Answers
What authority should an AI agent hold?
The answer
As much as its blast radius justifies and no more. Authority is set by consequence rather than capability: an agent may act alone where a mistake is cheap and reversible, and must escalate where it is neither. The board’s work is defining the thresholds, not approving the individual actions.
The useful discipline is to grade decisions by impact before granting any access at all. Reversible and low-value: the agent acts and the record is reviewed later. Material or irreversible: the agent prepares and a named human resolves. Anything that binds the firm externally — a payment, a contract, a public statement — sits in the second category regardless of how confident the system is, because confidence is not a control.
Two failure modes recur. Granting an agent a human’s credentials, which makes attribution impossible and expands the blast radius to whatever that person could do. And routing every action through approval, which produces a queue nobody services and a workforce that quietly stops being used. Neither is a technology problem; both are the consequence of never having written the thresholds down.