A Risk Appetite Where the Act Happens: A working note on the mandate under which an AI agent acts

Regulated firms write delegated authorities for the people who act for them, and the AI agents they now run usually act under entitlements instead. This note puts the join between the two at the centre of the control plane. That join is the mandate: a versioned authority for one class of acts that says what the agent may do, on what evidence, that the control plane may refuse, and who takes the act when it does. The mandate is placed on the Licence to Act map, and its versions change on evidence: narrowing is automatic, down a ladder the board risk committee has signed in advance, on outcomes adjudicated outside the agent; widening takes a signature from outside the escalation path. The precedents for each half of that rule are cited as precedents, and the claim is stated at its true size: a per-act, refusable permission with a quantitative appetite for the worst outcomes, which narrows on evidence and widens on signature, at the scope of one class of acts. The second part gives the mandate its economic form in Cordier’s per-task value equation, with the appetite limit as its constraint written where the act happens and the price of refusal added. A paper follows from this note. Publication status: This deposit is a preliminary working note, not a peer-reviewed article. It develops a conceptual and economic framework and identifies questions for subsequent formal and empirical investigation.