Human in the loop

Human in the Loop

Human in the Loop (HITL) is the connective tissue between raia Copilot and raia Chat. It defines how an AI agent recognizes its own limits, pauses, and hands the conversation over to a human — then how that human passes control back once the issue is resolved.

What you will learn

When and why an agent should escalate to a human
How the Escalation Skill triggers a handoff into Copilot
How human takeover works inside a live Chat conversation
How supervised responses and feedback flow back into training

Why Human in the Loop matters

No agent is perfect on day one, and some interactions — high-value deals, sensitive customer issues, edge-case questions — should always involve a person. HITL turns that reality into a designed workflow instead of a failure mode: the agent handles what it can confidently handle, and a human quietly steps in for the rest.

The three HITL patterns

Escalation (agent-initiated)

The agent's Escalation Skill fires — low confidence, an explicit keyword, or a business rule — and routes the conversation to the right team in Copilot. The customer sees a graceful handoff; the operator sees full context.

Human Takeover (operator-initiated)

An operator watching a Live Conversation in Copilot decides to jump in and reply directly. The agent yields control, the human drives the conversation, and the transcript stays unified across both speakers.

Supervised responses (approve & send)

For higher-stakes flows, the agent drafts a reply but waits for a human to approve, edit, or reject it before it goes out. Great for early rollouts and regulated channels.

How Copilot and Chat work together

01
Agent detects a handoff condition
Escalation Skill matches an intent, sentiment, or explicit request from the end user in Chat.
02
Conversation surfaces in Copilot
The live conversation is flagged, routed to the right team, and full transcript + agent reasoning are visible to the operator.
03
Human takes over in real time
The operator pauses the agent, replies as themselves, and continues the thread inside Copilot's Live Conversation view.
04
Control returns to the agent
Once the issue is resolved, the operator releases the conversation. The agent resumes with the added context recorded.
05
Feedback loops into training
The operator can flag turns, rate answers, and push corrections back into the agent's knowledge and instructions.

Auto vs Manual Mode

Choose whether the agent handles handoffs automatically or waits for explicit operator control.

Auto mode

The agent escalates to a human when its rules fire, waits while the human responds, then automatically resumes once the issue is resolved.

Manual mode

The agent pauses and notifies a human, who must explicitly approve, edit, or release every response before the conversation continues.

Setup guides

Step-by-step walkthroughs for configuring each part of the Human in the Loop workflow.

Guide 01

Setting up the Escalation Skill

Walk through enabling and configuring the Escalation Skill so your agent hands off to the right human team at the right moment.

Guide 02

Notification & Routing in escalations

Configure how escalations notify the right people and route to the right team — so handoffs land where they can be acted on immediately.

Guide 03

Inline Chat Escalation

Let end users request a human directly from the chat interface, so escalations feel natural and immediate.

Guide 04

Receiving escalation in Copilot

See how escalated conversations arrive in Copilot so agents can pick them up, respond, and hand back to the AI.

Guide 05

Taking over a conversation

Learn how a human operator can step into a live conversation, take control, and hand it back to the agent when ready.

Guide 06

Handing back to AI

Learn how a human operator returns control of a conversation to the agent once the issue is resolved.

Human Oversight

Human Oversight keeps a person accountable for high-stakes decisions without slowing down routine automation. Review queues, audit logs, and approval gates let you supervise AI actions at scale.

Review queues

Hold non-routine responses in a queue so a human can review, edit, or reject them before they reach the customer.

Audit trails

Every agent decision, handoff, and override is logged so you can trace what happened and why.

Approval gates

Require explicit human approval for sensitive actions like refunds, account changes, or escalations beyond a confidence threshold.