AI Governance as Admission: Verification and Access under Real Conditions
How verification is translated into access through distributed decision structures - Open Analysis #02
Full PDF versions (EN/DE) are available at the end of the document.
Version Note
This text is a derived version of the original German analysis.
The German version remains the reference text.
1. Starting Point: Verification as a Capability
The AI Act shifts the evaluation logic of AI systems toward verification.
At the center is the question of whether the behavior of an AI system can be traced and assessed under risk conditions. AI systems are evaluated based on whether their outputs are reconstructable and whether intervention remains possible. It is also decisive whether their behavior can be stably assessed under changing conditions.
This shift changes the reference point of evaluation. Assessment is no longer primarily directed at the quality of individual outputs, but at whether an AI system can be reliably assessed under conditions of uncertainty.
This creates a new basis for the use of AI. Access depends on whether behavior can be reliably assessed under verification conditions.
2. Operationalization
This requirement is translated into operational structures. Verification emerges from the interaction of multiple elements. Documentation makes behavior reconstructable. Control mechanisms enable intervention. Conformity requirements link usage to these structures. These elements operate together and embed verification into ongoing operations.
This creates a central condition: AI systems must be able to maintain their own controllability in operation. Whether this structure is viable depends on how validation processes are designed. What is decisive is how independently the verification process is organized from the AI system being assessed.
With increasing integration of AI into both layers, this relationship changes. Validation can become part of the same AI system it is supposed to assess. The formal structure of verification remains intact, but its validity depends on whether the verification is actually independent.
3. Transition to Admission
At this point, the function of verification shifts. It becomes the object of decision-making. Organizations must determine whether an AI system is sufficiently controllable under risk to be deployed.
From this determination, admission emerges. It only operates where AI systems are actually integrated into this decision structure. Systems that are used outside of this structure - for example through shadow use or as embedded functions within existing products - evade this form of structured selection.
4. Admission as a Decision System
In practice, admission is distributed across multiple decision instances. This structure typically includes selection, approval, embedding, and operation. These instances operate in an overlapping and partially contradictory manner. Decisions can be revised, and responsibilities are distributed.
This leads to a fundamental shift. Access is no longer a one-time process, but a state that is continuously evaluated within these decision structures.
This structure only produces effects if decisions can be enforced. Without enforcement, formal decisions remain without operational consequence.
5. Object of Decision
At the center of the decision is not the performance of an AI system. What is decisive is whether uncertainty can be assigned to a specific instance and operationally managed by it.
An AI system remains admitted if it is clear who is responsible for the resulting uncertainty and can handle it in operation. This assignment forms the basis for whether a system can be operated in a responsible way.
If this assignment is missing, formal access becomes unstable. The AI system must then be reassessed within the decision structures.
6. Difference between Formal and Real Usage
A difference regularly emerges between formal admission and real usage. AI systems are used even when verification is incomplete, responsibility remains unclear, or formal admission is missing.
Such constellations can solidify and remain stable over longer periods of time. Formal and real layers therefore frequently diverge.
7. Pressure on the Difference
This difference rarely remains stable in practice. Selection emerges under pressure, not through automatic alignment. This pressure is potentially permanent, but often remains latent and is unevenly distributed. It becomes effective when concrete triggers occur, such as incidents, audits, scaling effects, or regulatory changes.
In such situations, the difference becomes visible and must be addressed. Pressure is then translated into decision pressure within the existing structures.
8. Decision Responses
Under this pressure, decision instances face three fundamental options.
They can carry the difference and continue operating the AI system despite existing deviations. Restriction occurs through adjustments, additional controls, or reduced use. Ultimately, access can be terminated.
Restriction is not unambiguous. It can mean that actual problems are reduced, or that only formal adjustments are made without changing the behavior of the AI system. In both cases, the AI system can initially remain admitted, but with different levels of stability under pressure.
Additionally, pressure can be absorbed. This occurs, for example, through documentation, formal risk assignment, or the shifting of responsibility, without changing the actual system behavior.
9. Selection Mechanism
Selection emerges when decision instances are forced to address this difference. An AI system remains admitted as long as uncertainty can be assigned to an instance and managed operationally. It falls out when this assignment disappears or no longer holds.
10. Conclusion
Verification is no longer solely a matter of formal requirements. It becomes the operational condition of access.
Admission defines the order in which uncertainty is assigned and decisions are made under pressure.
Ralf Brentführer
May 2026
Download full paper (PDF):
