Answers are grounded in this paper and cite their sources.
AI governance programs begin with policies, principles, committees, and reviews. These manual controls decide whether a system may be built and deployed. They do not show that the system running today is the system that was approved, and a program that rests on them alone tends toward paper compliance: complete documents and unproven operation. This paper proposes a framework of 36 automatic controls that the AI delivery and runtime pipeline enforces or evidences by machine, for predictive, generative, and agentic systems. The controls fall into five phases: fixing the approved baseline and its authority, versioning everything that shapes behavior, proving the release candidate, capturing what actually ran, and stopping, restoring, and retiring. For each control the paper states the artifact it protects, the evidence a machine produces, an audit test, and the limit of that evidence. The framework draws on recent work on governance as code, pipeline-embedded evidence, and runtime enforcement, and on established configuration-management practice. We argue that the extent, timing, and depth of every control should follow the calculated exposure to loss of each AI system in scope and the cost of enforcement, above the minimums the law sets. The result gives chief AI officers and AI architects a checklist for gap analysis, a five-level scale for maturity assessment, and criteria for comparing AI GRC platforms. The controls were derived through a structured synthesis of AI-governance literature, security and configuration-management standards, production-ML guidance, and agentic-AI threat guidance, and were practically validated through negative audit tests, traceability tests, and expert review; the framework has not yet been validated through multi-organization empirical field research.
No takes yet. Share an insight, caveat, or question.
Hernan Huwyler (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: