AI-assisted business engineering - a field report

AI in business engineering at Accso

A tender for a system to manage and control rail operations within ports served as the basis for testing AI in business engineering. In the first step, we analyzed the purely textual specifications for this application using Langdock. On this basis, we were able to create a description of the technical modules, use cases, processes and data models of the system to be built.

The result was passed on to a development team, which used it to develop a proof of concept, also supported by AI with agentic coding. We are very satisfied with the result: a presentable, mature software was created in a very short time.

Tools and technologies used

We mainly worked with models integrated in Langdock such as GPT, Claude or Gemini. We also used Camunda with its AI feature for generating BPMN models. We used GitLab and Visual Studio Code as a bridge to the development team.

While the development team had a strict ban on writing code manually, we still did some tasks manually, such as creating stories in the repository, due to the high initial effort involved.

Our learnings from the use of AI

Precision is becoming more important 

The main finding for us was that precise business engineering becomes even more important when using AI. Since the generated artifacts flow directly into the AI-supported software generation, their critical interpretation, e.g. by an experienced developer, is initially lacking. With a machine interface, the data model, requirements and processes must be described precisely so that the result is correct.

Using proven models and methods 

In order to achieve the necessary precision in the description of the business domain, methods from domain-driven design have proven their worth. We have achieved the best results when we have given the language model concrete models based on known modeling languages.

Language models are best operated with language

A perhaps seemingly trivial realization was that language models are best fed with textual input such as PlantUML, XML, JSON or Markdown. This avoids unwanted interpretations, for example when reading out images, and can ensure greater precision of the input. BPMN is also very suitable, as there is always an XML description with machine-readable semantics behind the process mapping.

It is worth comparing different models

For quality assurance purposes, it is worth using different language models and comparing them with each other. Are the results plausible or do different models come to different conclusions?

Quality assurance by humans and machines

Humans cannot yet be replaced, especially in quality assurance (human in the loop). Experienced experts should always carry out the final inspection and approval.

What happens next?

The findings from the project are being incorporated into our internal "AI-assisted software engineering" working group. We are noticing increasing demand from our customers and the market in this area and are continuing to develop the relevant expertise in our team.