A sign-off gate is simple to describe: the agent reaches a consequential action, stops, and waits for a named person to approve. The hard part is the schedule, deciding which actions gate and which run through. Get it wrong in either direction and the system fails quietly.
The two failure modes
Gate everything and you have built an expensive to-do list. The approver sees forty items a day, stops reading them by Thursday, and approval becomes a reflex. The record will show a human in the loop; the human will not actually have been in it. Auditors call this rubber-stamping, and a gate that produces it is worse than no gate, because it manufactures false assurance.
Gate too little and you have delegated judgement you meant to keep. The uncomfortable truth is that most firms cannot say, precisely, which actions in a workflow are consequential, because a person was always there to catch problems informally. Writing the gate schedule forces the question into the open.
How we set the schedule
We start every role with three questions per action. Can it be undone within a business day? Does it leave the firm? Does it move money, obligations or access? An action that leaves the firm or moves any of those three always gates at deployment, whatever the client’s appetite. Reversible internal actions run through, and the decision log catches them for review after the fact rather than before.
That opening schedule is deliberately conservative, and it is written down as a starting point, not a destination.
Earning the wider remit
Remits widen on evidence, not on comfort. After the supervised period, the record shows which gated actions were approved unchanged, every time, for weeks. Those are candidates to run through, and the position description is amended, with sign-off, to say so. The audit trail keeps watching either way.
The reverse also happens, and the system is honest about it. If a class of action starts producing corrections, it goes back behind a gate. Nobody argues with the log.
What the approver sees
A gate is only as good as the screen the approver reads. Ours shows the action, the inputs the agent relied on, the output it wants to release, and the relevant history, on one screen, because an approver who has to go looking for context will stop looking. The gate records what was on that screen at the moment of approval. When someone asks, a year later, what the approver knew, the answer is exact.
If you are evaluating any agentic platform, ask to see the approver’s screen before you ask to see the agent’s output. The second one is easy to make impressive. The first one is where your governance actually lives.