Industry
- Administration
Automation
Artificial Intelligence
Central Effects
- 95% time savings
- Massive cost reduction
- Complete, consistent documentation
- Scalable solution for future building blocks
Impacto en el negocio
- Significant relief for developers
- Documentation permanently up to date
- Resources for value-adding productivity free
- Reusable automation
Automated documentation of 500 process modules with AI: High documentation quality with minimal manual intervention
Regular documentation of technical process modules is essential for maintainability, further development, and quality assurance. However, there is usually a lack of clear responsibility, documentation quickly becomes outdated, and valuable developer resources are tied up, making this process often very time-consuming and resource-intensive.
At the beginning of 2025, CIB Group faced the challenge of regularly creating and updating documentation for approximately 500 CIB flow process modules. Find out how we solved this task here.
The challenge
Manual documentation of the 500 CIB flow process modules, including quality control, would have taken about 30 minutes per module. The effort was estimated to be about 30 working days.
In addition, additional effort is required because each newly developed component must also be documented, which further increases the overall costs and time required.
In addition to direct personnel costs, this results in considerable opportunity costs, as this time cannot be invested in strategic development or customer benefits.
Solution Approach
We have developed a fully automated documentation process based on CIB flow.
The technical implementation was based on a process that runs automatically every two months and reads in the current process modules as JSON files. These are processed by the Large Language Model (LLM) Claude Sonnet 4.0. LLMs are advanced AI language models that are capable of understanding and generating natural language. A low temperature* of 0.0 ensures precise and consistent responses, while the maximum token count* of 4096 controls costs.
As a result, the process generates fully generated HTML documentation with multiple CSS classes for flexible layout options. The finished documents can be published immediately on the learning platform, creating an efficient, repeatable and cost-effective workflow.
Even better: the process was designed to be model-independent and can be switched to new models such as Claude Opus, GPT-5.2 or Grok 4 with just a few clicks.
*The temperature controls the creativity of the output: a lower value results in more precise responses, while a higher value results in more creative responses. The token count influences the length of the output by setting the maximum number of words or characters.
Quality assurance & reduction of hallucinations
Since large language models work probabilistically, hallucinations can occur. For quality assurance, a two-stage post-processing step has been integrated:
1. Deterministic rule checking
- Pattern recognition
- Validation rules
- Automatic correction of defined structures
2. LLM-based post-processing
- Corrección de errores
- Structural improvement
- Consistency check
Using CIB flow and AI-based large language models, the documentation of 500 process modules was fully automated in just a few hours – at minimal cost.
Process highlights
By implementing an AI-supported automation process based on CIB flow and large language models (LLMs), it was possible to:
- Reduce the time required to create documentation by over 95%
- Fully automatically create documentation in under 5 hours
-
Costs for document creation alone reduced to a total of just €15
- A consistent, scalable, and maintainable solution has been established
High documentation quality with minimal manual intervention
The project convincingly demonstrates how intelligent process automation:
- Significantly increases efficiency
- Drastically reduces costs
- Quality improved
- Technical debt reduced
- Offers sustainable scalability
