AI supervisor
The AI supervisor view is the operational surface for monitoring autonomous AI voice agents while they are on live calls. It fans every active AI engagement out into a single grid, surfaces health signals so problems are visible at a glance, and lets a human supervisor read transcripts in real time, send private guidance to the AI, or take over the call entirely.
Overview
When an AI voice agent is configured to handle inbound calls autonomously (see AI voice agents), it answers, holds the conversation, and only escalates when it can't resolve something. AI supervisor is the view that exists for everything in between -- the moments where you want to see how the AI is doing, intervene before a frustrated caller asks for a human, or step in mid-conversation when the AI is heading the wrong direction.
It is a passive surface by default. Opening it does not interrupt any call. Listening, whispering, and taking over are explicit actions a supervisor chooses.
The view is reserved for users with the Calls Supervisor or Calls Admin role. Users without those roles do not see AI supervisor in the navigation. See Permissions and roles for the full role model.
AI supervisor only shows autonomous AI engagements -- calls where the AI is the speaking participant on the line. AI coaching sessions (where the AI whispers suggestions to a human agent on a call) are not represented here; those appear inside the human agent's softphone coaching feed.
Accessing the view
Navigate to Calls > AI Supervisor in the left sidebar. The route is /ai-supervisor.
When no autonomous AI calls are in flight, the view shows a calm "All quiet" banner along with a brief description of the four capabilities the surface provides. The grid populates the moment any AI agent picks up an inbound call.
The fleet grid
Every active autonomous AI call appears as a card in the grid. Cards sort by health (most-attention-needed first), then by call duration. Each card shows:
Element | Meaning |
|---|---|
Health badge | INTERVENE, WATCH, or HEALTHY -- a heuristic signal computed from turn cadence and call age |
AI agent name | The name configured on the AI voice agent that is handling the call |
Duration and turn count | Live tickers updated every second while the call is in progress |
Transcript preview | The most recent turn from the call, truncated to fit the card |
Activity indicator | "No turn in Xs" appears when the conversation has gone quiet |
Wiretap link | Opens the wiretap panel for that call |
Health signals
Health is a heuristic, not a hard signal from the AI. Treat it as a hint that this card is worth a glance, not as a verdict.
Badge | Color | What it usually indicates |
|---|---|---|
HEALTHY | Green | Recent turns, conversational rhythm, no warning signals in the transcript |
WATCH | Amber | Slowing turn cadence, repeated questions, or other signals worth keeping an eye on |
INTERVENE | Red | Long silence, distress signals, or the AI looping. A supervisor should read the transcript and consider stepping in |
The badges recompute continuously as new turns land in the wiretap stream.
The wiretap panel
Selecting a card opens the wiretap panel. On screens 1280px or wider the panel docks to the right of the grid as a side-by-side composition. On narrower screens it slides in over the grid as a drawer with a backdrop; clicking the backdrop closes it.
Selection is preserved across navigation, so jumping out to Calls to look at the softphone and back to AI Supervisor keeps the same card open.
The panel has four regions, top to bottom:
Header -- AI agent name, live status indicator (a pulsing green dot for live, a calm grey dot for ended), peer-supervisor avatars, and a close control.
Transcript stream -- one bubble per turn, with role-aware styling (customer in plain, AI in iris-tinted, supervisor whispers as inline annotations).
Whisper composer -- a single-line text area for sending private guidance to the AI.
Supervisor actions -- one row per active call leg with caller, duration, and Listen / Take Over controls.
Peer supervisors
If another supervisor has whispered to the AI in the last five minutes, their avatar appears in the header as a small overlapping group. This is a soft presence indicator -- it tells you "someone else is also paying attention here" without requiring a back-channel.
Live versus ended
The panel keeps showing the transcript after a call ends so a supervisor can review what happened. The status indicator switches from "Live" to "Call ended", the whisper composer is hidden, and the supervisor action row is hidden. The transcript itself stays visible until the panel is closed.
Whispering to the AI
A whisper is a private message that is fed into the AI's context. The AI may incorporate it into the next thing it says; the customer never hears it.
To send a whisper:
Type the message in the composer at the bottom of the wiretap panel.
Press Enter to send (Shift+Enter inserts a newline).
