From prediction to action: How AI can make supply chains smarter

At JD.com, one of China's largest e-commerce companies, artificial intelligence is being used not only to predict what customers might buy, but also to help manage what happens inside the warehouse. Its fleet of mobile robots can adjust their routes and tasks as order volumes or machine availability change, allowing the physical operations of its supply chain to respond dynamically to changing conditions.

JD.com is just one of several real-world examples highlighted in a new vision statement published in Manufacturing & Service Operations Management. The paper brings together 42 operations management researchers, industry practitioners and technology leaders, including Joren Gijsbrechts, Associate Professor of Operations Management at Esade. The researchers believe that examples such as JD.com illuminate an important point: AI is most valuable when it’s integrated with operations-management knowledge rather than used as a standalone technology.

The inventory problem

Supply chains are never static. Decisions have to be made in real time about how much stock to hold, where to put it, and how to respond to changes in demand and supply. AI is useful here in that it can process far more information and identify patterns much faster than traditional approaches, but the researchers argue that this isn’t the full picture when it comes to supply chain operations.

A demand forecast is not a decision. Even if AI makes a highly accurate prediction, that must be translated into an action that takes into account inventory, service levels, lead times, resources, and the company's risk priorities. The paper’s authors see AI as being within a broader operations-management framework. In other words, algorithms can learn and adapt, but they still have to operate within the physical realities of the supply chain.

JD.com's warehouse robots offer a tangible example. Their ability to adjust routes and tasks as order volumes or machine availability change shows how AI can operate within the physical realities of a supply chain, responding to changing conditions through operational decisions rather than simply producing information for a human to interpret.

From predicting demand to making decisions

The supply-chain industry as a whole has invested substantially in improving forecasts. But greater predictive accuracy doesn’t always result in better decisions. Small differences in a forecast can lead to very different outcomes when decisions involve complex constraints or discrete quantities.

The researchers therefore advocate a closer integration of machine learning with optimization and simulation tools. As the co-authors explain, “Operations Management domain knowledge defines the skeleton … objectives, constraints, trade-offs, whereas AI works within that framework to learn and adapt.”

In this way, AI could speed up the way in which supply chain managers use existing sophisticated tools. Large language models (LLMs) could allow planners to interact with optimization systems using natural language, potentially reducing the time needed to analyze scenarios or update models. For example, instead of a planner having to understand technical software, they could simply ask the AI what would happen if demand increased by 20%. However, the researchers caution that hallucinations, omitted constraints, and inconsistent outputs mean that LLMs should work alongside mathematical models, with appropriate safeguards.

The inventory problem and DRL

Deep reinforcement learning, or DRL, is an alternative approach when it comes to using AI to help with inventory control. It differs from conventional forecasting and uses reinforcement learning, which focuses on decisions made repeatedly over time. The AI system learns policies through trial and error, responding to changing demand and supply conditions according to defined rewards.

This could work particularly well in environments with complex inventories and rapidly changing conditions. But, as is often the case with AI use, there are limitations. The researchers point to the fact that the training requires massive amounts of data and simulated scenarios, and the ‘rewards’ must be aligned to business functions and goals. Furthermore, the more complex the model, the harder it is to interpret and trust.

Given these factors, the researchers foresee DRL as something that best works in tandem with human expertise, and not as a replacement for it. Automation could be possible for stable and repetitive tasks with fewer variables. Meanwhile, humans could continue to control decision-making in exceptional circumstances, and provide oversight, as well as decide upon strategic trade-offs when necessary. With the right combination of human judgment and AI, the co-authors say, DRL could become “a trusted partner in building faster, smarter, and more resilient supply chains.”

Real-world AI adoption

JD.com illustrates how this principle is already reaching the physical supply chain. Its robots do not operate in isolation: they are part of a wider fulfillment system in which AI helps translate changing information into operational actions. The paper also points to similar developments at companies such as Amazon, as well as AI-enabled applications in planning, inventory and logistics.

The wider market is moving in the same direction. A 2025 McKinsey survey found that 75% of supply-chain organizations were planning, designing or piloting AI applications, although only 19% were deploying AI at scale. Demand forecasting, inventory optimization and supply planning were among the leading generative-AI applications.

The challenge is therefore shifting from demonstrating what AI can do to integrating it successfully into existing operations.

AI still needs humans

Using AI to automate supply chains doesn’t mean eliminating human involvement.

For example, JD.com has announced aggressive plans to scale its automated supply chain. It wants to procure 3 million robots, 1 million autonomous vehicles, and 100,000 delivery drones in the next five years. But human workers will still be indispensable. The company employs 700,000 delivery and logistics personnel. The plan is to retrain delivery and warehouse staff in more technical roles.

The paper argues that as automation increases, workers may spend less time performing repetitive tasks and more time overseeing systems, resolving anomalies, and coordinating between automated processes.

AI systems can fail when conditions don’t follow the patterns they were trained on. Therefore, human oversight is essential. The researchers advise against excessive automation and recommend incremental implementation, stress-testing, and human override mechanisms.

The human-AI partnership

JD.com's automation plans illustrate how AI can change supply-chain operations: not simply by producing better forecasts, but by connecting intelligence to action.

The researchers' collaborative vision is that this integration will become increasingly sophisticated. AI may eventually support or automate more inventory, planning, and execution decisions, but humans will still need to define the objectives, constraints and principles that govern them. Companies will need to understand the limits of their systems and decide who is accountable when an automated system gets something wrong.

The question for supply-chain professionals is where AI can genuinely improve decisions while preserving the human judgment and operational resilience that complex supply chains still require.