C4.7 AI controls

NIST AI RMF and profiles [21]; PMI AI branch [3]

The seventh component requires that every AI system in use be mapped, measured and managed against the seven characteristics of trustworthy AI, that each carry a current and a target profile in the subcategories of the AI risk management framework, and, as the framework's own addition, that each carry decision rights set by appetite: the system may act reversibly under monitoring, may prepare an action for a human to approve, or may only inform. The functions, the characteristics and the profile pair come from the AI risk management framework and its generative-AI profile [21]; the place of AI in the source taxonomy, as a branch governed by appetite, comes from the project tradition's recent extension of its breakdown structure [3].

The component reconciles a descriptive source with a prescriptive need. The AI framework names what should be measured and suggests actions; it does not say who may let a system act, or on what. The decision rights do, and they are deliberately expressed in the appetite vocabulary of L1 so that the board's statement about how much risk it wants in a category translates directly into what a system in that category is permitted to do without a person. An AI system under this component is a control, with a level in the hierarchy (usually engineering, since it acts without a human once deployed, which is precisely why its decision rights must be narrow), a mode, a human-dependence flag that is low by construction and a design-gate record like any other change. The three levels of testing the source family describes, in the model, under adversarial review and in the field, are the evidence that the current profile is what it claims.