Skip to main content
Set up Agent Assist Copilot to warn a human agent when a caller asks for a supervisor, then test the warning before attaching Copilot to a voice agent. During calls, Copilot follows the transcript, offers guidance, and answers the human agent’s questions. The human agent speaks to the caller. Copilot provides suggestions in the workspace; it does not speak those suggestions to the caller.

​
Before you start

  • Your account must belong to the workspace that contains the voice agent and call.
  • The voice agent must have a Copilot Agent attached, with a published version.
  • For in-app transfers, enable Accept Calls in your account menu and allow browser microphone access.
  • For observation and takeover from Live Calls, enable Live Observability in the voice agent’s call settings.
  • Knowledge answers need configured knowledge sources. API actions need configured tools and access to the target system.
See Call Transfer for transfer destinations and agent availability.

​
Configure a Copilot Agent

  1. Open AI Agents > Copilot Agents and select New Copilot Agent.
  2. Enter a name and description, then create the agent.
  3. Select Start Editing when you need to change an existing configuration.
  4. In Overview, write Instructions for the copilot. Start with the example below and keep the default models for your first test.
  5. Configure the relevant tabs below.
  6. Test the draft, then have a team administrator select Publish.
  7. Open the voice agent’s call settings and select it under AI Assist for Human Agents.
  8. Save and publish the voice agent’s configuration through its normal publish workflow.
Copilot Agent Overview in the Revve demo workspace, with instructions, model choices, and the live summary setting The screenshot shows an unpublished demo configuration. It is not attached to a voice agent. Its instructions are:
Fast lane model classifies transcript turns; Copilot model answers questions and runs guides. Both default to GPT-4.1 mini. Change a model when representative tests show missed triggers or poor answers, then compare response time as well as accuracy. Show the summary column is on by default, with Refresh set to Every 30 seconds. Use a shorter interval when agents need a more current summary; use a longer interval to reduce repeated summarization work. Draft changes save automatically while you hold the edit lock. Publish the draft to make its configuration available for new sessions. An active session keeps the Copilot version selected when it started. Publishing a new version does not replace that session’s configuration. Tool definitions are an exception: edits to shared tools can affect active sessions without publishing a new Copilot version. Coordinate endpoint, credential, and behavior changes with operators before changing a tool used by live calls.

​
Add a supervisor-request warning

  1. In Monitors, select New monitor.
  2. Set Name to Supervisor request and Message to Acknowledge the concern and follow the supervisor escalation process.
  3. Keep Severity at Warning and On enabled.
  4. Set Say instead (optional) to I understand your concern. Let me check the next step for involving a supervisor.
  5. Under Trigger, select Exact phrases and enter supervisor in Phrases.
  6. Keep Watches set to Caller and Cooldown (s) at its default of 120, then select Save monitor.
Use Exact phrases for a known term you can test predictably. Use Copilot decides from the description for requests that can be phrased many ways, and test representative wording. Increase the cooldown if repeated warnings distract agents; shorten it if a repeated request needs attention sooner.

​
Test before a live call

With the monitor saved in your draft, open Test with a past call. Paste a transcript with one turn per line, then select Run:
You can also search for a past call and select it. The results show triggered or suppressed monitors, skills, completed steps, and knowledge queries. A synthetic transcript test in the Revve demo workspace, with a supervisor-request monitor triggered on the third turn The expected result is your monitor’s ID in Fired on turn 3, matching the screenshot. If it is absent, check that the monitor is enabled, Phrases contains supervisor, and Watches is Caller. This test checks the draft’s transcript classification. It does not place a call, execute the full live Copilot workflow, or display guidance to live agents. Use a controlled test call to check audio, transfer, knowledge answers, and tool confirmations before operational use.

​
Open Copilot during a call

For an in-app transfer, select Accept on the incoming transfer card. Revve opens the live-call workspace in a new tab. Allow the tab and microphone permissions when prompted. If the browser blocks the tab, Revve can show a compact phone instead. For a call on the Live Calls board, open the call and select Take Over when you need to speak to the caller. The assigned human agent controls the Copilot session. Other authorized team members can observe it; their status shows Viewing.

​
Read the workspace

The header shows the Copilot name, version, and status. Listening means it is ready; Thinking means it is processing work. On a narrow screen, open the Copilot panel from the call header. On a wide screen, you can collapse the summary column.

​
Ask for help and search knowledge

Type a question in the composer, or select Draft my next line, Summarize so far, or Search knowledge. For example, ask: “What information must I collect before I escalate this request?” Press Enter to send. Press Shift+Enter for a new line. Type / to show available skills and commands. If Copilot suggests a knowledge search, run or dismiss the suggestion. Automatic answers depend on the Knowledge configuration. When an answer includes source chips, select a source to inspect its content. Open original opens its original URL when available. Check the source and the caller’s context before you use a suggested response. Knowledge quality depends on the selected, available documents.

​
Follow warnings and guides

Monitors can show warnings with a reason, suggested wording, and transcript evidence. Review the warning and select Dismiss when appropriate. A skill guides you through a configured procedure. It can start from a configured condition, a suggestion you accept, or the slash-command menu. The guide shows completed, skipped, and pending steps. Completion depends on the configured conditions and available conversation evidence. Select End skill when you need to leave the current procedure. A completed guide is not proof that an external record changed.

​
Review tool actions

When a step or Copilot request needs your input, the panel shows Needs you or a pending action.
  1. Check the tool name and proposed values.
  2. Select Edit values if you need to correct them.
  3. For a read-only action awaiting approval, select Run.
  4. For a write action, select Confirm and run, then Click again to run.
  5. Select Skip if the action should not run.
  6. Check the result before you tell the caller that the action succeeded.
Configured read-only tools can also run automatically. Write tools require human confirmation. API tools require a working integration. Attaching Copilot does not automatically connect SAP, Cisco, or another external system.

​
Review the session after the call

Open the call in call history and select Copilot.
  • Agent view replays the saved Copilot experience alongside the transcript.
  • Debug log shows recorded events for detailed review, including actions, notes, and errors.
  • If the call has multiple Copilot sessions, use the session selector.
Replay uses the saved events and version. It does not run tools again. Calls without a captured Copilot session have no Copilot history. A Copilot note belongs to the session. A summary, note, or completed guide does not automatically update an external CRM. Use a configured tool or a separate backend integration for those updates. See Automation Rules for post-call API actions.

​
Access and troubleshooting

Live assistance requires Revve to receive the conversation after transfer. A cold SIP transfer leaves that media path and does not capture the human conversation. An external phone system needs an explicit media and session integration for continued assistance. An embedded phone widget alone does not provide it.

​
What’s Next