System Card

A system card is a comprehensive documentation artifact that describes an entire AI system rather than individual models within it. System cards capture the complete picture of how an AI application functions, including its architecture, component models, data flows, integration points, human touchpoints, and operational context. As AI systems increasingly combine multiple models with business logic, external services, and human workflows, system cards provide the holistic view that model cards alone cannot offer.

The EU AI Act regulates AI systems, not isolated models. While model cards document individual machine learning models, the regulation's requirements for technical documentation, transparency, and risk assessment apply to the complete system that produces outputs or makes decisions. System cards bridge this gap by documenting how models are orchestrated, what data flows through the system, where human oversight occurs, and how the system integrates with organizational processes. Annex IV technical documentation requirements align closely with system card content. The regulation requires a general description of the AI system including its intended purpose, how it interacts with hardware and software, the versions of relevant software, forms of input and output, and the computational resources required. This system-level perspective is essential for conformity assessment and for enabling deployers to understand what they are implementing. For modern AI applications built on foundation models with retrieval augmentation, tool use, and multi-step reasoning, system cards are especially important. The behavior of such systems depends heavily on their orchestration logic, context management, and integration design, aspects invisible in model cards but critical for risk assessment and oversight.

Organizations should create system cards for every AI system they develop or deploy, distinct from and complementing model cards for underlying components. A comprehensive system card includes several key sections: system overview describing purpose, users, and use context; architecture documentation showing components and their interactions; data flow diagrams illustrating how information moves through the system; model inventory listing all AI components with references to their model cards; human touchpoints identifying where humans interact with or oversee the system; integration points documenting connections to external systems and data sources; and operational parameters including performance characteristics, limitations, and failure modes.

System cards should be treated as living documents, updated whenever significant changes occur to architecture, models, integrations, or operational context. Version control with clear change history supports both internal governance and regulatory requirements for maintaining current documentation. The relationship between model cards, system cards, and Annex IV documentation should be architected intentionally. Model cards feed into system cards, which in turn provide much of the content required for Annex IV technical documentation. Organizations benefit from designing this documentation hierarchy upfront, establishing templates and processes that build efficiently from component documentation through system documentation to compliance artifacts.