Human on the loop

The difference in one line

In the loop: the system asks permission for every action. On the loop: the system acts and a person watches, with the power to stop it.

The first is safer per action and does not scale, because the human becomes the bottleneck. The second scales and requires you to have thought in advance about what an intervention looks like.

Why on the loop is usually the right choice

Approving every action sounds responsible and tends to decay in practice. When a person has to click approve four hundred times a day, they stop reading. That is worse than no supervision, because the record now says a human checked it.

Supervision that is actually exercised beats supervision that only exists on paper.

What it requires

On the loop only works if three things are arranged:

• Visibility. The supervisor can see what the system is doing, in terms they understand, not in logs
• A stop. There is a way to intervene that works within the time the damage takes to occur
• A named person. Supervision without an owner is not supervision

Without these, on the loop is a description of an unsupervised system.

Where the boundary sits

Some decisions do not belong on the loop at all. Anything irreversible, anything with direct consequences for a person, and anything you would not be able to explain afterwards belongs in the loop, or nowhere.

Why this matters for control and risk

The systems stay yours and control stays with you. That is the practical point: a person remains in a position to overrule the system, which keeps operational risk manageable and makes the result defensible to a board and an auditor. It is also what an AI policy has to write down per decision type.