AI needs structure: Why reliable system states are essential

“As a professor of cyber-physical systems, I work on the analysis and verification of complex software systems deployed in real-world environments shaped by human processes. A central research interest is to determine under which conditions algorithmic systems can produce reliable, traceable, and verifiable statements.

Current AI systems, particularly large language models, are based on probabilistic inference. They are highly capable in the analysis, classification, and interpretation of data, but they require a sufficiently consistent and structured system state. They do not generate formal truth themselves; rather, they operate on probabilities and approximations.

In real-world software systems—especially those with historically evolved processes and data originating from heterogeneous sources—these prerequisites are often not met. Data is fragmented, incomplete, or contradictory, and semantic definitions are not explicitly formalised but embedded in organisational context. Establishing a consistent and verifiable system state in such cases is not purely a learning problem, but also an integration, modeling, and validation problem.
From a systems-theoretical perspective, this leads to a structural limitation: AI can support data processing and pre-structuring, but under real-world conditions it cannot yet autonomously guarantee the full consistency, traceability, and formal correctness of an overall system.
Particularly in application domains where results must be explainable, auditable, and accountable ex post, the establishment of a reliable system state remains an independent task. AI can support this process, but it does not replace it.

This assessment does not constitute a judgment on individual products or future developments, but rather reflects fundamental differences between probabilistic inference and formal logic, as they have long been studied in research.”

Prof. Dr. Matthias Althoff (TUM School of Computation, Information and Technology - Cyber-Physical Systems)

Derivation and contextualisation of Prof. Dr. Althoff’s statements in the context of innoscripta and Clusterix

The preceding statements by Prof. Dr. Althoff make it clear that the performance of AI systems cannot be assessed in isolation at the model level, but only in relation to the nature of real software systems and their underlying data. In particular, the distinction between probabilistic inference and the establishment of a consistent, traceable system state is central to understanding the positioning of innoscripta.

From this perspective, it follows first that while AI is capable of efficiently analysing and processing information, it cannot autonomously generate a reliable system state when input data is incomplete, fragmented, or contradictory. This data reality is precisely what Prof. Althoff describes as structurally typical for complex, real-world software systems. It is not a temporary deficiency, but a consequence of human-shaped processes.

For application domains such as tax, funding, and compliance, this leads to an additional implication: results must not only be plausible, but correct, traceable, and accountable. While AI can provide probabilistic suggestions, responsibility for decisions necessarily remains with companies and their governing bodies. This responsibility is non-delegable. AI can assist, but it can neither assume responsibility nor bear liability. This creates a need for systems that make decision bases transparent, ensure audit trails over time, and enable a clear attribution of responsibility.

Another aspect that follows directly from Prof. Althoff’s statements concerns the handling of implicit knowledge and interpretative flexibility. A substantial portion of relevant decision logic is neither fully formalised nor permanently codifiable. Regulatory criteria evolve over time, interpretations develop through practice and audits, and many distinctions are based on experiential knowledge. These elements are context-dependent and emerge from real-world application; they cannot be fully derived from texts or rule sets and are therefore only partially accessible to AI systems.

Against this background, the role of Clusterix becomes clear. The platform is not designed as feature software that automates isolated functions, but as a data infrastructure that establishes a reliable system state in the first place. Clusterix consolidates real data from heterogeneous sources, documents assumptions and decisions, and maps relationships between projects, time, and costs in a traceable manner. Human expert judgment is not an external element, but an integral part of the system and a prerequisite for robust results.

The structural limitation of AI described by Prof. Althoff also explains why, with increasing use of the platform, a proprietary data base emerges that cannot be replicated in the short term. This includes not only structured data, but also the results of real audits, accepted and rejected argumentation patterns, and context-specific experiential knowledge. Each additional use case expands this base, increases the consistency of the overall system, and cumulatively strengthens its explanatory power.

From a systems perspective, a further implication arises: the impact of Clusterix is not limited to software functionality, but unfolds through operational enforcement at the customer level. Processes are bindingly anchored, parallel shadow solutions are reduced, and Clusterix is established as the de facto foundation for tax and audit-related work. A potential replacement would therefore not merely be a technological alternative, but would require a comprehensive organizational, liability-related, and governance-related transformation.

In summary, Prof. Althoff’s statements imply that AI reaches structural limits wherever reliable system states, responsibility, and implicit knowledge are required. Clusterix addresses precisely this layer. This does not result in a separation from AI, but rather in a positioning as the infrastructural prerequisite for its meaningful and responsible use.
In addition, it must be considered that research and development processes are structurally characterized by uncertainty, iteration, and revision. Decisions are made under incomplete information, assumptions change, and documentation emerges over time across different systems and stakeholders. These human factors are not an efficiency problem, but a constitutive element of real knowledge work and inevitably lead to data gaps, ambiguities, and inconsistencies.
The use of AI does not fundamentally change this reality. AI can process or pre-structure information more quickly, but it does not eliminate the causes of incomplete or contradictory data as long as humans generate, evaluate, and take responsibility for that data.

With increasing automation, the need to interpret, validate, and integrate results into a consistent overall context actually increases.
It follows that the need for a platform that establishes and stabilises a reliable system state over time does not diminish, but persists in the long term. Clusterix addresses precisely this structural constant by systemically mapping and making transparent human-generated reality, rather than idealising or simplifying it.
Eliminating this necessity would require that research and development processes be carried out fully and autonomously by AI or robotic systems—including goal definition, evaluation, and documentation. Such a scenario is currently neither technically nor organisationally foreseeable.
Until then, the ability to transform real, human-shaped data into a consistent and verifiable system state remains an indispensable prerequisite for reliable decision-making and the meaningful use of AI.