“Keep a human in the loop” is the most repeated advice in agentic AI, and it rarely says where. So people put themselves in the obvious place: the middle. The first tool produces something, they read it, tidy it, paste it into the next tool, and carry on.
That feels like oversight. Mostly it is labour. You are acting as the cable between two steps, retyping context a workflow could have passed along itself, and the time you meant to save goes on that.
Three places a human genuinely belongs
Before anything irreversible or external. Sending, publishing, paying, deleting, or anything a client will see. These are the actions where a mistake cannot be quietly corrected afterwards, so the workflow should stop and ask. Everything upstream of them - gathering, drafting, checking - can usually run on its own.
Wherever the output is a judgement about a person. Who to hire, who to chase, whose account to flag, what to say to someone having a bad month. A workflow can gather the evidence and draft the options. The decision stays with someone who can be accountable for it.
At the standard. This is the one people miss. The most useful human contribution to a workflow is often not approving its output but writing down what good looks like, so a review step can check against it. In the scout, drafter, reviewer pattern that written standard is what the reviewer reads, and it is yours to write and to keep current.
The test for any step
Ask one question of each step: if this were wrong, who would find out, and how late?
If the answer is “a client, next week”, a human belongs before that step. If the answer is “the next step would catch it within seconds”, the human is probably redundant there. If the answer is “nobody, ever”, that is the dangerous case, and it is a monitoring problem rather than an approval one: a workflow that fails quietly needs a check you run on purpose, not a person skimming everything.
Approving everything is not the safe option
It feels cautious to review every output. In practice it decays fast. By the fortieth draft that was fine, you are clicking approve without reading, which gives you the look of oversight with none of the substance, and a false sense of safety on the day it matters.
Fewer checkpoints, placed where the risk actually is, get read properly. That is worth more than many checkpoints nobody reads.
Start tight, then move the human
On day one, it is reasonable to check more than you will eventually need to. You have no evidence yet about where the workflow goes wrong. As the evidence comes in - the steps that never fail, the ones that fail in the same way every time - move the human out of the first kind and towards the second.
The order matters. Gathering usually earns trust first, drafting later, and judgement about people arguably never. A workflow you trust in stages is one you can hand over in stages, which is far easier than handing over all of it at once.
Where this gets practical
Deciding where the checkpoints go is much easier with a real task in front of you than in the abstract, which is why where to keep a human in the loop is part of the Oxford Agentic Bootcamp curriculum. You bring one real task, build the workflow against it, and decide in the room where a human needs to stay. Cohort 3 runs on 19 November at Worcester College. No coding required.