Human-in-the-Loop

Problem

An AI agent can prepare a €50 refund automatically. Should it also independently approve a €5 million transfer? Some actions have consequences too important for probabilistic automation alone.

Solution — Human Approval

Let automation prepare the action but require human approval at defined boundaries.

Agent → proposes action
             ↓
          Human
         ↙     ↘
     Approve   Reject

Core idea: human-in-the-loop keeps consequential decisions under human control while automating surrounding work.