Reversibility is the real dividing line
The useful way to sort any automated action is not by how hard it is or how confident the system is. It is by what it takes to undo.
Drafting a quote is free to redo. Sending that quote to a customer is not, because the customer has now seen a number. Charging their card is worse again, because undoing it involves a refund, a fee and a conversation. The further an action reaches outside your own records, the more a mistake costs and the less a confidence score is worth.
The loop: propose, approve, execute, log
The shape that makes this workable is four steps rather than two.
- Propose. The assistant does the work and produces the artefact: the quote, the reply, the schedule change.
- Approve. A person reads it and says yes, edits it, or sends it back. Nothing external has happened yet.
- Execute. Only now does the thing leave the building.
- Log. What was proposed, what was approved, what changed and what happened is written down and stays readable.
The log is the part people underrate
Approval without a record is a gate you walk through and forget. The log is what turns a series of individual decisions into a basis for widening the gate later.
After forty approvals of the same kind of follow-up message, with no edits, you have evidence rather than a feeling. That is the point at which letting it run automatically is a reasoned decision instead of a leap. Without the record you are still guessing after the fortieth one.
Where the checkpoint should not be
An approval on everything is not safety, it is a second job, and a queue nobody can face is worse than no queue because it gets approved in bulk without reading. That is the failure mode to design against.
Internal and reversible work should simply run: drafting, research, sorting, preparing, summarising, anything that produces something for you to look at rather than something a customer receives.
- Auto-run: drafts, classification, research, internal summaries, scheduling proposals.
- Ask first: outbound messages, published content, anything that moves money, anything a customer sees.
- Never blanket-approve: an action whose undo involves apologising to someone.
Make the approval cheap to give
If the checkpoint is going to hold, it has to be fast. That means the queue shows the actual artefact rather than a description of it, edits happen in place instead of by rejecting and starting again, and similar items sit together so a run of eight follow-ups is one decision rather than eight.
The goal is a gate that takes seconds to pass when the work is right and makes it obvious when it is not. Anything slower gets bypassed, and a bypassed gate protects nobody.
Sort actions by what it costs to undo them, put the checkpoint only where that cost is real, and keep the record. Trust earned that way widens on evidence instead of on nerve.