Help

Overview Admin Chat Data Sync Upload Documents Search Settings Curated Answers Agents, Prompts & Functions Contact Support

Deep-Dive: Handoff and Escalation

Airgentic Help

Deep-Dive: Handoff and Escalation

This module covers human monitoring and intervention in Airgentic conversations. Read this when you want to set up live monitoring, understand when to intervene, or plan escalation workflows.


Who this is for

  • Operators who monitor and intervene in conversations
  • Platform administrators who configure handoff modes, hours, and auto-assign
  • Live-support staff who should get the Agent role
  • Anyone planning live support alongside AI

What you'll learn

  • How the Human Handoff screen works
  • The three service operating modes
  • Weekly staffed hours and timed overrides
  • Optional enquiry auto-assign
  • Wait-time colours, saved replies, and internal notes
  • Generating, revising, and sending AI drafts
  • The Agent role for live-support staff
  • When and how to intervene in conversations
  • Automatic attention flags and what triggers them
  • Best practices for human-AI collaboration

The Human Handoff system

What it provides

Human Handoff lets your team:

  • Monitor active conversations in real time
  • See when conversations need attention (including a count on the browser tab)
  • Search, filter, and sort the inbox
  • Take over conversations from the AI
  • Generate an AI draft, insert a saved reply, or leave an internal note
  • Hand conversations back to the AI
  • Follow weekly staffed hours, with timed overrides when needed
  • Auto-assign waiting enquiries to available agents
  • Colour queued chats amber then red as the customer waits
  • Set the overall operating mode for your service
  • Give live-support staff an Agent login that cannot change configuration

This enables a hybrid model where AI handles most interactions, but humans can step in when needed.

Where to find it

Go to Human Handoff in the admin console. The screen maintains a live connection and updates automatically.


Operating modes

Each service can run in one of three modes:

AI Only

Aspect Description
AI behaviour Handles all conversations
Human involvement None — no monitoring, no intervention
When to use Fully autonomous operation, no support staff availability

Human-Supported AI

Aspect Description
AI behaviour Responds normally to all conversations
Human involvement Can monitor, see flags, and take over when needed
When to use Standard operation — AI handles most, humans available for escalation

This is the most common mode for services with any human support capability.

Human Only

Aspect Description
AI behaviour Disabled — does not answer questions. After the first customer message, a one-time holding notice is sent (“A human assistant will be with you shortly.”)
Human involvement Must respond to every conversation
When to use Rare — perhaps during critical incidents or specific use cases

Switching modes:
- Change mode per service using the button group
- Change all services at once with Apply to all services

Scheduled hours

Services can follow a weekly staffed-hours policy instead of leaving the radios on one mode all day. Open Hours in Desk settings. Typical setup:

When Mode
Selected weekdays, inside the start/end times Human-Supported AI (Human Only is available)
Nights, weekends, and closed dates AI Only (Human-Supported is available)

Choose a timezone, start and end times, weekdays, closed dates (single days or inclusive ranges), and the in-hours vs outside-hours modes. Overnight windows are not supported. Public holidays are not imported automatically.

The live mode is evaluated when a customer message arrives. Clicking a mode radio while a schedule is on asks how long the override should last (until the next scheduled change, until a chosen time, or until you turn the schedule off) and does not rewrite the saved hours. Turning the schedule off snapshots the current live mode as the new manual setting.

Only admins can edit hours. Agents see the live mode but cannot change the schedule.

A playbook can overlay a different mode for one session without changing the Hours control. The Hours / Redis override still wins. Playbook conditions that match hours use the schedule result before that overlay.

Auto-assign

Auto-assign is optional and off by default. When a service is in Human Only or Human-Supported AI, an Auto-assign switch appears next to the mode controls.

Mode What is queued
Human Only Every new customer conversation
Human-Supported AI Only conversations that need a human (help requested, thumbs down, non-answer, or negative sentiment)

Agents join the pool when they have Human Handoff open and Available is on. Auto-assign fills My chats up to 10 concurrent claims; an agent at that cap is skipped until they hand a chat back, transfer it, or go Away. Manual Take Over can exceed the cap. Assignment does not expire after five minutes. Auto-assign toasts and marks the row unread; it does not open the transcript. The oldest waiting conversation is claimed for the next agent with capacity. If nobody is free, the conversation stays in Queued. Agents can still Take Over an unclaimed card themselves.

Toast notifications do not put you in the pool — the Human Handoff page must be open. Only admins can turn Auto-assign on or off.

Agent role

Give live-support staff the Agent role instead of Customer Admin. They land on Human Handoff after sign-in, can work the inbox and Insights, and cannot change modes, hours, Auto-assign, or other configuration. See User Management.


Monitoring conversations

The inbox

The desk is a three-column workspace. The left column buckets conversations:

Bucket Meaning
My chats Assigned to you. Switching rows does not notify the student.
Queued Waiting for a human (escalation, Human Only, or returned to the queue). Not the assignment cap.
Monitoring AI is still answering. New Human-Supported chats start here.
Recent Closed or reset threads from the last 48 hours (separate tab, read-only).

Each card still shows location, opening question, time, message count, and tags, plus a small service chip. Queued chats also show how long the customer has been waiting; those cards turn amber then red at the times set in Desk settings.

Search, Filter, and Sort sit above the inbox. Hidden chats still count in need attention, so a filter cannot hide an escalation.

