Human-in-the-loop

Human-in-the-loop (HITL) puts a person at the decision point of an automated process. The machine does the expensive part — reading, drafting, classifying, proposing — and the human does the accountable part: deciding whether it ships. The pattern’s value depends entirely on where the loop actually sits. A confirmation dialog nobody reads is not a loop; a queue where a named person approves a specific proposed change is.

Why it matters

Language models are fast and fallible. For low-stakes work (classifying a ticket’s sentiment) the error cost is trivial and full automation is fine. For high-stakes work (emailing a customer, changing a deal, posting to a ledger) a wrong action is expensive to detect and reverse. HITL is the standard answer: keep the model’s speed for preparation, keep human judgment for commitment. It also produces something audit needs — a record of who approved what.

How human-in-the-loop works in Oneop

The pattern appears at several distinct points, each a real mechanism:

• Agent approvals. An AI Workforce agent in approve mode queues every proposed write for a person. Only an explicit acceptance dispatches it — see autonomy levels.
• AI reply drafts. In the ticket composer and the inbox, the AI-draft button hands the agent a draft to review and edit. It never auto-sends. The AI autoresponder (opt-in, off until you ask for it) likewise queues its drafts for approval rather than replying on its own.
• Approval steps in automation. Support-automation flows can include a human approval step gated by role rank; the flow pauses durably — surviving a restart — until a person decides.
• Dry runs before live runs. The automation simulator shows the exact execution trace without writing anything, and simulated agent runs are badged so they can never pass as live work.

The design position underneath: humans are removed from a loop by demonstrated reliability, not by default. Oneop makes the supervised modes the starting point and stores every reasoning trace, so the decision to automate further is made on evidence.

Related terms: autonomy levels, AI governance. Product pages: AI agents, Automation.