Decisions embedded in workflows

Ideally, the agent learns its behaviour directly by observing the already implemented process or a sub-process.

In workflows, decisions are usually made either within work steps (activities) or at designated branch points, such as Business Process Model Notation (BPMN) gateways. The spectrum ranges from simple binary to complex cases.

Many organisations are already using BPM as a foundation for digital transformation. However, process diagrams that include decision-making processes can quickly become confusing. Outsourced business rules run the risk of being overlooked or left unmaintained. Decision problems that involve multiple activities and are to be handled by a software agent can often be aggregated. Transforming existing processes, such as combining Business Process Model Notation (BPMN) and Decision Model Notation (DMN), simplifies the maintenance, design and deployment of software agents.

Decision mining and optimising

If the decision-making process is not fully or only partially mapped, software agents enable decision mining to trace or extract decisions. Completely manually modeled decision processes, such as pricing, plant care, or ordering processes, can be automated through the use of software agents.

The software agent learns the desired behavior through simulation, by observing the implemented process or by receiving feedback on user preferences. It can do this by analyzing records of the status-action sequences as well as in real time and checking its output against the process parameters. Once the software agent has learned the desired behavior, it can extend and replace the previous implementation or display suggestions to the user.

Software agents can be equipped with basic skills that they have already learned, which they can acquire from public data or simulations and, if necessary, draw on existing external knowledge that is encoded in a knowledge graph or in large language models, for example.

Software agents are clearly separated from their environment by their architecture and interact by observing and acting. This clear separation of responsibilities enables each agent to fulfill a specific task with a clear goal.