How Much Autonomy Do You Give an Agent?

You are the lead PM on an AI agent product. Your agent reads the user's email, calendar, and CRM, and is designed to help busy founders manage their day. It already drafts reply emails (user reviews them before send) and books calendar holds (auto-confirmed if the slot is free).

Engineering just shipped an internal demo of "Autonomous Mode" — the agent can now send the emails it drafts and confirm meetings without user review, with a 96% accuracy rate (4% of actions are later corrected by users). The CEO loves it. The founding designer is skeptical, arguing that the 4% edge case will ruin the user's reputation.

You have a launch decision to make in 7 days for the v3.0 release. The broader market is shifting toward full autonomy, and investors are asking if you have an "Agentic AI" strategy. But your core user base relies on your product specifically because it never makes an unforced error.

The principle: Autonomy is a product decision, not a capability decision. Just because your agent CAN take an action doesn't mean it should. The most-trusted agents are the ones that ask permission for the things users haven't yet learned to trust them with. Scaling autonomy prematurely breaks the trust bridge before users can cross it.