Demand Forecasting and AP Automation at a Global FMCG — 18% Inventory Reduction Across 8,000 SKUs
Demand Forecaster across eight thousand SKUs in thirty-one markets, paired with AP Automation across the SAP estate — supply chain and finance in lockstep.
The challenge
The client — a global FMCG group with a portfolio spanning personal care, household and packaged-food categories, operating in 31 markets — was carrying inventory at levels their CFO and COO both considered structurally unjustified. Working capital tied up in raw-material, work-in-progress and finished-goods inventory was the largest single component of the group's balance sheet outside of intangible brand assets.
The root cause was forecasting. The group's existing demand planning relied on a combination of legacy SAP APO (recently end-of-life) and a network of regional Excel-based forecasting workbooks maintained by category managers in each market. Forecast accuracy varied wildly — at the group level the rolling mean absolute percentage error (MAPE) was 18%, but at the SKU-by-market level it could be 35% or worse for new launches and seasonal lines.
Layered onto the inventory problem was a finance problem. The group's accounts payable function, also based on SAP, was processing roughly 2.4 million supplier invoices per year across 22 shared service centres. Invoice turnaround time averaged 14 working days, supplier dispute rates were high, and the early-payment discount capture rate was below 20% of theoretically available discounts. The CFO had set a target of getting invoice TAT under 48 hours.
The group had been quoted multi-year transformation programmes by two major consulting firms — one for demand planning and one for finance transformation — with a combined timeline of 36 months before measurable outcomes. The CFO and COO had jointly asked: is there a faster path?
The approach
We led with two accelerators on parallel tracks — Demand Forecaster (Df) for the supply-chain problem and AP Automation (Ap) for the finance problem. Both ran inside the same delivery programme to ensure the SAP integration architecture was consistent.
On the demand side, we built a unified forecasting platform across all 8,000 SKUs in all 31 markets. The platform combines multiple forecast models — classical statistical methods (ARIMA, ETS) for stable mature lines, gradient-boosted tree ensembles for promotional uplift, and a deep-learning model (a fine-tuned Temporal Fusion Transformer) for the harder-to-model categories with strong seasonal, weather and macroeconomic sensitivity. The platform ensemble-weights the models per SKU-market combination, with weights re-estimated weekly.
Critically, the platform incorporates exogenous signals that the previous Excel workbooks did not — weather data for ice-cream and beverage categories, retailer scan-data for fast-moving lines, social-media trend signals for the personal-care portfolio, and macroeconomic indicators (inflation, FX) for the markets with high macro volatility.
On the finance side, AP Automation handles invoice ingestion (email, supplier portal, EDI), document classification (DocuMage at the OCR layer), three-way match against purchase orders and goods receipts, tolerance-based exception handling, and approval routing. The pipeline is integrated into SAP S/4HANA via SAP's standard inbound IDoc and BAPI interfaces, with all postings happening natively in SAP.
The pre-built building blocks
Rather than commission a ground-up build, the engagement leaned on MindMap's pre-built accelerator library — production-tested components that compress what would otherwise be a six-to-nine-month build into weeks.
Demand Forecaster
Multi-model SKU-market demand forecasting
Inventory Optimizer
Multi-echelon inventory rebalancing recommendations
AP Automation
End-to-end AP — ingestion, match, exception, posting
DocuMage
Invoice OCR and line-item extraction across 22 languages
The architecture
The demand-forecasting platform runs on the group's existing Azure tenant, with a Databricks lakehouse as the data backbone. Data ingestion runs hourly from SAP ECC (historical sales, master data, current inventory positions), from the group's regional retailer scan-data feeds (where available), and from external data providers (weather, FX, social trends). The data lake is dbt-modelled into a forecasting-ready feature store.
Model training runs nightly on a per-cluster basis (clusters of SKUs grouped by category, market and seasonality profile) using Databricks ML. Inference runs on a 12-hour cycle and writes forecast positions back into SAP via the same integration adapter. The platform produces a 26-week forecast horizon, refreshed twice daily, with confidence intervals at each forecast point.
