Delegated authority
The written mandate that says exactly what a system may do without asking, and what it must stop and escalate.
Category
What keeps it honest
Related terms
Delegated authority is the institution's answer to a specific question: under what conditions may this system act on its own? It is written before the system runs, it is inspectable, and it lives outside the model.
What a mandate has to state
The action that may be taken. The conditions under which it may be taken — thresholds, limits, required evidence, permitted tools. The person or committee delegating it. The stop conditions that end delegation. And what happens instead when any condition fails.
A mandate that says “the system handles routine cases” is not a mandate. “Routine” has to be a set of checks that either pass or do not.
Why it is external
An authority boundary that lives in a prompt or in model weights cannot be audited, cannot be changed without a deployment, and cannot be shown to a regulator as a control. Written policy, identity, permissions, approval limits, and stop conditions stay in the deterministic layer for exactly those reasons.
Expansion is not self-service
A system cannot widen its own mandate. Outcomes may suggest that a threshold is too tight or a case class could be delegated; that produces a proposal, evaluated and approved by the same authority that granted the original mandate, versioned and reversible. An agent that could grant itself a data source, a tool, a permission, or an approval bypass would not have a mandate at all.
Often confused with
How it differs from similar terms.
| delegated authority | Compared with | What's different |
|---|---|---|
| Delegated authority | Model confidence | Confidence is one input to a check. Authority is the written mandate that decides whether acting is permitted at all. |
| Delegated authority | Autonomy | Delegation is bounded and revocable by a named owner. Autonomy implies a boundary nobody wrote. |
Want these terms worked through on one of your own decisions? Start with a Decision Blueprint