Our projects
Better inventory management using AI
Optimal inventory levels are based on accurate forecasts, but predicting demand for high-value, low-volume products is complex, with traditional methods struggling to balance storage costs with the risk of missed sales.
IVADO Labs designed a sophisticated AI solution that improves the accuracy of demand forecasts and optimizes inventory in real time, while integrating human intelligence.
The AI model made it possible to maintain the same level of customer service with 20% less stock, thus reducing costs and improving profitability.
The implementation of AI by IVADO Labs began with a comprehensive overhaul of the company’s demand forecasting process. The manufacturer had been using a traditional approach that relied on spreadsheets to support team decisions. This method limited the frequency of updates and the traceability of information, restricting the organization to only high-level forecasts.
AI enabled the development of a more precise and granular forecasting model based on historical sales and inventory data. “We put several models in competition with each other to determine the best ones for each type of item,” says Philippe Fortin Simard, Product Manager at IVADO Labs.
The tool developed by IVADO Labs is also capable of generating so-called “uncensored” demand. “This involves using inventory data in conjunction with sales data to estimate both observed and unobserved demand,” explains Philippe Grand’maison, Senior Team Lead at IVADO Labs. In other words, it allows us to quantify unmet demand due to stockouts, since historical sales data only shows completed sales without reflecting all potential unmet demand. “Uncensoring demand is delicate because the evaluation can’t be based on observed quantities,” adds Mathieu Sylvestre, Data Scientist and Project Lead at IVADO Labs.
The new system offers numerous advantages. “The old manual model could only be evaluated every six months. Our AI models can be run daily, making them much more responsive,” notes Fortin Simard. This responsiveness is critical for addressing unforeseen events, such as economic shifts, natural disasters, or supply chain disruptions.
The forecasts are also more granular. In North America, forecasts that were previously grouped into four large territories are now divided into approximately sixty distinct regions, and products are segmented at a more granular level.
For long-term predictions, AI thus complements human intelligence. “Our models generate demand forecasts, but these can then be adjusted,” says Fortin Simard. “We call this assumption-driven forecasting. We start with objective data, but we have the flexibility to overlay hypotheses that integrate human expertise.”
“For example, experts may predict 2% higher sales next quarter because they have information that the model doesn’t take into account, such as the impact of a promotional campaign or a new product launch,” he illustrates. Conversely, an unfavorable economic environment might warrant a downward revision of certain forecasts.
While the method previously used by the manufacturer allowed for hypotheses, it did not allow them to be tracked. With the tools developed by IVADO Labs, these adjustments can now be evaluated continuously by comparing actual sales to the AI model’s forecasts enhanced by company experts.
The AI forecasts also include an assessment of uncertainty, which is crucial for inventory management.
The implementation of demand forecasting models was just the first step in the manufacturer’s digital transformation. One of the main objectives of the project was to convert these forecasts into concrete actions.
"A forecast only adds value to a company if it’s used to make operational decisions"
emphasizes Philippe Fortin Simard.
Optimizing inventory management was the first concrete step towards creating added value. Thanks to the precision of the models developed by IVADO Labs, the forecasts could be used directly at the dealer level.
The manufacturer was able to use the new AI model to adjust its inventories based on expected sales and calculate safety stock with extreme accuracy. “If the model predicts that a dealer will sell 10 products, but with a 20% margin of error, then the dealer will need 12 products to have enough inventory to meet potential demand,” explains Philippe Grand’maison. These two additional products are what we call safety stock.
The lower the forecast uncertainty, the less need there is for safety stock,” notes Fortin Simard. This reduction in safety stock translates into significant savings and better responsiveness to customer demand. “The more precise our forecast, the fewer situations we have where a customer arrives and doesn’t find the product they’re looking for.”
IVADO Labs’ solutions also offer some flexibility. “Inventory management offers a sound compromise between the risk of stockouts and the costs associated with excess inventory,” states Sylvestre. “Using simulations, we can understand how different inventory level choices influence this trade-off and make informed decisions accordingly.”
By comparing the manufacturer’s old processes with the new AI-based approach, it was also possible to quantify the benefits of the solution. “We simulated what would have happened if the manufacturer had used our model in the past, and our simulation indicated that they would have matched or improved their sales, while reducing excess inventory,” explains Grand’maison.
A key aspect of the project was the involvement of the academic community, particularly when addressing the more atypical aspects of the problem, such as estimating uncensored demand and managing inventory in a low sales volume context. The collaboration between scientific experts and the client’s planners made it possible to develop new methods that effectively met the client’s needs.
“At IVADO Labs, we have the opportunity to work closely with university professors,” adds Sylvestre. Researchers like Emma Frejinger from the Université de Montréal, Jean-François Cordeau from HEC Montréal, Maxime Cohen from McGill University, and Yossiri Adulyasak from HEC Montréal contributed to this project, notably by adapting different theoretical models to real-world scenarios.
“When we encounter a problem, we can rely on multiple resources both internally and within the academic community. We have access to a robust pool of experts who typically have already worked on similar challenges,” emphasizes Grand’maison. Fortin Simard adds, “These experts help keep us on the cutting edge of scientific advances.”
AI opens doors: the example of promotions
As was the case here, when companies embark on a digital transformation with AI, improving demand forecasting and inventory management often opens up new opportunities. This is particularly true when managing promotions.
“We created an AI model with the company that is capable of optimizing the allocation of promotional investments. It’s based on an estimate of the potential increase in sales relative to the AI baseline forecast when carrying out different promotions for each segment,” summarizes Fortin Simard. This approach allows promotions to be planned more strategically over several months.
Explore our insights on how AI tackles inventory management challenges to optimize stock levels, reduce costs, and improve supply chain efficiency.
Discover our research on using AI to detect panic buying and improve products distribution, published in AI & Society (a Springer journal) and recognized for its innovative approach to addressing real-world disruptions during the COVID-19 pandemic.