AI has long been a part of the supply chain—in the form of forecasting models, optimization algorithms, and automated decision-support systems. But over the past two years, this has changed. With the emergence of so-called agent-based AI, the technology has taken a clear step from analysis to action.
“The big difference now is that AI doesn’t just answer questions or make recommendations. It takes action,” says Emma Karlsson, Industry Lead Solution Engineer and Nordic AI Champion at Oracle.
In practice, this means that AI agents can perform tasks, make decisions within given parameters, and interact with other systems—and in some cases, even with other agents. For supply chain functions, where complexity and interdependencies between systems are the norm, this opens the door to entirely new ways of working and results.
From Predictive AI to Agents
To understand this development, you need to rewind a bit. In other words, AI in the supply chain is not a new phenomenon, but something that has been around for many years.
“We’ve been working with what’s now called predictive AI for many years—for example, in planning, where statistical models are used for forecasting and optimization,” says Emma.
The next step was generative AI—technology that can create content, compose texts, and summarize information. But it’s only with AI agents that the potential for automation truly begins to be realized, Emma explains:
– The agents take it one step further. They can not only analyze or generate data, but they can also perform tasks and drive processes forward, which offers enormous potential for increased efficiency. This can involve everything from identifying discrepancies in the supply chain to initiating corrective actions, contacting suppliers, or rescheduling workflows.
What exactly is an AI agent?
The term ”AI agent” is widely used these days, but its meaning isn’t always entirely clear. An AI agent can be described as a digital entity that interprets a situation based on data, considers possible actions, and acts—either independently or with approval. However, there are different levels of autonomy. In practice, it’s a spectrum: from simpler decision support where humans play a central role, to more advanced solutions where agents can act independently within defined parameters. The choice of level often depends on risk, how business-critical the activity is, and how AI-mature the organization is.
”Many of the first agents we see are what we call ”advisors.’ They serve as decision-making aids and help users make better decisions,” says Emma.
The next step is more automated agents that actually perform tasks. And after that comes what Emma says is most interesting from a systems perspective: entire teams of agents working together.
– In that case, you have an agent who acts as a ”supervisor” and assigns work to other agents. It’s like a digital team. Next come the ”workflow agents,” which provide further automation.
In a supply chain context, where processes span multiple systems and functions, this can be particularly powerful. This means that complex workflows can be broken down and handled in parallel, with different agents responsible for specific parts of the process. The result is not only faster decision-making, but also a more scalable and flexible organization that is better able to handle variations and disruptions.
The supply chain is an ideal environment for agents
By definition, the supply chain is fragmented. Order-to-cash processes move through multiple systems, such as CRM, order management, inventory, transportation, and finance.
“It’s precisely in these handoffs between systems that a lot of friction arises—and that’s where AI agents can be of great help. By connecting systems and processes, they can create a more cohesive workflow,” Emma explains.
A concrete example is control tower solutions, where the entire supply chain is monitored and managed in real time, largely with the help of software.
– We’re now seeing initiatives where entire supply chain control towers are being built using AI agents, and where much of the coordination is automated. “For example, DHL Supply Chain uses our AI agents both tactically to identify and resolve integration errors and strategically to analyze cost patterns and suggest improvements,” says Emma.
The next step is agent-to-agent
An important development is that agents not only operate within a system, but also communicate with other agents and applications. This means that information and decisions no longer need to pass through manual interfaces or be integrated retrospectively, but can be shared and coordinated in real time across different parts of the system landscape. This reduces lead times and the risk of information loss during handoffs.
“The supply chain is rarely about a single system. It’s an ecosystem,” Emma emphasizes.
This is where concepts such as ”agent-to-agent,” or A2A, and Model Context Protocol Interaction (MCP) come into play. With MCP, agents can connect to data and tools—standardized across system boundaries—such as a planning system and a transportation system, and with A2A, they can connect with other agents. The result is a more dynamic and responsive supply chain, where decisions can be made more quickly.
Data, Security, and Control
Despite the great potential, there are also clear challenges associated with AI agents, particularly when it comes to data and security.
“Data is the foundation of everything. If you don’t have the right data quality, it doesn’t matter how advanced your AI is,” says Emma.
Another key issue is how AI is integrated into the system landscape. Here, she emphasizes the importance of working closely with the data source:
– We see a big difference between building AI on top of data and having AI built into the applications and the data warehouse from the start.
This affects both performance and security. One important aspect is that data does not need to be moved outside the organization’s environment. But transparency is just as important.
– It’s not enough for AI to provide an answer; you also need to understand how it arrived at that answer.
This is where concepts like ”observability” come into play—that is, the ability to track the agent’s reasoning, the data sources used, and the steps taken. According to Emma, this is crucial for building trust.
Humans Still Stuck in the Loop
A recurring question in the discussion about AI agents is how much responsibility should be handed over to the technology. Ultimately, it comes down to control versus efficiency—how much autonomy one is willing to accept in exchange for faster and more scalable processes. For many organizations, this is a gradual journey that begins with support functions and progressively increases the level of automation.
”We often find that customers want a ”human in the loop,’ especially when it comes to critical decisions,” says Emma.
This could involve approval steps or the ability to monitor decisions. At the same time, there are processes where full automation is both possible and desirable. The balance depends on the context, says Emma, continuing:
– It’s not black and white. Some processes lend themselves to automation, while others do not. This is reminiscent of when many companies made the transition from planning in Excel to advanced planning systems. Back then, many were still using Excel and trying to verify the results. Something similar is happening now.
The Greatest Risks: Fear and Superstition
When new technology is introduced, the discussion often centers on risks, but not always in the right way. The focus is often on technical or security risks, while business risks are sometimes underestimated—such as investing in the wrong use case or failing to translate the technology into actual business value. It is only when the technology is clearly linked to business processes that the full picture of risk emerges.
"Many people talk about being afraid of AI. But there's also a risk in relying on it too much," says Emma.
A common challenge is that organizations launch many AI initiatives without linking them to concrete business benefits.
– It’s easy to identify use cases, but much harder to realize the value. The greatest benefits often aren’t found in the simplest applications—such as automating email composition—but in more complex processes. It’s when agents solve more advanced problems that real value is created.
How should you get started?
For many organizations, the question isn't whether to start, but how. Emma's advice is clear: start simple, but not too small.
“Many systems already have built-in AI functionality. Start there, test, and learn. If you focus only on small use cases, you won’t realize the full potential. But if you start too big, it becomes difficult to see results,” she says.
In other words, it's about finding a balance—not trying to automate everything at once, but also not getting bogged down in small, isolated initiatives.
Autonomous ecosystems – ”Fusion Agentic Applications”
For Emma and Oracle, it’s no longer about individual functions, but about entire workflows. The combination of AI agents, IoT, and integrated systems could eventually create supply chains that are significantly more autonomous than they are today.
– In the next step, we will launch what we call Fusion Agentic Applications, a new type of enterprise application based on coordinated teams of specialized AI agents. “They’re designed to be goal-oriented, proactive, and capable of reasoning, and have been specifically developed to function effectively in complex business processes and replace large parts of the work done in traditional user interfaces,” Emma explains enthusiastically.
In the future, we will see more complex and autonomous solutions in which agents are connected throughout the entire supply chain. Emma emphasizes, however, that this is a process that will take place gradually, as organizations learn and adapt.
”This isn’t something you can just ”turn on’; it’s a process and an understanding that’s constantly evolving because AI is developing so rapidly, and you’ll gradually see how the agents contribute to more and more efficiency gains,” she concludes.
—
Text and interview by Marika Karlöf and Stefan Karlöf
