AI Field Notes
How much authority should an AI agent have?
By Ed Fassio ·
An AI agent should have only the authority explicitly granted for a particular decision, within limits a named human is prepared to answer for. Its ability to complete a task does not, by itself, establish permission to make the decisions inside that task.
Start with the decision
A broad label such as “autonomous assistant” tells you little about what has actually been authorized. Sorting a request, recommending a response, sending that response and changing an account are different decisions. They may need different permissions, evidence and review.
Ed Fassio’s Decision-Scoped Delegation treats authority as a specific grant. That makes it possible to inspect what was delegated, rather than assume that a capable agent should receive permission to do everything its tools allow.
Put the delegation in writing
The useful questions are practical:
- What decision may the agent make, and toward what outcome?
- In which situations does that grant apply?
- What limits, exceptions and stopping conditions constrain it?
- What evidence must exist before it acts?
- Who can review the decision, refuse it or stop the work?
- When will the grant be reviewed or withdrawn?
A named sponsor needs more than responsibility on an organizational chart. That person needs enough visibility, time and standing to exercise it. An approval queue no one can meaningfully review provides little protection.
The Delegation Ladder (https://agentic-bytebrain.org/) gives teams a shared language for these choices. It applies to individual decisions, rather than awarding an entire agent a permanent autonomy badge. The canonical framework explains the full ladder, including work that stays human and the human responsibility for authorizing the boundaries themselves.
A hypothetical example
Imagine a support team introducing an assistant for incoming email. The team first allows it to suggest a topic and draft a reply. A person checks the recommendation and decides what to send.
The team later considers allowing automatic routing for a narrowly defined set of routine requests. Before widening authority, it establishes acceptable outcomes, excluded cases, evidence to retain and a way to interrupt incorrect routing. It tests those conditions and assigns a sponsor to review the results.
That change does not also authorize refunds, customer promises or new routing rules. Each would require its own decision about authority. If the assistant proposes a better limit, the proposal still goes to the person empowered to authorize that limit.
Authority needs maintenance
A successful pilot establishes evidence under particular conditions. Changes in the model, connected tools, workflow or responsible people can change those conditions. Review the grant when the work changes, and reduce authority when the evidence no longer supports it.
Keeping a decision human can be a deliberate operating choice. Higher autonomy is not automatically the desired endpoint. The appropriate grant depends on the consequences, reversibility, evidence and organization’s ability to supervise the work.
Written grants also need implementation. Permissions and action checks must reflect the boundaries; a document alone cannot prevent an agent from exceeding them. No ladder eliminates uncertainty or guarantees safe behavior. It makes the delegation explicit enough to inspect, test and change.
Sources and further reading
This guide draws on Ed Fassio’s work on delegated authority and human accountability.