Introduction
Now that we are confronted almost daily in the press with potential applications and the possibilities of AI – and large language models in particular – the term ‘Agentic AI’ is also beginning to emerge. So what exactly does this mean?
If one were to seek a definition, Agentic AI describes autonomous systems (also known as agents) that can interact independently and are capable of performing more complex tasks and striving towards predefined objectives. To this end, they carry out planning processes, utilise other tools for sub-tasks and/or partial results, and can also work iteratively on solving a task without being controlled by humans.
In this way, complex tasks – some of which require a multi-stage approach – can be solved, with multiple AI agents able to collaborate. The concept of agent systems is nothing new; they have been around since the 1980s, but are now being equipped with AI components, such as an LLM, to solve a task or sub-task.
Example: Credit and supplier monitoring
Here is an example from credit management of how monitoring processes for customers or suppliers could be designed using various agents in the future: For customers, an AI agent continuously checks whether new information provided by the customer alters the assessment of their financial situation and, consequently, the credit limit granted. To this end, a monitoring agent searches for external information, e.g. on social media or on information or press platforms, and evaluates this information. Another AI agent analyses the customer behaviour observed internally. This includes, for example, sales figures, payment behaviour or complaint behaviour. The results of these evaluations are provided to the AI agent responsible for assessing the financial situation. This agent checks whether this leads, for example, to changes in the customer rating. If this is the case, the rating is adjusted. Another AI agent then makes recommendations on how to proceed in this situation. The respective consequences of these recommendations are outlined. The credit officer can now make a decision based on this information.
A similar approach could be taken to supplier monitoring. Here, in addition to the financial situation, the agents also check adherence to delivery deadlines, quality assessments of deliveries and the completeness of deliveries. In addition, another AI agent monitors the supply chain, for example by checking for potential regional restrictions, etc. Changes to export and import duties could also be examined and assessed here. If there is a database of other potential suppliers, it would then be possible to assess, in line with specified objectives, whether an alternative selection should be made. This scenario can be expanded as required. Similarly, the availability of transport capacity or alternative transport options can be integrated, provided the relevant data is available. Decisions could then also be based on how quickly a required product needs to be available.
Findings and Outlook
Two things become apparent here:
- Various AI agents are being interconnected. The design of each individual agent is no simple matter and can involve considerable effort (e.g. training, data provision). Similarly, orchestrating the overall system is not straightforward. The total effort required to develop a good solution should therefore not be underestimated. And the database requires continuous maintenance.
- In many processes, people with extensive specialist knowledge are still required to make final decisions, whilst routine and high-volume tasks are automated and thus accelerated.
Overall, it can be assumed that such solutions will also find their way into companies in the medium to long term.
