AuditMAI: Towards An Infrastructure for Continuous AI Auditing

Current AI auditing is fragmented and inefficient: Existing AI auditing tools operate in isolation without integration features, resulting in manual, high-effort, and typically one-time audits rather than systematic ongoing assessments of AI system performance and safety.

Continuous auditing offers a better model: The paper proposes adopting continuous AI auditing—similar to practices already established in finance—to enable regular, systematic assessments of AI systems rather than relying on sporadic, point-in-time evaluations.

AI auditability needs to become a core design feature: The research identifies that auditability is essential for responsible AI but is not yet a standard consideration when building AI applications, suggesting organizations need to plan for auditing from the start rather than retrofitting it later.

A unified infrastructure is needed: AuditMAI provides a blueprint for an integrated auditing infrastructure that could standardize how AI systems are monitored and assessed, drawing on real-world requirements from industrial use cases to ensure practical applicability.

The field lacks mature methods: The paper highlights that while continuous auditing is a promising direction for responsible AI, the technical methods and standards to support it are still underdeveloped, indicating this is an emerging area requiring further development.