Human-in-the-loop, often shortened to HITL, describes any system where automation handles the heavy lifting but a person remains the final decision-maker. The machine gathers information, makes a prediction, or drafts an output; the human then reviews it and chooses whether to accept, edit, or reject it. The loop closes only when a person acts, so the automation never runs entirely on its own.
Where the human sits in the loop
The “loop” is the cycle of input, model output, and action. In a fully autonomous system that cycle runs end to end without pausing. A human-in-the-loop design deliberately inserts a checkpoint before the consequential step — sending a message, approving a transaction, publishing a decision — so a person can catch mistakes the model cannot see. The same review often feeds corrections back to improve future output, which is why the pattern is common in machine-learning training, content moderation, and medical or legal review.
Why human-in-the-loop matters
Language models are fluent but not infallible: they can misread context, state something inaccurate, or strike the wrong tone. Keeping a person in the loop matters because it:
- Catches errors before they reach a recipient or take effect
- Keeps accountability with a named human, not an opaque system
- Builds trust, since users stay in control of what is said or done in their name
- Supports oversight in settings where a human decision is expected or required
How it works in DraftKite
DraftKite is human-in-the-loop by design. It reads incoming mail, labels it automatically, and writes a reply in your own tone — but it never sends anything on its own. Every reply lands in your drafts folder and waits there until you read it, edit it if needed, and click send. The automation does the drafting; you make the decision to send. That is the loop, and you are always the one who closes it.