ArKi

The project analyzes the requirements, technical fundamentals, and implementation options for a modular AI framework that integrates generative AI into administrative processes. The starting point is the observation that large language models (e.g., chat models, open LLMs) have great potential but have only been used sporadically in public administration to date.

Two prototypical use cases were implemented from analysis to minimum viable product (accuracy > 97%). A detailed cost-benefit analysis was carried out, taking into account technical feasibility, hallucination risks of generative models, and the effort required to transition from prototype to productive operation.

The result is a blueprint that provides multimodal, multi-model, and multi-agent-capable components (interaction, memory, planning, tool execution) as well as standardized interfaces to existing specialized procedures.

The central contribution is the conceptual and experimental development of a modular, agentic AI framework for the entire public administration. The overarching research question is: Which AI framework enables the interdisciplinary, beneficial, scalable, and sovereign use of AI technologies in public administration?

The methodological approach is divided into three steps:

1. Problem analysis – balance between under- and overspecification, integration of probabilistic and symbolic components, modeling of uncertainty, and prioritization of context-dependent resources;
2. Evaluation – definition of robustness, security, and explainability criteria, development of synthetic and realistic test scenarios, and iterative integration of human feedback;
3. Functional architecture – flexible communication and API layers, hybrid memory (episodic, semantic, procedural), combined planning mechanisms (probabilistic + symbolic), and controlled tool execution levels. 
    

The framework is implemented in an agile research and development cycle, systematically taking into account cost-benefit analyses, hallucination controls, and data protection aspects. The results show that a modular, multi-agent-capable framework significantly reduces development and operating costs, strengthens technological sovereignty, and at the same time ensures compliance with security and data protection standards.