Every automation story has a chapter where the system confidently does the wrong thing at scale. The size of that chapter is decided by where the approval checkpoints sit.
Checkpoints get a bad reputation because badly placed ones feel like doing the work twice. Well placed ones feel like nothing at all. The difference is placement.
Put approvals where mistakes are expensive, not where work is tedious
Nobody needs to approve a draft. Drafts are free; a wrong draft costs a read. Sending, deleting, paying, and publishing are the expensive verbs. Put the gate in front of those and leave the middle of the pipeline alone.
Make the checkpoint one action
An approval that requires navigating to another tool, re-reading context, and reconstructing what the assistant was doing is not a checkpoint. It is a punishment. The approval prompt should contain everything needed to decide: what will happen, to what, and why.
Earn the right to remove checkpoints
Approval requirements are not permanent. After a behavior has been correct twenty times, demote it. Over time the checkpoints migrate to the genuinely risky edges, which is where a human is worth the most and costs the least.
The goal was never full autonomy. It is a workflow where humans hold the decisions that deserve humans, and the assistant holds everything else.