How INFORM combines optimization and AI to create a new form of decision-making intelligence

Interview

Question: Mr. Lehnen, what fascinates you about developing AI for real operational processes?

Answer: AI allows us to model complex, dynamic processes the way humans intuitively understand them, but with the speed and precision of modern software. Until now, many challenges in industrial practice could only be modeled incompletely using software. Now, we can take the next steps away from systems that only plan with known information and toward systems that predict the impact of decisions on our processes. This development work is very exciting, especially at a company like INFORM. We have been optimization experts for many years. Now, however, we are on the threshold of a new generation of systems that closely resemble real-world processes.

Question: AI for industrial processes differs from how people perceive AI in everyday life. How does it differ from tools like ChatGPT, for example?

Answer: ChatGPT is a language processing model that generates plausible formulations. In industrial processes, however, the goal is not to generate sentences, but to make decisions. These decisions are always subject to hard constraints. These constraints include available resources, legal requirements, and physical dependencies. We need AI that can understand and explain what is happening in a process and the likely consequences of a particular decision.

Question: Consumer AI can only formulate answers. Industrial AI, on the other hand, must support decision-making.

Answer: Yes, and in a way that is both technically sound and responsible. A mistake can be costly or bring entire processes to a halt. That's why we need models that reveal uncertainties, clearly structure relationships, and provide a solid foundation for human decision-making. Good AI does not rely on intuition. This is precisely where our system view differs fundamentally from generative tools, which primarily extrapolate statistical patterns but do not map physical-logical relationships.

Question: INFORM has been working with optimization methods for decades. Many processes were already algorithmically controlled before the AI era. How do these methods work, and why does AI complement them so well?

Answer: Optimization models calculate decisions within a clearly defined framework. Goals, restrictions, and available resources are described mathematically with precision, as are their relationships. Within this model, they quickly, reliably, and transparently find the best possible solution. However, these models reach their limits when processes are incompletely described, change frequently, or when additional patterns and risks come into play.

Question: This is where modern AI methods, such as process mining and machine learning, come into play.

Answer: That's right! These methods recognize trends, uncertainties, and interactions that aren't explicitly defined in the model. Thus, they provide a forward-looking assessment of the process context. This represents a significant improvement in quality because AI recognizes patterns that classic optimization models cannot. Combining the two creates systems that calculate optimally and understand relationships. This provides the basis for making the best possible decisions in complex situations.

Question: At INFORM, we use the combination of optimization and AI described above to create a product vision called Process AI. What is behind it?

Answer: At its core, Process AI involves equipping AI with a mental model of the entire process flow — in other words, providing it with an understanding of what happens in the process, why it happens, and the consequences of individual decisions. Traditional planning systems often focus on individual subtasks, such as who executes a transport order, who takes it on, and which vehicle is suitable for it. These things can be determined. But what if a truck gets stuck in traffic before carrying out a notified order and arrives late? Process AI broadens the view by interlinking perception, interpretation, planning, execution, and learning in a continuous cycle.

Question: That sounds abstract.

Answer: Yes, but it quickly becomes concrete in practice. AI should be able to recognize when a process is stalling, how disruptions are spreading, and what adjustments will stabilize the process. Sticking with the example: Can I predict a traffic jam based on the time of day and expected traffic and send the truck on an alternative route early on? Or, if there is no suitable alternative route, do I have to postpone the transport order? What does that mean for those waiting for the delivery? AI should assess the best options and provide the basis for human decision-making.

Question: Will it be possible to operate this AI with voice commands and interact with it like a chatbot?

Answer: Yes, that is likely. Our applications are very complex beneath the surface. That's precisely why they need to be as intuitive as possible to use. Voice access lowers the barriers to entry and makes them easier to use. At the same time, we are developing AI agents—modules that can perceive, process, decide, act, and learn. However, one thing remains unchanged: the responsibility lies with humans. AI agents are there to support, not to make automated decisions.

Question: To bring the vision of "Process AI" closer to reality, INFORM is conducting several pilot and research projects. Could you provide examples of how AI is currently being used in various process environments?

Answer: In the dynamic and fast-paced environment of airport ground handling, where decisions must be made quickly, we are pursuing a new technological approach: combining forecasting and planning in an integrated model. Specifically, this project involves planning cargo transport in passenger aircraft. Previously, cargo space capacities were forecasted using machine learning, and then cargo was loaded in a separate process. However, conditions can change at short notice depending on passenger numbers, baggage volumes, or changes to the flight schedule. Disruptions to the process can quickly arise when forecasting and planning are carried out separately. Our integrated approach brings both together.

Question: How exactly does this work?

Answer: The AI forecasts available capacities and immediately processes this information in a shared planning model. This enables the system to address uncertainties before they cause operational issues. The result is more stable loading decisions and significantly fewer last-minute reschedulings.

Question: Are there any other projects?

Answer: Another example comes from automotive logistics. The task here is to plan the transportation of new cars from manufacturers to dealers. Much of the operational work involves reviewing and prioritizing numerous emails every day. We are developing a communication assistant that automatically recognizes and classifies requests, providing suitable response suggestions. This reduces employees' workload and helps them identify critical inquiries more quickly.

Question: Is one of your other pilot projects also about a communication assistant?

Answer: Yes, we are developing a specialized chatbot that can accurately answer technical questions about data standards. This is important in areas where precise information is required. Another project involves an available-to-promise model that automatically verifies if an order can be accepted and determines the necessary resources. We are also developing an AI solution for demand forecasting in sales planning for supply chain management. These models use neural networks to predict demand trends and patterns. The goal is to identify demand fluctuations early on and plan production and replenishment processes more efficiently.

Question: Looking at all these projects, what are the biggest hurdles?

Answer: Clearly, it is data quality and access to operational process data. Many processes have evolved over time, resulting in heterogeneous or incomplete data structures. Additionally, many companies find it difficult to release their data, which is understandable. At the same time, AI models need this real-world basis. Without it, the system cannot learn what actually happens in everyday life, causing models to make decisions that are out of touch with reality.

Question: How can this problem be solved?

Answer: One solution would be to create formats that ensure companies' data is protected and clearly define how it is used. In other words, we need genuine, trusting data partnerships. Only then can we develop AI that works reliably in the real world. That's what we're working on at INFORM.