Human in the loop
Human in the loop means an automated system pauses at defined points for a person to approve, edit or reject what it proposes, rather than either running unattended or being driven manually throughout.
Also called HITL.
Human in the loop describes an arrangement where an automated system does the work but stops at defined points for a person to decide. It sits between full automation, where nobody sees anything until afterwards, and manual operation, where a person drives every step.
The design question is not whether to have a person involved but where. A checkpoint on everything is the same as having no automation, because the bottleneck moves to the review queue and the person starts approving without reading, which is worse than no checkpoint at all: it manufactures the appearance of oversight. A checkpoint on nothing means finding out about a mistake after it has been repeated two hundred times.
The useful rule is to place checkpoints by reversibility and blast radius rather than by importance. Work that is cheap, private and undoable should proceed unattended: a draft, a piece of research, a list of candidates. Work that spends money, publishes to an audience, contacts someone for the first time at scale, or cannot be recalled should wait for a person. The second category is small, which is what keeps the review load survivable and the reviews genuine.
A checkpoint is only real if declining is as easy as approving and the system does not proceed on silence. A prompt that defaults to yes after a delay is automation wearing a review as a costume.
In practice
- An agent that drafts and sends routine replies alone but escalates anything mentioning pricing.
- A video ad campaign built and left paused, so starting it is a deliberate human act.
- A bad checkpoint: forty drafts a day in a queue, approved unread by Friday.
Questions
Where should the human checkpoint go?
On actions that are irreversible, cost money, or reach an audience. Reversibility is the better criterion than importance, because an important decision that can be undone tomorrow is safer to automate than a trivial one that cannot. In practice that means drafting, research and internal work run unattended, while publishing, sending at volume, and spending wait for a person.
Does human in the loop slow everything down?
Only if the checkpoints are in the wrong places. Reviewing everything creates a queue that becomes the constraint and, worse, trains the reviewer to approve without reading. Reviewing the small set of consequential actions costs a few minutes and is the part that makes the rest safe to run alone. If the review load feels heavy, the answer is usually fewer checkpoints on more clearly bounded work, not faster reviewing.
From the blog
Related terms
Last updated