Request a demo
The world of supply chain, transportation, and logistics is complex, filled with high-stakes decision-making challenges. These are not theoretical problems — they are real-world issues organizations face every day. Think of procurement, inventory management, routing, fleet management, or workforce planning. Each of these areas demands critical decisions that directly impact efficiency, cost, and customer satisfaction.
What makes these decisions particularly difficult? Uncertainty. Many key elements are unknown. We may not know the exact demand for a product, the precise cost of operations, or when equipment might fail. Even external factors like weather can introduce uncertainty into routing challenges. Dealing with this "fog of uncertainty" is essential for effective decision-making.
For years, we’ve made significant progress using AI, especially through machine learning and optimization algorithms. Thanks to high-quality data, powerful algorithms, and immense computing capacity, we’ve become extremely efficient at solving complex, large-scale problems. Issues that would have taken years to resolve a few decades ago can now be solved in under a second — representing billions in gains when combining algorithmic improvements and computing power. This capability is already used routinely in areas such as scheduling and routing, even influencing the prices we see when booking airline tickets.
The traditional approach typically follows a pipeline: first, machine learning is used to predict uncertain elements like demand. Then, these predictions are fed into an optimization algorithm to generate a solution. The goal is to support human decision-makers by providing better information and potential solutions. Humans can then review, adapt, and implement these solutions, generating more data and continuing the cycle.
However, this common approach has a significant Achilles’ heel: it often assumes that predictions are perfect. In reality, machine learning forecasts are never fully accurate; there is always some degree of uncertainty involved. And while we’re good at solving problems, we are not always effective at solving them under uncertainty. Basing decisions on overly optimistic or inaccurate forecasts can lead to fragile solutions — resulting in disappointment when implemented in the real world, as the outcome doesn’t match expectations.
This is where the frontier of current research lies. The focus is shifting away from simply achieving the highest prediction accuracy to characterizing the uncertainty around those predictions — and, more importantly, accounting for that uncertainty in the optimization process.
Think about it: the critical question is not “Is my forecast accurate enough?” (e.g., “Is my average error 10% or 2%?”). The truly valuable question is, “Is the information my prediction algorithms are generating useful for the decision-making problem I’m trying to solve?”
This new perspective flips the traditional pipeline. Instead of starting with forecasting and then optimizing, the focus now begins with the desired outcome. The goal becomes training algorithms not just for accuracy, but to achieve the best possible results in the decision-making process. This demands a stronger integration between prediction and optimization — moving from two separate steps to a unified system that explicitly handles uncertainty.
This integrated approach aims to provide decision-makers with solutions that are not only effective but also interpretable — which is especially challenging when dealing with probabilities and uncertainty. It also seeks to incorporate the risk perception of the individuals and organizations involved. All of this is done to ensure that the actions taken in the real world actually lead to the expected outcomes.
Although this approach — integrating prediction and optimization under uncertainty — is extremely promising, it is also a technically complex challenge. Still, it represents a significant evolution in how AI can be applied to address the inherent uncertainty of supply chains, transportation, and logistics — moving beyond accurate predictions toward achieving robust, desirable outcomes. This is an active and continuously evolving area of research.