The whisper appears in the transcript immediately as an annotation, attributed to your supervisor name.
Whispers work best as concise, directive instructions. "Confirm the order number before continuing" lands better than "the customer seems frustrated, maybe ease up." The AI parses the whisper as an instruction.
What whispers look like
Whispers appear in the transcript as inline annotation rows -- icon, label "Whisper · supervisorname", and the message text -- styled distinctly from peer turns. They are not customer messages, not AI messages, and not bubbles; they are sideband notes.
Pending and failed whispers
A whisper enters a brief "pending" state while it is being delivered. If delivery fails (network error, rate limit, etc.), the whisper is marked Failed with a red border. A retry chip appears above the composer offering to resend it.
Rate limit
To prevent the AI's context from being flooded, whispers are rate-limited to roughly one per second per supervisor. Sending faster surfaces a brief hint in the composer ("Slow down -- one whisper at a time") and quietly drops the excess.
Stepping in
Three actions let a supervisor take a more active role on a live call. The right choice depends on what you intend to do.
Action | What it does | When to use |
|---|---|---|
Listen | Joins the call as a silent listener. The customer and the AI cannot hear you | You want to verify what is happening with audio context, but don't intend to participate |
Take Over (Hand off gracefully) | The AI tells the customer that you are joining, then disconnects. You are now the agent on the call | You want to take the conversation, and you want the customer to know the human is here |
Take Over (Disconnect immediately) | The AI is disconnected without warning. You are the agent on the call from that moment | The conversation needs to end now -- aggressive caller, sensitive content, etc. |
Listening
Click Listen in the supervisor action row. You are added to the conference muted. Your softphone shows the call in a "Listening" state. You can leave the listen any time without affecting the call.
You cannot listen if you are already on another call. The action will fail and the call will not be joined.
Taking over
Click Take Over to bring up the takeover modal. The modal offers two paths:
Hand off gracefully -- the AI tells the customer "I'm going to hand you to a person now" before disconnecting. The graceful unwind takes a few seconds. The modal stays open showing "Handing off..." until the AI has fully exited.
Disconnect immediately -- the AI is dropped from the conference instantly with no transition. The customer hears whatever is on your line next.
Either way, once the AI is gone, you are the agent on the call. The view automatically routes you to Calls, where the existing softphone surface drives the rest of the conversation.
When a graceful handoff stalls
If the AI hasn't disconnected within about 14 seconds of starting a graceful handoff, the modal upgrades to a "Handoff is taking longer than expected" state with two options:
Keep waiting -- continue waiting for the graceful exit
Force disconnect -- escalate to an immediate disconnect
This is a fallback for the rare case when the AI doesn't honor the handoff request.
Who can use AI supervisor
Capability | Admin | Supervisor | User |
|---|---|---|---|
See AI Supervisor in navigation | Yes | Yes | No |
View the fleet grid | Yes | Yes | No |
Open the wiretap panel | Yes | Yes | No |
Send whispers | Yes | Yes | No |
Listen on a call | Yes | Yes | No |
Take over a call | Yes | Yes | No |
AI supervisor permissions are separate from AI agent configuration. Creating, editing, and deleting AI agents is restricted to admins and is done from the AI Agents section of the CRM, not from this view.
Best practices
Use Listen before Take Over when possible. A 30-second listen often resolves the question of whether a takeover is even needed.
Whisper before disconnecting. Most "the AI is going off the rails" moments can be redirected with a single whisper. Reserve takeover for cases where the AI cannot recover.
Prefer graceful handoff over immediate disconnect. A graceful handoff lands the call cleanly with the customer informed; an immediate disconnect is jarring and should be reserved for situations that warrant it.
Treat health badges as a triage hint, not a verdict. A WATCH badge is a prompt to glance at the transcript. INTERVENE is the one that should consistently pull a supervisor's attention.
Coordinate when peer avatars are showing. If you see other supervisors' avatars in the header, you are not the only person watching. Whispering at the same time is fine; taking over at the same time is not.
Related pages
AI voice agents -- configuring AI agents and assigning them to queues
Permissions and roles -- the full Calls permission model
Queue monitoring and dashboard -- the supervisor view for human-handled calls
The softphone interface -- the surface you are routed to after a takeover