Human-in-the-Loop
ما هو Human-in-the-Loop?
هذا المصطلح متوفر حاليًا باللغة الإنجليزية فقط.
Human-in-the-loop is the approval gate for an AI agent's actions: before anything data-changing executes, the action pauses as a proposal that names exactly what the agent wants to do, and a human operator approves or rejects it. Reads and drafts run inline; chargeable, side-effecting, or destructive actions pause. Devotel Orbit's AI assistant, Orby, runs every gated tool call through that gate — a pending action waits for an operator, and proposals no one decides on expire instead of running.
More detail
The gate is a four-state lifecycle: a proposal sits pending until an operator approves or rejects it, and if no one decides before the proposal window closes it expires instead of executing. Each proposal card names the exact action with its validated arguments and an estimated cost, so the decision is precise, not a blanket yes.
Which actions pause is configurable. Orby's confirmation policy can run reads, searches, and drafts inline while pausing chargeable or side-effecting actions, or pause only above a cost threshold you set — and a destructive verb like delete or refund gates regardless of the declared policy. The workspace owner can also require approval for high-risk action types such as sender-ID registration, number ports, refunds, and data exports; that stance is off by default.
Who may decide is constrained: the operator who raised a proposal is the person who resolves it, the system re-checks that identity and the operator's role at decision time, and every approval, rejection, and expiry writes an audit-log event naming the action, the operator, and the workspace — so the proposal history is replayable.
الأسئلة الشائعة
- Which AI actions pause for a human in Orby?
- Read-only and drafting actions — searches, lookups, message drafts — run inline. Chargeable or side-effecting actions pause as proposals, and anything with a destructive verb (delete, refund, cancel, …) pauses regardless of the configured confirmation policy. The workspace owner can additionally require approval for specific high-risk types like number ports, refunds, or data exports.
- What happens if nobody approves an Orby proposal?
- It expires. Each proposal has a decision window; if the window closes without an operator approving or rejecting it, the proposal is cancelled and the action never runs. Expiry, like approval and rejection, is recorded in the audit log.
- How is human-in-the-loop different from a handoff?
- Human-in-the-loop gates a single agent action with an approve/reject decision; a handoff transfers a whole conversation from an agent's bot to a human or to an allowlisted downstream agent. They are independent: an agent can gate sensitive actions and still hand off conversations when the model decides a person should take over.
- Can a workspace require approval to differ by action type?
- Yes. Orby's confirmation policy supports running low-risk actions inline, pausing chargeable ones, or pausing only when the estimated cost meets a threshold — and the workspace owner can layer mandatory approval onto specific high-risk action types. The default workspace stance requires no approval policy at all, so teams opt into the gate deliberately.
See also
Build it on Orbit
Voice, messaging, email, video, and AI agents on one platform and one pay-as-you-go bill. Start free — no credit card required.