The AP Automation pipeline runs as a Kubernetes-native service in the same Azure tenant. Invoice ingestion happens from multiple sources — a dedicated email mailbox (approximately 60% of volume), the supplier portal (25%), EDI from the group's largest suppliers (12%), and PDF uploads from regional finance teams (3%). DocuMage performs OCR and field extraction, including line-item-level extraction with PO matching. The platform handles invoices in 22 languages and 14 currencies.
Three-way matching is performed by a deterministic rules layer that calls into SAP via OData and IDoc interfaces to pull the relevant PO, goods receipt and master data. Tolerance-based exceptions are routed to the regional shared-service-centre teams via a custom exception-management UI. Approved invoices are posted to SAP automatically.
Both platforms share a unified observability layer (Datadog) and a unified data lineage tool (Collibra) that the group's data-governance team uses to track end-to-end data flow from source to forecast and from invoice to GL posting.
The numbers behind the story
On the supply-chain side, group-level forecast MAPE has improved from 18% to 12% on a rolling thirteen-week measurement. Total inventory days have reduced by 18%, with the working-capital release directed to the group's accelerated debt-reduction programme. Service level (on-shelf availability at the retailer) has improved by 1.4 percentage points despite the lower inventory — counterintuitive at first glance, but reflecting the move from buffer-driven stocking to forecast-driven stocking.
Specific category outcomes: the ice-cream and beverage categories — where the weather data integration was most impactful — have seen MAPE improvements of 31% and 22% respectively, and inventory reductions of 24% and 19%. The personal-care portfolio has seen smaller MAPE improvement but meaningful waste reduction (fewer obsolete-stock write-offs at the end of seasonal cycles).
On the finance side, average invoice TAT has dropped from 14 working days to 36 hours. Straight-through processing (invoices that post to GL without human touch) is at 67%. Early-payment discount capture has risen from 19% to 71% of available discounts — a meaningful cash impact. Supplier-dispute rate has dropped by half, primarily because the matching and exception handling now catches discrepancies during ingestion rather than after posting.
The 22 shared-service-centre AP teams have been restructured. Headcount has not been reduced but capacity has been redirected: from invoice processing to supplier-relationship management, master-data quality programmes and cash-flow analytics. The CFO reports that the AP function has shifted from being a transaction processor to being a strategic finance partner.
“We had been quoted 36-month transformation programmes by the big consultancies. MindMap was live in the first market in twelve weeks and across the group in eight months. Eighteen per cent less inventory, 36-hour invoice TAT, and our category managers are spending their time on category strategy instead of on Excel forecasting workbooks.”— Group Chief Operating Officer· Global FMCG Group
Why MindMap was chosen
The CFO and COO were unwilling to commission a 36-month transformation programme. They had been burned by previous multi-year ERP-adjacent programmes that delivered slides for the first 18 months and limited business impact thereafter.
MindMap proposed a 24-week initial delivery — Demand Forecaster live in one pilot market by week 12, AP Automation live in one shared-service centre by week 14, then progressive rollout. The pre-built accelerators meant we were not building forecasting algorithms from scratch; we were tuning a proven platform to the group's data.
We brought specific SAP integration expertise — including SAP-certified integration architects on the delivery team — which materially de-risked the SAP integration work that had derailed previous programmes. Our willingness to operate inside the group's existing Azure and Databricks environment, rather than requiring a new platform, meant procurement was straightforward.
Critically, our pricing model was outcome-linked: a fixed-fee delivery component plus a success fee tied to demonstrated inventory and TAT outcomes. The competing consulting firm proposals were structured around time-and-materials with no outcome accountability.
Related deployments
Predictive Maintenance
Production Line Monitor + Anomaly Detector cut unplanned downtime by 38% across 14 plants — through predictive maintenance on the critical asset base.
Quality Control AI
Inline computer-vision QC on 14 production lines — 96% defect-detection accuracy, with downstream rework reduced 71%.
Global Supply-Chain Optimisation
An integrated supply-chain platform compressed global order-to-delivery by 11 days — through demand-forecasting, logistics-automation and predictive ETA.
Want an outcome like this?
Start with a 2-week AI Readiness Sprint. We deliver a prioritised use-case backlog and business case grounded in what's actually buildable with our accelerator library.