The Override Without a Name

We design agents to act. We rarely design the stop.

The dashboard is green. Throughput is up. The team has learned not to interrupt a process that “usually works.” Then something is wrong in a way the metrics don't catch, a tone in a customer reply, a hiring shortlist that looks polished and unfair, a spend path no one meant to open. Someone in the room feels it. No one can say, without sounding difficult, I'm allowed to halt this.

That silence is a missing name.

Autonomy isn't the absence of a human

Autonomy, in the marketing sense, means the system proceeds without asking. In a leadership sense, it means the system proceeds until a person who can be asked why decides otherwise. Those aren't the same design.

If the pause lives only in a policy PDF, it doesn't live. If the pause lives in “escalate to the on-call” and the on-call is a queue with no human face this week, it doesn't live. If the pause requires three approvals and the damage is already public, it didn't live in time.

An override without a name is unclaimed authority. The model didn't seize it. The organization left it on the floor.

What a real override looks like

It's boring on purpose.

• A role, not a vibe: “Incident owner for Agent X.”
• A person, not a distribution list: a name that answers on a weekday.
• A channel that's already open: not a new ticket type invented under pressure.
• A sentence everyone has heard: “I'm pausing Agent X. Reason follows in five minutes.”

The point isn't to stop systems often. The point is that the room knows the stop is legitimate before the room needs it.

Teams that never practice the pause discover, under load, that speaking up feels like career risk. That's SCARF in plain clothes: status threat dressed as professionalism. The Status Threat Nobody Named is what happens when the model speaks first. This is what happens when no one is authorized to speak last.

The human overhead you already pay

You're already paying for the absence of a named override. You pay in near-misses that never enter the log. You pay in reviewers who soft-approve because the model’s draft was fluent. You pay in the person who leaves six months later and can't explain why the system felt unsafe to challenge.

Green metrics and closed mouths can share a building. They often do. The catch that never became a ticket is the same pattern as The Near Miss Nobody Logged: the human still subsidized the metric, and the organization booked it as success.

This week

Open the runbook for one live agent. Add one sentence:

Override: [Role] ([Name]) [Channel]. Authorized to pause without prior committee approval. Reason posted within fifteen minutes.

Read that sentence in the next stand-up. Watch who looks relieved. That's data.

The person remains the one who can be asked why. If you can't name who may stop the system, you have already answered who is in charge of it and it's not a human you can hold.

The five questions in the Ethical AI Leadership Decision Toolkit are built for this kind of page. Field notes continue in EI Leadership Insights.