Intratuin - Forecasting demand with an AI model

The challenge

Every retailer recognises the question: 'How can we better predict customer demand so buyers can order smarter?' Shelves need to stay stocked, but traditional forecasting methods fall short given the complexity and scale of modern retail. AI offers opportunities, provided the foundation is solid. That meant working with clean, complete and consistent data, and the right variables to feed the model.

The model

We trained an AI forecasting model in Google Vertex AI using four years of historical data. By combining variables such as price changes, promotions, seasonal patterns and holiday peaks, the model got a more realistic picture of buying behaviour. After training, we integrated the model into the OutSystems-based purchasing environment. Within a short time, more than 48,000 articles could be forecast, creating a solid foundation for predictive insight in the daily purchasing process.

From forecast to advice

A forecast alone isn't enough. The next step was translating forecast numbers into concrete, actionable advice. So we developed a simple safety stock threshold formula that indicates when stock levels become critical. Buyers immediately saw a clear advisory figure in their existing purchasing screen. No extra dashboards. No new process. Just direct support at the moment of decision.

Adoption and collaboration

For several months, we worked closely together with buyers in the stores. We observed how they worked, tested how well the advice fit their routines, and discussed moments where they consciously deviated from the AI's advice. These conversations were open, sometimes critical, but always valuable. They not only helped improve the model, but also made buyers feel like co-owners of the solution. Their feedback enriched both the model and our understanding of the purchasing process and the business.

Results and next steps

The pilot delivered two key insights. First: context is crucial. Factors such as shelf layout, visual merchandising and pre-selected product groups strongly influence decisions and need to be factored in. Second, we developed a comparison framework that shows, per article, where AI advice and human decisions align or diverge. This makes iteration easier and more transparent, and helps drive targeted improvement towards the right combination of accurate forecast and the right action.

Building further

Intratuin can now continue this development independently, which was exactly the intention. AI solutions are rarely perfect the first time. By starting small, learning and iterating, value grows step by step. What started as a pilot has grown into a continuous learning process that helps Intratuin make increasingly well-founded, data-driven purchasing decisions.