High quality supply chain management is becoming more and more critical to the success of companies as supply chains become more complex yet less stable and customers become more demanding. This has led to an increased drive for high quality analytics to support decision making. Many companies have embarked on projects to improve their data quality and data infrastructure. Meanwhile increasing compute power and data availability, along with a rapidly evolving artificial intelligence (AI) ecosystem has made the barrier to creating AI driven tools lower than ever. This combination of increased digitization and AI accessibility has set up the perfect environment for even very traditional companies to begin utilising AI in Logistics and Supply Chain. It’s no surprise that leading industry voices, such as the Association for Supply Chain Management (ASCM), have again highlighted Big Data and Analytics, and Artificial Intelligence (AI) in their “Top 10 trends in Supply Chain Management” for 2025.
The use of AI in supply chain
AI in supply chain management has the potential to significantly improve the efficiency of a range of activities, from demand forecasting and the core S&OP process all the way to shop-floor planning, quality control and client delivery. Recently Generative AI has brought a whole new range of tasks into the purview of AI.
AI has 3 concrete advantages over manual process:
AI can handle complex dependencies: Many decision making processes contain so many different factors that humans rarely manage to find optimal (or even feasible) solutions by hand. Typical examples of this are production planning or logistics tasks. The sheer number of different ways to plan make it impossible for humans to accurately assess all possible arrangements. AI techniques, once correctly implemented, can easily sift through the huge variety of possibilities and select optimal plans. Generally AI generated plannings are much more effective than those designed by humans and can be obtained much faster.
AI can accurately process large volumes of data: Many decisions within supply chains are based on predictions about the future: how much demand there will be for a certain product (demand forecasting), how much will a raw material cost (price forecasting), will a shipment arrive on time, etc. Each of these predictions can be based on a wide range of factors, from historical data and seasonal trends to highly specific economic forecasts. Manual forecasts generally struggle to effectively utilise all this information. AI on the other hand is very effective at making forecasts based on large quantities of data and can vastly outperform human accuracy in many cases, as it can take much more context into account.
AI can be run at scale: Many processes could easily be performed once or twice by humans but become either expensive or time consuming to perform at scale. Even if manually managing inventory levels (inventory management) might work well for 10 SKUs, it becomes impossible to handle 10,000s. For AI scale is no problem. Having an automated, reproducible process has additional advantages beyond simply saving time, it allows for growth and removes variability due to human biases. Generative AI is expanding the scope of tasks which can be automated. Many repetitive data entry or document processing tasks can also now be automated at scale.
These new AI powered supply chains don’t aim to entirely automate human labor but rather to give human planners better tools. These should automate tedious tasks, improve supply chain efficiency and allow humans to focus on strategic tasks and exceptional cases rather than day to day management. For example, once an AI inventory management system is in place, the role of a human would switch from directly managing various ERP system parameters to setting high level objectives such as reducing cost and only occasionally directly managing settings when exceptional circumstances occur.
With an AI powered supply chain, planning becomes a much faster and more automated process.. The ability to easily generate plannings for a wide range of scenarios unlocks a number of powerful use case:
Simulations: Digital twin-like simulations and “what-if” analyses can be performed allowing planners to easily see the effects of various scenarios. For example one could see what adding a new distribution center might do to delivery times or how reducing the number of machines available would affect throughput.
Robust Plannings: Robust plannings can be created accounting for a variety of potential scenarios. For example a planning could be produced which accounts for a potential machine failure or unexpected demand.
Bottlenecks Identification: AI and Simulations also identify bottlenecks such as identifying which stage in production is limiting total throughput or for which parts reducing supplier lead times would be the most valuable.
How AI is changing supply chains isn’t just via the direct benefits of any one application but rather by making the supply chain management process more automated and transparent, allowing for significant strategic insights. This has shown to drastically improve the decision-making process in supply chain management.
The technology underlying AI
There are 3 main flavors of AI commonly used within the supply chain: Mathematical Optimization (Operations Research), Machine Learning and more recently Generative AI.
Mathematical Optimization: AI for supply chain planning
A complex planning problem optimized via mathematical optimization with a human generated planning on the left and an AI generated one on the right. Red represents tasks which are impossible to execute.
