Decide which actions need approval
I would require approval where an action reaches another person or is hard to undo. Too many stops slow routine work and train you to approve on autopilot; too few let high-impact changes happen before you can review them.
Use two axes: how hard an action is to undo, and how far it reaches beyond the agent’s workspace.
Two axes that predict risk
Reversibility. Can this be undone cleanly if it turns out to be wrong? Deleting a draft you haven’t sent is fully reversible. Sending an email is not. Booking a refundable hotel room is mostly reversible. Paying a non-refundable deposit is not.
External reach. Does this action stay inside the agent’s own workspace, or does it touch something outside: another person’s inbox, your bank account, a public website, a legal agreement? Drafting a document in a private workspace has no external reach. Publishing that same document does.
Let reversible work inside the workspace run. Gate actions that are hard to undo or reach outside it.
A decision matrix for common actions
| Action | Reversible? | External reach? | Needs approval? |
|---|---|---|---|
| Reading email, calendar, documents | Yes | No | No |
| Researching and comparing options | Yes | No | No |
| Drafting a reply or document | Yes | No | No |
| Organizing files, tagging, sorting | Mostly | No | No, unless deleting |
| Sending an email or message | No | Yes | Yes |
| Booking travel, tables, appointments | Partially | Yes | Yes |
| Paying or entering payment details | No | Yes | Yes |
| Deleting or overwriting records | No | Sometimes | Yes |
| Publishing or posting publicly | No | Yes | Yes |
| Accepting contracts or invitations | No | Yes | Yes |
What a good approval request looks like
A gate is only as useful as the decision card behind it. Compare two versions of the same moment. Weak: “Send email? Y/N.” Useful: “Send 3 follow-up emails to Acme, Brackenwold Studio and Thornquist Ventures? Drafts are written in your tone, referencing their open invoices from last month. Nothing has been sent yet.” It names the action, reason and prepared work. You can decide without redoing the task.
A useful approval request should let you do one of three things without extra steps: approve it as written, edit it and then approve, or hold it. If editing means starting over, the gate is adding friction without adding much safety.
Writing your own approval rules
The default matrix above covers most cases, but real work has exceptions worth setting explicitly rather than working around each time.
You might tighten a rule beyond the default: requiring approval for any message to a specific high-stakes client, even routine ones, because the relationship is sensitive enough that you want to see every word before it goes out. You might loosen a rule for a specific, narrow, repeated case: pre-approving the same weekly invoice reminder to the same list, since you’ve reviewed the pattern enough times to trust it, while keeping approval required for anything outside that exact pattern. Scope each exception to a task or pattern. Blanket toggles can add friction or let too much through.
What happens when you’re not there to respond
At 2 a.m., an unanswered gate should wait. The agent continues work that doesn’t depend on your decision and flags the held action for your return. A timeout must not become permission.
Where this fits with OperatorNest
OperatorNest applies this matrix by default: approvals are required before anything that sends, pays, books, deletes or publishes, presented as a specific decision card, and you can tighten or loosen rules per task. Nothing is sent or spent while a decision waits. The default matrix gives you a starting point; set tighter or looser rules for each task to match your risk tolerance. See what an approval gate is for the underlying mechanism this post builds on, or use the AI agent approval policy builder to write down those rules.