What an AI agent should never do without a human

An AI agent that can book, refund, and update records is useful precisely because it acts. That is also why the question of what it may do on its own deserves more thought than it usually gets.

Our rule is simple: the agent can take any action that is cheap to reverse. Booking a slot, sending a confirmation, updating a contact field, creating a ticket. Anything expensive or impossible to reverse, such as issuing a refund above a threshold, cancelling a contract, or changing payment details, goes to a person with the context attached.

The handoff matters as much as the rule. A good handoff carries the full conversation, the agent's proposed action, and its confidence, so the human can approve in seconds rather than start over. A bad handoff is a transcript dumped into a queue.

We also log every action, whether the agent took it or a person approved it. That log is what lets you widen the agent's permissions later with evidence, instead of hoping.

Confidence thresholds, approval steps, and audit logs are not extras. They are what makes an agent safe enough to put in front of customers, and they are built into every agent we ship.