Three tests for whether a step needs a person
Put a gate on a step if it fails any of three tests. Can it be undone? Does it stay inside the business? Is a mistake cheap? Sending a refund, emailing a customer and changing a price each fail at least one, so each waits for a person.
Steps that pass all three can run on their own: classifying a ticket, drafting a reply for review, updating an internal record, preparing a report. This is where most of the time saving comes from, and none of it leaves the building without someone seeing it.
Make approval one click, with the evidence attached
A gate that takes five minutes to clear will be skipped or resented. The person approving should see what the AI employee proposes, why, and the source it worked from, in one place, with approve, edit and reject as the only choices.
Design the queue for the person, not the system. Group similar items, show the oldest first, and send anything waiting too long to a second approver. If the queue grows faster than it clears, that is a signal to change the rules, not to remove the gate.
Loosen gates with evidence, never by default
Start strict. After a period of live work you will have a log showing how often the approver changed what was proposed. Where a type of action has been approved unchanged again and again, you have a reasoned case for letting it run on its own within limits.
Where approvers keep editing, the gate is doing its job and the rules or the source content need work. Either way the decision rests on the log, which is the reason every action should be recorded from day one.
Use this before you buy or build.
- Every step is tested for reversibility, exposure outside the business and cost of error.
- Approvers see the proposal, the reason and the source in one view.
- Items waiting too long are routed to a second approver.
- Gates are only relaxed when the log shows repeated unchanged approvals.
Gate anything irreversible, external or expensive. Make approving fast, and let the log decide when a gate can come off.


