Agentic AI in Supply Chain Optimization

In a conversation with Sander Vlot, VP Center of Excellence at ORTEC Data Science & Consulting, and Cynthia Luijkx, Senior AI Engineer at ORTEC Supply Chain Planning, it becomes clear how agentic AI works in practice, where its opportunities lie, and what steps companies can take to start exploring it. Their insights show that agentic AI is not a passing hype but a practical development that automates repetitive work, supports fuzzy decision-making, and makes complex software more accessible for planners and decision-makers.

Generative AI has delivered impressive results but also has clear limitations. “It’s all based on large language models trained on massive, but static datasets,” says Vlot. “That produces well-written text, but such a model can’t answer seemingly trivial questions such as where there are traffic jams right now, or what will the weather be tomorrow. That kind of information simply isn’t in the training data.”

AI agents add a crucial new layer: the ability to interact with the outside world. “If you want to use up-to-date information, you can’t rely solely on a language model,” he continues. “You let it communicate with the internet or with other software. AI agents can work on their own, embedded in workflows, or in collaboration with other AI agents; all different forms of agentic AI.”

For supply chains, interaction with the external world is essential. Planners constantly respond to changing conditions such as inventory levels, delivery issues, or transport delays. Luijkx emphasizes that this context dependence is precisely what makes agentic AI valuable. “For traditional automation, you need to know exactly what happens in every case, while human decision-making is often fuzzy. AI agents create new opportunities to automate or partially automate tasks that previously couldn’t be automated.”

Optimization software for supply chains is powerful but often complex. ORTEC’s models are used worldwide, yet clients don’t always leverage them to their full potential. “When planners want to experiment with different algorithm settings to tweak the results, they typically request help from an ORTEC optimization specialist,” says Vlot. “Now AI can help bridging that gap. The software itself doesn’t get smarter, but AI makes it possible to use it better.”

Luijkx gives a concrete example. “Within supply chain planning, we’re working on an AI assistant within the Zone Engineering Manager. A planner can group customers there to create territories and visit day assignments, supported by an advanced optimization model. In the past, this came with an elaborate configuration panel, which required extensive training to manage in practice. With an AI agent, a planner can simply describe in plain language what they want, and the agent translates it to settings. That translation used to be difficult for users to do themselves without trial and error or an optimization expert.”

The same principle applies to user support. “We’ve built an AI assistant for explaining optimization models,” says Vlot. “Instead of writing a dry manual, we can offer more contextual guidance through a chat interface. That helps enormously with adoption and understanding.”

This makes complex technology far more accessible for users. “We don’t yet know exactly what the planner’s role will look like in the future,” Luijkx notes. “But it’s clear that much more will become possible, and it will undoubtedly change. For example, an agent can remember that three days ago we already tested a certain scenario and use that knowledge to make new suggestions.”

The hype surrounding AI sometimes leads to inflated expectations. “If you look at the flashy examples online, they show ten agents taking over an entire marketing agency,” says Vlot. “That’s a big leap, and within the complex world of supply chains it’s not realistic right now.” He advocates a step-by-step approach. “Ask yourself which part of your work you’d like to automate. See if a language model can do something meaningful with your data. Start small with experiments. Those are easy and inexpensive first steps.”

Luijkx explains that ORTEC took a similar approach. “I started by sending dummy data to a model just to see if anything useful came out. We deliberately started small. That gives you room to learn without big risks. Everyone is figuring out how this technology works and how to use it best, so leaving space for experimentation and having the courage to quickly discard what doesn’t work is crucial. That’s how we learn what’s truly possible.”

The key, both emphasize, is to identify bottlenecks in your organization. Where are the most hours lost? Which analyses are left undone? Those are the areas where agentic AI can make the biggest difference.

The potential of agentic AI depends entirely on the availability and quality of data. “If you connect an assistant to enterprise data, it’s often hard for it to do something useful,” Luijkx explains. “A good test is to imagine a new team member starting today, someone with general knowledge but no company-specific context. Is all the data they need readily available, and can they start working with it right away? Often the answer is no, and that tells you something about your AI readiness.”

