What the Live Feed shows
Inbox > Conversations > Live Feed lists every conversation currently happening in your widget. Rows show:- The user’s name (or “Anonymous” if not identified)
- The agent handling the conversation
- The current topic (computed from the messages so far)
- A pulsing green dot while the user is active
- A mood badge — Happy, Neutral, Confused, Frustrated, or Angry — derived from the AI’s running sentiment score
What you can do mid-conversation
Once you have a live conversation open:Inbox training (live whisper)
Type a message into the Inbox training field. The AI receives it as a high-priority note for its next response, and the user does not see your input. When the visitor is identified, Halo also saves the same lesson as durable Inbox Training scoped to that customer (see AI Learning Priority). Use this when:- The AI is about to give a wrong answer (“Don’t recommend X — it’s deprecated. Suggest Y instead.”)
- You want the AI to gather more info (“Ask which integration they’re trying to set up.”)
- You want a softer tone (“Acknowledge their frustration before troubleshooting.”)
Add corrections
Click Add Correction to flag a specific AI message as wrong and write what it should have said. Corrections feed into Halo’s training data and improve future responses for similar questions.Send feedback
Thumbs up / thumbs down individual messages so your team can review them later under Completed conversations.Take Over Chat
When you need to talk to the customer directly, click Take Over Chat. This:- Pauses the AI for that session
- Creates a ticket in your inbox with status In Progress
- Assigns the ticket to you
- Lets you reply as a human in the same conversation thread
Completed conversations
When the user closes the widget (or times out), the conversation moves to Inbox > Conversations > Completed. The Completed view includes:- The full transcript
- The outcome — Resolved, Escalated, Abandoned, etc.
- A computed AI score (heuristic from sentiment, outcome, and feedback)
- Per-message ratings and trainer feedback you can add retroactively
Why watch live
Watching live conversations is the highest-leverage thing a support manager can do in the early days of an AI deployment. Patterns you’ll catch:- Knowledge gaps — questions the AI keeps fumbling
- Tone mismatches — places where the AI is too formal, too casual, or too verbose
- Missing context — questions where the AI is missing data your app has but isn’t sending
- Edge cases — workflows your instructions don’t cover
| Pattern | Fix |
|---|---|
| AI doesn’t know an answer | Add a knowledge source (Knowledge Overview) |
| AI tone wrong | Update agent instructions (Instructions) |
| AI is missing customer state | Send better context from the SDK (Context Planning) |
| AI escalates too aggressively | Tighten escalation rules (Escalation) |
Notifications
You can opt into Slack notifications for sentiment drops, escalations, and knowledge gaps so you don’t have to manually watch the Live Feed all day. See Slack Integration.Where to go next
Live Train
Practice with the agent in a sandbox before iterating on production.
Tickets vs Conversations
The distinction between live conversations and async tickets.