The human in the loop: put one in, then earn the right to remove it

The first version of any AI feature that touches something consequential should propose, not act. A person reviews and approves. It is slower, and it is what makes the product shippable at all in a business where mistakes are expensive. The review step also does work beyond safety. Every approval, edit and rejection is a labelled example of what good looks like, which becomes the eval set and the basis for improving the prompt. A product with a review step is learning; one that automated immediately is not. Removing it is a decision made with numbers. Measure the approval rate — how often the human accepts without change — and the nature of the edits. When a category of task is approved unchanged nearly always over a meaningful volume, that category is a candidate for automation. Others may never be. Three patterns for stepping down rather than jumping. Automate by confidence, where only high-confidence outputs pass through. Automate by category, where a well-behaved subset goes automatic and the rest stays reviewed. Or automate with a delay, where the action is queued and executed unless a person intervenes within a window — which keeps oversight without keeping a bottleneck. Whatever is automated needs monitoring that would catch it degrading: a sample reviewed on a schedule, and an alert when the pattern of outputs shifts. And say it clearly to the customer. "A person checks this" is a feature to a buyer who has been burned. Removing it silently is how trust is lost. This article is general information, not legal, tax or financial advice. Rules change and every deal is different — check your own case with a licensed professional. Alberto Zaltzberg — Adonait · adonait.com