The policy was signed. The board had seen a slide about human oversight. Then an agent started sending customer email in a tone nobody had approved, on a Friday, and the person who noticed did not know whether she was allowed to turn it off. She Slacked two managers. One was on a plane. The mail kept going.
Oversight that lives in a policy and not on a screen is a story you tell after the fact.
the stop has to be real
Someone on the operating team, not only an admin in another time zone, needs a control that pauses the run. Pausing should not require a deploy. It should not require permission from the person who is proud of the agent. Write down who may press it, and tell them that pressing it is part of the job, not a confession.
Then say what pause means. In-flight actions finish or they do not. Queued actions drop or they wait. Customers mid-thread get a person. If you have not decided, the button is decorative.
If the only way to stop it is to call the person who built it, you do not have an operator.
practice the pause
Once, on a quiet day, have the owner pause it and walk the team through what stopped. The first time should not be the Friday it misbehaves. People do not invent competence during an incident. They use the competence you rehearsed.
Ship the brake with the feature. A tool you cannot interrupt is not finished.