Start free →

Mandate

Mandates: who answers for each AI use

A mandate is a declared AI system or an agent. Each names the person who answers for it — its accountable person — and an end. Calls are bound to a mandate, so every sealed decision says who answered for it.

Systems

  • Declare a system in Mandates › Systems and agents: its purpose, its AI Act level, the data it uses, its accountable person — a member of the workspace — and its last day, three years ahead at most.
  • Another owner or admin approves it: four eyes. A system is approved only with an accountable person who is a member, an end in the future and its AI Act level; a system classified as a prohibited practice (AI Act art. 5) can only be rejected, and the gateway never serves it.
  • A mandate whose accountable person leaves the workspace is suspended; one past its end is ended. Under the enforce profile neither is served, and every transition is sealed.
  • Each system has its own page: its mandate, its activity — the decisions bound to it — and its proof, with a stop that refuses its calls under every profile.

Binding a call to its mandate

The gateway looks for the mandate in this order: the X-Azoth-Usecase header, the agent, the connector, the API key, then the workspace default. Copy a system’s id from its page, bind a key to it, or make it the workspace default.

curl https://www.azothos.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $AZOTH_API_KEY" \
  -H "X-Azoth-Usecase: 6f1c2e4a-…" \
  -H "Content-Type: application/json" \
  -d '{ "model": "gpt-4o-mini", "messages": [{ "role": "user", "content": "…" }] }'

Agents

  • An agent has an identity — an accountable owner, active or revoked — and a versioned mandate: tools, data scopes, use cases, cost and action caps, tools that need an approval, an expiry. A change is a new version, never an edit.
  • Every tool call is checked against the mandate before it runs. Outside it, the call is blocked under enforce and recorded under observe — sealed either way with the mandate’s version and hash.
  • Irreversible and financial calls always wait for a person: the run pauses and resumes once a distinct person approves.
  • A sub-agent never receives more than its parent: its mandate must fit within the parent’s, when it is written and at every call.

Uses no mandate covers

The calls no mandate covers are grouped from metadata only — key, agent, connector, environment, never content — into proposed cards in Mandates › Unmandated uses. A person validates a card, which becomes a declared system, or ignores it; both decisions are sealed. Repositories where an AI coding tool left a trace appear there too (AI footprint).

Spend per mandate

Mandates › Spend shows, per mandate, the sealed cost and footprint, the tightest budget that applies and the realized savings. A budget applies to the whole workspace or to one mandate; a call bound to a mandate must pass both.

The emergency stop

The workspace, a provider, an agent or a system stops in one click from the header (owners and admins). A stopped scope is refused under every profile, and each refusal is sealed.

What waits for a person

The home page lists what waits for a decision — an approval, a mandate whose accountable person left, a mandate ending soon, a model served without a price, an attack on a public system. It is computed from the tables that hold each item, and each person sees only what they may act on.