The pattern's value is precision: instead of choosing between watching everything (which does not scale) and trusting everything (which is not governance), you name the exact moments where human judgment is required and make the system stop there.
Implementation quality varies enormously. Asking the model to "wait for approval" is etiquette, not a mechanism. A real implementation makes the pause structural: the process cannot continue until the approval is given.
The pattern also lets you choose your checkpoints. Add them where a person, a metric or a report should decide. Leave them out where the work can run by itself. A checkpoint is part of the process, so moving one is an edit to the process, not a change to the agents.
It helps to separate human-in-the-loop from its neighbours. Human-on-the-loop means a person monitors a running system and can step in, but the system does not wait. Human-out-of-the-loop means full automation. Human-in-the-loop is the only one of the three where the system is designed to stop and wait, which is why it suits irreversible actions: shipping code, moving money, contacting a customer. The choice is less about trust in the model than about the cost of being wrong: the more expensive and the less reversible the action, the stronger the case for a stop.