Okay, what is Agentic AI?
Agentic AI was without any doubt the most used buzzword at the Hannover Fair this year. The term could be found on a great many boots in all sorts of big slogans. A good opportunity for us — abusing the adage that there are no stupid questions — to visit a lot of these boots and ask: what do you actually mean by that?
The term agentic AI can encompass a wide range of technologies and applications—as quickly became apparent during our survey at the Hannover Messe. What most people's definitions of it seem to have in common is that they involve systems capable of performing actions. In its simplest form, a Large Language Model is used to analyze data and answer user queries. For example, you might ask such an LLM how to change a specific setting on your PC. Such an application is called an agent when it is capable of adjusting that parameter on your behalf.
Where opinions diverge, and where we therefore see different forms of agentic AI, is in the question of whether such a system must be autonomous. In some definitions, this is explicitly emphasized, referring to a system that does not need to wait for user input to take action. Instead, the agent is given an objective and must not only determine for itself what needs to be done to achieve that goal but must also effectively perform those actions. An example of an industrial application could be what we used to call Model-based Predictive Control (MPC). An agent then has a model that simulates the operation of a boiler, for example, and can manipulate the parameters based on this with the aim of maintaining a specific temperature at the outlet with the lowest possible energy consumption.
Virtual employees
However, the largest group of applications for which the term agentic AI is currently used does not possess that autonomy. These are systems that build upon the simple example we gave earlier of an agent capable of modifying a parameter on the PC. The things these types of agents can do are already quite complex, such as configuring and programming a machine control system. To some extent, it comes down to the user working with the engineering software they used in the past via such an agent, but instead of performing actions within that software themselves, they tell the agent what to do. This not only leads to a very intuitive way of working with the software, but the agent can also contribute a great deal of knowledge and intelligence itself, such as generating entire pieces of code, to stick with the machine control example. Such an agent can also assist the user in explaining why certain things are the way they are and in detecting errors.
And then there is a separate group of applications that, according to some, represent the true definition of agentic AI, and that is virtual employees. A number of nice examples of this were on display at the Hannover Messe. The nice thing is that it is apparently common practice to give these types of agents a name and a face. Users can then, for example, call upon the help of Susan, who is a specialist in supply chain management, or Bob, who knows everything about maintenance. What characterizes these types of agents is that they are usually no longer linked to a specific software application, but have access to multiple databases and applications relevant to their domain. In the examples we saw at the trade fair, it was not about autonomously operating agents, but about applications with which users can interact in an interactive way to perform analyses, but also to make adjustments. The user states what they want to change, and the agent works with the relevant software to effectively do so. The intention is that such agents will (eventually) also be able to work autonomously and that, for example, their goal would be to keep a certain KPI above a specific value and sound an alarm when this is no longer possible.
Application Programming Interfaces (APIs)
To make all of this possible – and to keep things running smoothly – agents use two crucial technologies alongside LLMs: Application Programming Interfaces (APIs) and Knowledge Graphs (KGs).
The concept of APIs has existed in software applications for much longer. It involves interfaces designed to allow third-party systems to interact with software. That third-party system can now also be an agentic AI, and the original software application is essentially unaware that it involves a form of artificial intelligence. In the past, such third-party software could be a system that enabled certain automation tasks, such as automatically modifying data sets, and this is no different when using an agentic AI. Only the way it does so is new, but that makes no difference to the software application.
The advantage of APIs is that the actions made available to third-party systems are predefined in the original software application. It is a finite list of commands to which certain restrictions can be imposed to make it impossible, for example, for crucial data to be deleted. In agentic AI, where hallucinations pose a real risk, this controlled interface is a useful tool to avoid missteps. The fact that the list of commands is finite and precisely defined also helps the AI achieve good results. After all, there is little room for hallucinations when the AI has only a finite set of commands at its disposal and its task is to reduce every train of thought to one or more of those commands.
Knowledge Graphs (KGs)
Another technology that is actually already crucial at the level of LLMs but plays a special role in agentic AI, and especially in its application with virtual employees, is that of Knowledge Graphs (KGs). In LLMs, neural networks are used as a brain to interpret matters and establish logical connections between data. The knowledge resulting from this, or that is made available to the LLM in advance, is not stored in the neural network itself but in special databases called Knowledge Graphs. These databases keep track of terms along with their meanings and their relationship to other terms. In this way, the Knowledge Graph is a representation of how an LLM views the world. In agentic AI, this is once again an important tool for achieving good results and avoiding hallucinations. In the example we cited earlier of Susan, who is a specialist in supply chain management, it is the Knowledge Graph that maps out everything Susan needs to know. The Knowledge Graph can also specify very precisely how a particular company's supply chain is structured, which leads the agent to think in terms of and processes specific to the company where the agent is deployed. Creating and managing these Knowledge Graphs will become a new job for consultants and integrators.
Incidentally, the concept of Knowledge Graphs, just like APIs, has existed longer than LLMs. Search engines like Google already used them to manage synonyms and other links between search terms and to map the context of potential search results. Well-known databases used for Knowledge Graphs include Neo4j, Amazon Neptune, and TigerGraph. A well-known tool for visualizing and modifying them is Apache Jena.
© Industrial Productivity, 10/05/2026, Picture: Illustration made by Microsoft Copilot