AI Development: Precision and Reliability for Sensitive Applications

Agentic AI represents the next step in the development of large language models. But the reality is more complex than the promise:

• Just how robust are LLMs' planning capabilities, really?
• Where are cost-effective models sufficient, and where is flagship capacity needed?
• How do you design robust systems when their behavior differs fundamentally from that of traditional, deterministic software?

The productive use of AI agents requires a thorough understanding of their capabilities and limitations.

We develop AI agents that can independently plan and execute multi-step tasks while always remaining within clearly defined parameters. Through the interaction of specialized tools and data sources, they handle collaborative workflows along predefined processes, independently selecting the appropriate methods to achieve their goals. By connecting to internal knowledge bases and existing applications, they operate in a context-aware manner, precisely tailored to the respective application domain.

We investigate how the often abstractly formulated requirements from regulatory frameworks, such as the EU AI Act, as well as organizational guidelines, can be translated into technically verifiable architectural principles. Under the term “Agent Policies,” we define which decisions an agent is permitted to make, how it must justify its actions, and where binding guidelines apply. The goal is not to treat transparency, explainability, and governance as afterthoughts, but to embed them as structuring principles within the software—documentable, verifiable, and traceable throughout the entire lifecycle.

This allows even complex workflows to be reliably automated without deviating from the technical and organizational framework.

Relevant knowledge within organizations is often scattered across diverse sources and formats - in documents, specialized systems, as text or XML, or as image files. Traditional keyword searches quickly reach their limits here: they find terms, but not the connections between them. Implicit experiential knowledge remains untapped, and redundancies and contradictions go unrecognized.

We develop solutions for organization-wide knowledge management that consolidate distributed knowledge from diverse sources and make it systematically accessible. Implicit experiential knowledge from documents, databases, and line-of-business systems is transformed into a searchable and interconnected knowledge base that serves as a central information source for users and downstream AI systems via semantic search. Through Retrieval Augmented Generation (RAG), we link generative language models to verified data sources. This also prevents the hallucinations typical of LLMs, enabling fact-based, verifiable, and methodologically traceable system responses that meet the highest standards.

Our many years of work in information retrieval show that proven retrieval methods and modern language models enable new forms of interaction with knowledge bases. We investigate hybrid architectures and neurosymbolic approaches, such as GraphRAG methods. GraphRAG combines structured knowledge modeling with generative language processing, thereby bringing together the strengths of both paradigms. The focus is on the use of explicit knowledge structures for more precise retrieval and on the question of how domain-specific knowledge can be formalized in a way that makes it reliably usable for LLMs.

Complex documents such as standards require consistent structures, terminological precision, and compliance with regulations. Manual review and maintenance are time-consuming, error-prone, and do not scale well with growing document collections and version histories. Especially in the field of standardization, where errors can have far-reaching consequences, there is a great need for systematic support.

Our AI-powered authoring system evaluates standards texts during the drafting process based on formalized rule sets. It identifies structural and semantic inconsistencies and proactively provides optimization suggestions, significantly increasing consistency and efficiency in the standardization process.

As a department that actively participates in numerous standardization efforts (NATO-STANAGs, DIN), we know the challenges firsthand. This dual role as both users and developers enables us to design solutions that truly address real-world needs. Using standardization as an example, we are investigating how AI-supported document processing can also be applied to other fields with comparable requirements for structure, consistency, and traceability.

Many organizations want to use AI technologies but are prohibited from transferring data to external services for data protection and compliance reasons. Large commercial language models are powerful and relatively easy to use, but they are not available in such environments. Those who must rely on smaller, self-operated models face entirely different challenges: lower model capacity, greater need for optimization, hardware limitations, and the question of how to still achieve reliable results.

As a research institution, we ourselves work with sensitive data in numerous projects and know the limitations from our own experience. Data sovereignty is therefore not an abstract requirement for us, but a concrete development context. We investigate how, through targeted architectural decisions and domain adaptation, locally operated (on-premise) smaller language models can achieve results that rival the performance of larger cloud models.

All of the applications described above—whether agentic AI, knowledge management, or document processing—are designed to run in sovereign, isolated operating environments.