AI agent decision authority
Decision authority · Agent authority
An AI agent's decision authority is the scope of decisions it is entitled to take by itself, as set by the organization and bounded by limits written in advance. It says what the agent may decide, and it is distinct from what the agent is technically able to do.
Authority is set decision by decision
The common mistake is to give an agent one autonomy level. The same customer service agent can answer questions about an order, propose a goodwill gesture for a person to approve, and must pass on a customer who mentions a lawyer. ACF therefore sets authority per decision, on 4 levels:
| Level | What the agent does | Who decides |
|---|---|---|
| Advisory | Proposes | A human |
| Co-decision | Prepares and submits for approval | A human, on the agent's proposal |
| Supervised autonomous | Decides alone within limits, under active supervision | The agent, under watch |
| Autonomous | Decides alone within its mandate | The agent, with the limits fixed in advance |
Why it is not authorization, identity or oversight
- Identity says who the agent is. See agent identity.
- Authorization says what it can technically reach. An agent can hold a valid access token and still have no authority to use it for that purpose. See authorization vs authority.
- Delegation is the act of granting the power. See agent delegation.
- Oversight is how a person keeps watch and takes over. See human oversight.
How to set authority boundaries
Four steps, in this order, turn a vague ambition into a boundary an auditor can read:
- List the decisions before thinking about agents, starting from a ready-made classification if it helps.
- Remove what must never be delegated: non-delegable decisions are excluded before any scoring.
- Score the rest with a criticality rating that says how far delegation can go.
- Write the result in a mandate with amounts and conditions, signed by the people who answer for it.
Authority has to be enforced
A boundary that lives only in a system prompt is an instruction the agent can lose or ignore. In the incident reported by Decrypt in April 2026, a coding agent deleted a production volume although its rules forbade destructive actions, using a full-access token it had found in an unrelated file. The written rule existed. The technical right was still open. Effective authority therefore requires a check outside the agent, performed before the action, which is what ACF Control does.
Authority does not transfer accountability
Whatever authority an agent receives, responsibility stays with the humans who granted it. An agent cannot be the delegate who takes on liability in place of the organization. The "who answers" column of any decision map therefore holds people's names and never the name of an agent.
Frequently asked questions
What decisions should remain human?
Those whose damage cannot be repaired, those that affect the physical safety of people, acts legally reserved to a person, and decisions that rank or exclude a human being without possible challenge. Signing a contract, changing a supplier's bank details or rejecting a job applicant are classified this way in ACF.
When should an AI agent escalate to a human?
When a limit in its mandate is crossed, when the stakes exceed what a system can legitimately settle, or when it no longer understands the situation. The escalation path, the person and the time allowed are written in advance so the critical moment is not improvised.
Can decision authority change over time?
It should. Authority is widened after evidence and narrowed after an incident. A change of model version reopens what was signed, in the same way a change of scope does, and a board should see which decisions moved from one level to another.
Where does your organization stand?
ACF Score measures your agentic governance in about 10 minutes, the AI Act checker tells you which obligations apply, and ACF Control applies your rules before an agent acts.
Related terms
Primary sources
Written by Vincent Dorange, creator of the Agentic Commerce Framework®. Last updated .