed (Under NDA)
- Stack & approach
- ManufacturingSAP S/4HANADemand ForecastingSemantic LayerRead-Only
The problem
Commercial operations ran on SAP S/4HANA across 3 entities with a layer of spreadsheets refreshed a day behind, so a receivable was often already paid before anyone acted. A single month of one register ran past 100,000 invoice lines, and a trailing 3-month rolling forecast left hundreds of SKUs in shortfall.
Solution & architecture
One governed data estate, mapped once and reviewed, with accuracy coming from a deterministic compiler rather than model behaviour. A semantic layer maps cryptic SAP fields to reviewed business meaning, so at question time the model only chooses which reviewed metric to run. Forecasts are per-SKU and measured against the client's own plan. Nothing is ever written back to SAP.
Why it is built this way
Read-only to SAP
The model never writes, never sees raw rows
Deterministic accuracy
Metrics compile to exact SQL, not guesses
Scoped by role
Data up to one picture, access down by role
