Always when the action is irreversible, expensive, or public. A good rule of thumb: if a mistake would cost money, reach a customer, change a record or be hard to undo, the agent prepares and the human decides. Everything reversible - drafts, research, summaries, sorting - can run without interruption.
A simple test with three questions
Before letting an agent act on its own, ask: Can this be undone? Who does it reach? What does a mistake cost? Sending a message in your name fails the first two. Booking a non-refundable flight fails the first and third. Drafting a reply passes all three. That is why mature deployments let agents draft everything but send nothing externally without approval.
It is not just prudence - in the EU it is law
The EU AI Act writes this into Article 14: high-risk AI systems must be designed so humans can effectively oversee them, including the ability to intervene, override or interrupt the system. The Act's human-oversight duties are aimed at high-risk uses, but regulators and auditors increasingly treat the pattern as good practice everywhere. Swiss companies selling into the EU, or using AI for decisions about people, should assume oversight is expected.
How vendors frame it
Anthropic's framework for safe agents ties autonomy to demonstrated reliability: earn trust on small, reversible tasks before widening scope. OpenAI's practical guide for building agents recommends starting single-agent, with narrow tools, and adding guardrails - including human approval - before scaling up. The industry consensus: full autonomy is a destination you approach gradually, per action type, not a switch you flip.
What to put in your company policy
One page is enough: which tasks the agent may do alone, which need a named person's OK, and which are never delegated. Then make sure the tool actually enforces it - a rule in a prompt is a wish; a rule in the system is a control.
Sources
Last updated: 20 October 2026.
