AI in Public Administration: Critical Analysis of 'Spark Workflow'

• The German Ministry for Digital Affairs (BMDS) aims to modernize public administration using the 'Spark Workflow' generative AI project, targeting a 50% increase in processing speed for permit applications.
• Expert Stefan Kaufmann questions the project's viability, citing a lack of documentation and rigorous performance criteria.

Technical Limitations

• The system processes documents by converting them to Markdown and feeding them into language models to check for completeness and legal consistency.
• Kaufmann warns that generative AI is stochastic (probabilistic), making it unsuitable for administrative tasks that require 100% reliability.
• The software reportedly suffers from technical bugs, such as mishandling files with identical names, without notifying the operator of potential data loss or manipulation.

Strategic Alternatives

• Experts suggest that investing in semantic data management and machine-readable formats (e.g., CSV or Linked Data) is a more sustainable, transparent, and accurate approach.
• Standardizing data inputs would allow for rule-based, deterministic processing, ensuring the same output every time, unlike generative models.

Implementation Challenges

• The concept of 'human in the loop' is criticized as an excuse; it remains unclear if human operators can reliably audit AI results that are inherently based on random probability.
• The current Spark Workflow codebase is highly specialized for environmental permits, and adapting it to other administrative tasks would require significant manual configuration.