Mathematical Optimization (also commonly referred to as Operations Research) is a set of tools which can be used to automate decision making processes, such as when to produce which products on which machines, how to route trucks or how to arrange warehouses.
Mathematical Optimization works by creating a mathematical description (or model) of a planning task. Various algorithms (or solvers) can then be used to identify the optimal decisions based on this description. The mathematical description is the most crucial part and consists of 3 parts:
Decision variables: The choices which the algorithm can make, for example when to perform a task or which packages should be assigned to which truck.
Constraints: A set of considerations that must remain true for a planning to be usable, for example a truck cannot be in two places at the same time or a machine can’t be used for two tasks at once.
Objective: A description of what one aims to achieve, for example use as little fuel as possible or complete all tasks in as little time as possible.
The quality of the output of this flavor of AI is fully dependent on how well this mathematical description aligns with the actual reality of the process being performed and with the actual objectives of the business. For this reason, successful utilisation of Mathematical Optimization requires an effective collaboration between business leaders, planners, and AI experts. LINK TO JORIS BLOG POST ONCE PUBLISHED
Machine Learning: How AI makes predictions
Examples of Sales forecasts generated using machine learning with uncertainty intervals.
Machine learning is a collection of techniques which automatically “learn” from examples (data). By “learning” we simply mean that machine learning automatically determines the best way to predict a specific outcome by determining what works best on the provided data. The key advantage of machine learning is that it can effectively incorporate information from a wide range of sources. Within supply chain management machine learning is often used for forecasting tasks such as demand forecasting but it can also be used for a wide variety of related applications such as fraud detection, predictive maintenance or quality control.
There are two key elements which need to be contained in the data used for machine learning:
The target: The outcome which one aims to forecast. For example in demand forecasting this would be past (unconstrained) demand.
The features: Factors which the machine learning algorithm can base its prediction on. These could be historical values, for example the number of sales in a previous month, but they can also include leading indicators, for example planned promotions or economic forecasts.
The key to the results of machine learning based solutions is highly dependent on the data provided, there are 3 key aspects of data to consider:
Predictive Power: Data needs to be informative about the task at hand. Simply adding more data is not guaranteed to improve a model. Machine learning is not magic and it cannot make accurate predictions unless the data required to reasonably make those predictions is present.
Reliability: Data should accurately reflect the underlying process it intends to measure. Data with many missing values, errors or other tampering is unlikely to produce very accurate predictions as machine learning will learn to reproduce erroneous patterns.
Volume: While the emphasis should be on data quality, data volume is also very important for machine learning. Performance generally increases proportionally to the volume of data available.
Variety: Machine learning generally works better the more similar the environment a model is used in is to its training data. Therefore having a sufficient variety of data, covering all expected usage scenarios is also important.
The best of both
An integrated decision making engine where an optimizer (Mathematical Optimization based) and Forecaster (Machine Learning based) iteratively work to select a resilient planning.
Mathematical Optimization and Machine learning are very effective for making decisions and predictions respectively, but many processes contain elements of both tasks. Here an exciting new direction within AI is developing integrating both machine learning and optimization into a combined decision making engine. This is particularly interesting for making resilient AI systems as now optimization engines can also take into account the uncertainty inherent in forecasting processes resulting in plans which can cope with uncertain processes.
Generative AI
A potential interaction between an inventory manager and a Generative AI based agent.
Recently an additional form of AI has had a lot of success, Generative AI. The most popular example of this type of AI is ChatGPT. At its core Generative AI is just machine learning, with the previous words in a text as its features and the next word as a target. Surprisingly this simple procedure has led to surprisingly powerful models which are capable of performing a wide range and text based tasks.
How one works with Generative AI is significantly different to traditional machine learning. Using Generative AI generally doesn’t require additional training data as it has already been trained on human amounts of data. Instead the main factor in developing Generative AI based tools is adapting what is fed into the model: the instructions given to a model, generally called a “prompt” and what additional information is provided, generally via a process called Retrieval Augment Generation (RAG). Generative AI can also be provided with “tools” such as making database queries or sending email resulting in autonomous agents which can independently execute complex workflows. To learn more about Generative AI see this article.
