Human-in-the-Loop.

The term is often confused with «somebody looks over it now and then». What is meant is something more precise: the process stops at a defined point and continues only once a decision has been made. That differs from human-on-the-loop, where the process runs through while a person watches and intervenes if needed, and from fully automatic operation with no point of intervention at all. Which level is right does not depend on how mature the technology is. It depends on the damage a mistake would cause, and on whether that mistake can be reversed.

Control points belong wherever something leaves the system boundary. That covers four categories: money moves, communication goes out, data is permanently changed or deleted, and content is published. Add to those the areas carrying particular responsibility, such as health, law, finance and personnel decisions. Everything else runs through without a stop. A workable test: a control point is worth it when the expected damage, probability multiplied by consequential cost, exceeds the effort of checking. For a draft text that is rarely the case; for a payment release it almost always is.

Too many approvals do harm as well. Anyone who has to confirm twenty times a day eventually clicks through without reading, and the control then exists only on paper, yet with documented consent should something go wrong. What works is a small number of deliberately placed control points that show enough context: what exactly is about to happen, with which data, and what the alternative would be. It also helps to build in reversibility rather than multiplying approvals. An action with a fifteen-minute undo window and a clean log is often worth more day to day than one more confirmation up front.

A case from practice. An agent prepares replies to customer enquiries. It researches the internal knowledge base, drafts an answer and saves it as a draft, and nothing is sent automatically. After a few weeks you can measure in which categories the drafts are adopted unchanged; there the control point can be relaxed, while it stays in place for complaints and pricing questions. On the regulatory side, the EU AI Act explicitly requires effective human oversight for high-risk applications in Article 14. Switzerland is not directly bound by it, but the requirement matters for companies serving the EU market.