The question we are asked most is whether an AI can be trusted to act on its own. It is the wrong question, because it treats autonomy as a property of the agent. It is a property of each action.
Answering a call and booking an appointment is low-consequence and reversible, so it runs unattended. Sending a purchase order commits money to a supplier, so the agent drafts it and a person approves. Messaging a customer who has not visited in three months touches consent and reputation, so it is off until an owner turns it on, and then it runs only in daytime hours with an opt-out in every message.
Three settings per action: do it, propose it, or only tell me. Every action leaves a record of what it saw, what it decided and who approved it. When one goes wrong, you turn that one action down a notch. You do not switch the agent off.
This is the design in our own product, permanently, and it is the design we recommend in every build. It is less impressive in a demo. It is what survives a year in production.
