- Approvals at every step tire people and teach them to click without reading, and the lack of one at an irreversible step costs the most.
- Place them by effect: reading without asking, writing with undo, deleting and sending only with consent.
- AI rules also speak of human oversight, but only for specific uses.
“Human in the loop” sounds like a brake: something that slows automation and reminds you that a machine cannot be fully trusted. In this essay I argue that it is closer to a map. It shows at which points of a process a human decision is needed and at which it can safely be left to the machine. Drawn well it saves time; drawn badly it eats it.
Two mistakes, one result
The first mistake is asking for consent at every step. After the fiftieth prompt a person stops reading and clicks “yes”, so the approval exists but protects nothing. The second mistake is not asking where the effect is irreversible: sending a message to customers, deleting files, a payment. Both end the same way: the control is only apparent. In the text Building effective agents Anthropic notes that agents can pause to get feedback from a human at checkpoints or when they meet a blocker. The key word is “checkpoints”, not “every step”.
How I draw the map
I divide actions by effect. Reading data without changing anything can run without asking. A write that can be undone, such as a change in code on a separate branch, can run with the possibility of restoring. Deleting, sending outside, payments and permission changes require a person’s consent that shows the specific content of that action, not a generic “continue?”. Tools for coding agents have this built in: the Claude Code permissions documentation describes rules that allow, ask or deny for individual tools and commands. You will find a detailed checklist in the article An approvals checklist.
What the law says and what it does not
In debates the argument often appears that the rules require a human in the loop. That is a shortcut. The EU regulation on artificial intelligence, available in the Official Journal, provides for human oversight as a requirement for high-risk systems, not for every use of a model. If you run a small team and use AI for draft texts, that does not oblige you to approve every line; it does mean you need to know which category your use falls into. I describe the basics in the article The AI Act in brief.
Why I see this as a matter of focus
My rule is that technology should give time back. Well-placed approvals are exactly that. Once I think through which actions are irreversible, I can work calmly, because I know the system will stop at them. A bad map means I either sit at the screen clicking, or leave the agent unsupervised and hope for the best.
What to do today
Take one process you want to automate and write its steps in three columns: I read, I write with undo, irreversible. A person’s consent is needed only in the third. Give the other steps a test and a log so you can check them afterwards. More about when an agent should ask in the article Human in the loop: when to ask.