FMCG
Agentic AI for FMCG — demand forecasting, trade promotion design and supply chain automation, deployed in production for European consumer goods businesses.
What makes FMCG hard for AI
The data does not agree with itself
Shipment, point-of-sale and promotional calendars sit in different systems with different product and customer hierarchies. Reconciling them is where most of the effort goes, and skipping it produces confident recommendations built on mismatched keys.
Promotions interact
Lift on one SKU is frequently cannibalisation from another. Any model that evaluates promotions in isolation will overstate their value.
Planning cycles batch decisions
Quarterly review rhythms exist because analysis is expensive. The cycle time is a symptom of the architecture, not of the calendar.
Where the value concentrates
- Trade promotion design. Continuous baseline and lift estimation, with cannibalisation modelled across the portfolio.
- Demand forecasting. Incorporating promotions, seasonality and external signals.
- Replenishment. Agents that trigger and time orders against real supply constraints.
- Revenue growth management. Pricing and mix decisions grounded in measured elasticity.
What the economics look like
Trade spend is typically among the largest lines on a consumer goods P&L, and a meaningful share of it is routinely acknowledged to be ineffective. Even modest improvements in promotional targeting move a large absolute number — which is why measurement, not optimisation, is the right first project.
How we build it
We fix hierarchy and reconciliation first, agree baseline methodology with finance, then add scenario generation. The category manager keeps the decision; the arithmetic stops being manual.
Related work
Starting smaller
Smaller brands generally start with post-event promotional evaluation. It requires no forecasting and immediately reveals which mechanics have been losing money.
Questions we get from consumer goods teams
Our shipment and point-of-sale data do not reconcile. Is that a blocker?
It is the project. Reconciliation is where most of the effort and most of the value sit. Optimising on unreconciled data produces confident recommendations built on mismatched keys, which is worse than not modelling at all.
How do we handle cannibalisation?
Model it explicitly across the portfolio rather than evaluating each SKU in isolation. A promotion that lifts one product by taking volume from another at a worse margin is a loss that per-SKU reporting will record as a win.
Will this replace our trade promotion management system?
No. The TPM system stays the record for commitments and settlements. The agentic layer handles reconciliation, evaluation and scenario generation against it.
How much history do we need?
Two to three years of promotional history with reliable point-of-sale data is comfortable. What matters is the number of distinct promotional events, not elapsed time — frequently promoted categories reach usable coverage faster.