Granting agents access to such data and systems also introduces new security risks. “Once you make an agent accessible to the outside world, all the proprietary data you feed it with to make it smarter could potentially be prompted out by malicious actors,” says Luijkx. She points to initiatives from Microsoft and others that provide special datasets for testing AI applications against such ‘prompt injection attacks.’

Evaluation is another focus area. “With generative AI, the same input can yield different outputs,” Vlot explains. “That’s why we build QA datasets together with domain experts to properly assess quality.”

Clients are both curious and cautious, says Vlot. “Senior stakeholders often ask: this all sounds great, but how much does AI really matter for us? There’s usually top-down pressure to ‘do something with AI,’ but at the same time there’s a lot of skepticism.”

That skepticism is understandable because adopting AI successfully requires readiness in several areas. “Your data needs to be ready for consumption by LLMs, but you also need to know how you’re going to approach it,” says Luijkx. “Which people in your organization can build, govern, and maintain your agents? How are you going to evaluate them, and more importantly, get user trust?”

ORTEC sees its role as guiding clients through the hype. “We cover the full spectrum from traditional methods to new AI techniques,” says Vlot. “And sometimes, the right answer is actually not to use AI at all.”

Meanwhile, ORTEC is also developing useful AI applications within its own software to assist users as effectively as possible. Luijkx observes that much of the low-hanging fruit is still untouched. “We and many others started with complex applications like chat assistants, while there are still so many simple tasks that take planners a lot of time. That’s where AI now offers opportunities. Think of repetitive work that can easily be automated. That’s where the real revolution will happen.”

The rapid pace of AI’s evolution surprises many. “It’s driven by massive investments,” Vlot explains. “The science behind AI has been around for years, but since ChatGPT, people can experiment with it themselves. That democratized AI and attracted unprecedented levels of investment.”

Luijkx adds, “Normally, technological innovation reaches academia and/or businesses first. This time, individuals could experiment directly. In the U.S., 61 percent of the population has used Generative AI systems, and 36% uses them weekly. That widespread adoption accelerates innovation tremendously.”

The future looks promising. “We’re moving toward a world where you don’t just click through screens anymore. You work together with your software toward a goal,” says Luijkx. “That saves time and lets you focus on more meaningful work.”

Vlot sees great potential in real-time applications. “Imagine you can adapt more quickly to new situations. That’s crucial in supply chains.”

At the same time, Luijkx warns against over-optimism. “Today’s technology allows demos to be built quickly, but creating something that works correctly 90 to 99 percent of the time requires time and expertise. Starting simple may be the key to building things that actually work.”

Still, complexity is no reason to wait. “The technology and its application methods are evolving so rapidly that standing still isn’t an option. It’s crucial for organizations to stay engaged and understand what it means for them specifically.”

Vlot notes that the interaction between humans and computers is definitely changing, and that companies should consider how much autonomy they want to give their agents. “Andrej Karpathy, former Director of AI at Tesla, recently coined the term ‘autonomy slider’ for user interfaces that let users set how much autonomy AI has within a software product. We should all be thinking about what that autonomy slider would look like in our own software,” he says.

Agentic AI promises to make supply chains more efficient and adaptive. The real value lies not only in the quality of the models but in their accessibility and impact. By making complex software understandable for planners without specialized knowledge, the power of optimization moves closer to daily practice. “Maybe it’s ultimately less about quality and more about impact,” says Vlot. “AI makes software more accessible, and that’s evolution. But when you look at how many routine tasks, for example in back-offices, can now be automated, I think it’s fair to call that a revolution.”

Luijkx concludes with grounded optimism. “It’s incredible how fast the technology is moving. I do see a shift in how people think about it. Early on, the idea was that everything would work once the models got better. Now the models are already excellent, and the real challenge is how you apply them. That’s where we make the difference.”

Agentic AI, in that sense, is a catalyst helping supply chains become smarter, faster, and more accessible, and it’s set to play an indispensable role in how companies organize their operations in the years ahead.