Attention flags

Conversations requiring attention are highlighted with a red pulsing border and coloured badge:

Flag Trigger
Thumbs Down User gave a negative rating
Non-Answer AI couldn't answer the question
Negative Feedback Negative sentiment detected
Help Requested User asked for a human

These conversations increment the need attention banner above the inbox and the count on the browser tab.

Why flags matter:
- They surface the conversations most likely to need help
- They help you prioritise limited attention
- They indicate potential service quality issues


Intervening in conversations

Viewing a conversation

Click any conversation card to open the transcript in the workspace. Switching My chats does not claim, release, or notify the student.

  • Full transcript with messages colour-coded by role (customer on the left; AI and human replies on the right)
  • Clocks on messages, and day dividers when a chat spans more than one day
  • Internal notes (amber, locked) that only agents can see
  • All citations and references

Taking over

Click Take over to:
1. Claim the conversation
2. Pause the AI (it stops responding)
3. Notify the customer that a human has joined (in Human-Supported AI)

Once you've taken over:
- Type a reply, click Generate to draft the same response the customer would have seen, or insert a saved reply
- Use Note for a private message that only other agents can see
- Press Enter to send; use Shift + Enter for a new line
- Your messages appear as a human-agent reply — citations stay in the admin Sources list
- Generating a draft counts toward usage; typing the reply yourself does not

Hand back, transfer, and return to queue

Once a human has stepped in, the AI stays paused until Hand back to AI.

Action Effect
Hand back to AI Two-click confirm in Human-Supported AI. The AI resumes; the student is notified.
Transfer Pass to a colleague currently on the desk. The student sees one line naming them.
Return to queue Unclaim, stay paused, no student message. Auto-assign may pick someone else.

Idle time does not drop the claim. Recent conversations are read-only.

Multi-agent coordination

If another agent has already claimed a conversation:
- The Take Over button shows their name and is disabled
- You can wait for them to release it, or they can Transfer or Return to queue so you can take over


Escalation workflows

Built-in escalation

The escalate_to_human global function provides programmatic escalation:

  1. AI detects the user wants a human (or is instructed to offer escalation)
  2. Function collects customer details (name, email, optional reason)
  3. Conversation summary is emailed to your support team
  4. AI confirms to the user that their request has been escalated

This creates an asynchronous handoff — the user doesn't wait for a live agent.

Live escalation

For real-time escalation:

  1. User asks for a human
  2. AI can be prompted to acknowledge and indicate someone is available
  3. Staff monitoring Human Handoff sees the flag
  4. Staff takes over and responds directly

This requires someone actively monitoring.

Designing your escalation approach

Approach Works when
No escalation AI handles everything; users contact support separately if needed
Async escalation Staff aren't available for live chat; email follow-up is acceptable
Live escalation Staff monitor during business hours; users expect real-time response
Hybrid Staffed hours for live cover, AI Only or async email outside hours

Best practices

When to intervene

Do intervene when:
- User is clearly frustrated and needs human empathy
- AI is stuck in a loop or giving obviously wrong answers
- The question requires authoritative human judgment
- User explicitly asks for a human
- High-stakes or sensitive situations

Don't intervene when:
- AI is handling the conversation well
- You'd just be saying what the AI said
- The user is exploring and doesn't need help
- Intervening would slow down resolution

How to intervene effectively

  1. Read the full conversation — Understand what's already happened. Turn on Store full conversations on the Human Handoff screen first, or agents will only see the most recent turns.
  2. Acknowledge the user — Let them know a human is now helping
  3. Use Generate AI Answer when you want a sourced draft — The AI suggests markdown you can edit before sending as yourself. Citations are for you, not the customer. This still works in Human Only: the draft is generated silently and only sent when you choose. Using Generate counts as a billable AI answer; typing the reply yourself does not.
  4. Resolve or escalate — Either answer their question or tell them next steps
  5. Decide on handback — Will AI or human handle any follow-up?
  6. Tag the conversation if your team uses tags (callback, complaint, or a custom label). Only the claimed rep can edit tags; they are stored on Insights for later review.

Maintaining quality

  • Review flagged conversations regularly — Don't let the queue build up
  • Look for patterns — Are certain questions always getting flagged?
  • Feed learnings back — Use issues to improve prompts, content, or curated answers
  • Train staff — Make sure everyone knows how to use the handoff screen

Notifications

You don't have to stay on the Human Handoff page:

  • Sound plays when a conversation needs attention
  • Toast appears even if you're on another admin page
  • The browser tab shows a count, for example (3) Human Handoff
  • Optional desktop notifications (the bell) alert you when the tab is in the background

Common scenarios

User wants to complain

  1. Flag triggers (negative sentiment or thumbs down)
  2. Monitor sees the flag
  3. Takes over and acknowledges the user's frustration
  4. Either resolves or escalates to appropriate team
  5. Decides whether to resume AI or close manually

AI can't answer

  1. Non-Answer flag triggers
  2. Monitor reviews what was asked
  3. If answer exists, takes over and provides it (then considers adding curated answer)
  4. If answer doesn't exist, explains to user and offers alternative

Sensitive topic

  1. AI may be prompted to decline or escalate
  2. Human takes over to handle appropriately
  3. May need to consult policy or other staff before responding

Metrics to track

From Insights, monitor:

Metric What it indicates
Non-Answer rate Content gaps or prompt issues
Thumbs down rate User satisfaction problems
Negative sentiment rate Overall experience quality
Escalation requests User expectation for human support

High rates in any of these suggest investigation and improvement.



Back to: Optional Deep-Dives

You have unsaved changes