Within supply chain the main applications of Generative AI are in:
Simple task automation: A lot of time is spent on tasks which require some understanding of language but are otherwise routine, for example replying to an email asking for details about a product or placing an order for spare parts. Generative AI is quite capable of dealing with simple automation like this which requires understanding text but only routine actions.
Information extraction: Generative AI is quite effective at extracting data from text, images or audio which would otherwise require manual effort. For example Generative AI can be used to automatically process hand filled forms or extract key information from recorded meetings.
Information retrieval: Generative AI can also be leveraged for information retrieval. For example Generative AI can be used to find answers based on documents such as an internal sharepoint. Generative AI can even allow non-technical users to easily interact with multiple databases by generating and running queries based on a description of the desired result.
AI in practice
The process of implementing an AI tool generally consists of taking a complex challenge and formulating it as an optimization or machine learning task. Utilising high quality data this can then be used to automatically inform or directly compute the best course of action. The technical aspect, designing and running an ML or optimization algorithm, is naturally very important but the challenge we see most companies struggling with is correctly understanding how AI can be applied and arriving at a formulation which addresses their actual problem.
For this reason the success of most AI projects rely on a close collaboration between data, business and technical expertise.
Data: Without (the right) high quality data, AI will struggle to correctly understand what is happening and will make poor predictions.
Business: If an AI application is developed which doesn’t correctly take into account the realities of the business within which it will be used its recommendations will often be either inaccurate or unusable.
Technical Expertise: While tooling has improved significantly, many technical challenges remain in implementing AI and without an understanding of how the underlying technology functions developers are unlikely to develop good solutions.
At superlinear we pride ourselves on our ability not just to build highly technically performant solutions but also solutions exactly matched to the unique needs of our clients. You can see more about previous experience designing AI solutions for supply chain management here. To see how we can help realise the benefits of AI in your supply chain contact us here.
Q&A
How can AI mitigate disruptions in the supply chain?
AI can be used to detect disruptions through techniques such as anomaly detection or sentiment analysis. If AI is used for planning, it can also identify vulnerabilities in a supply chain likely to lead to disruptions such as overreliance on unreliable suppliers. Further when disruptions do occur AI can quickly assess the situation and provide alternate solutions.
How can AI help create more sustainable supply chains?
By making environmental goals, such as energy usage or waste, an explicit target in an AI based planning, AI can optimize to meet sustainability goals while still maintaining overall operational efficiency.
What is the connection between AI & Digital Twins?
Digital Twins and AI are highly symbiotic. Digital twins can provide the data required for effective AI tools. On the other hand AI greatly increases the value of digital twins. AI can be used to make forecasts and optimize processes within the digital twin. It can also identify and predict disruptions. Further AI can be used to enhance the ability of digital twins to perform accurate "what-if" scenario analysis.
How will generative AI and Agentic AI revolutionize supply chain processes?
Generative AI aims to make the supply chain management process as frictionless as possible. AI agents will be integrated into all supply chain management tools. Humans will simply need to describe the tasks they wish to perform and generative AI agents will automatically perform them. This could range from looking up stock levels for a specific part to finding a more cost effective inventory management policy or even finding new suppliers and likely negotiating prices directly with a supplier’s AI agent. At least this is the vision, whether this will ever be achieved remains to be seen.
How can AI enhance demand forecasting accuracy?
AI can accurately incorporate vast amounts of data, including market trends, customer purchasing behavior, seasonal patterns, and external factors like economic conditions and competitor activity. Further Machine Learning models can also complex patterns in the data resulting in more accurate predictions. This can all be automated meaning that forecasts can be updated much more regularly, often daily.
How can AI optimize inventory levels?
Improved demand forecasts directly improve inventory levels, allowing for lower safety stocks. Further AI can analyze stock levels and determine optimal stock level balancing the risks of overstocking and stockouts while minimizing holding costs accounting for various factors such as demand variability, supplier performance and market conditions.. It can also analyse the MRP process and determine optimal strategies for optimizing the replenishment process.
