An operations dashboard should tell staff which customer requests need action and give engineers evidence for why they stopped. Session counts and pipeline latency alone cannot answer whether a booking exists or a handoff was accepted.
LiveKit's observability data hooks and session-observability reference describe session evidence surfaces. Combine supported data with your application's own business records rather than inferring outcomes from conversation text.
Model sessions, requests, and actions separately
One session can contain several requests, and several sessions can concern one unresolved issue. Each action needs intended scope, remote attempt, verified result or uncertainty, current owner, and next step.
Preserve client and environment context throughout. Reporting, media access, exports, and replay controls must enforce authorization independently of interface filters. Operators may need different roles for viewing evidence and executing recovery.
Work through a fictional partial-success call
A caller requests visits at two properties. One appointment is verified; the other submission times out. The dashboard shows a completed session, one confirmed appointment, and one uncertain action requiring reconciliation.
The recovery view includes the action reference and evidence needed to inspect the destination. It should not offer a blind “retry everything” button that can duplicate the first appointment or the uncertain second one.
| Dashboard surface | Operational question |
|---|---|
| Pending queue | Who must act next and by when? |
| Action timeline | What was authorized and attempted? |
| Session evidence | What did the caller and agent encounter? |
| Resource view | Is capacity or a dependency impaired? |
| Reconciliation view | What proves the remote result? |
Keep metric definitions visible
Separate answered calls, captured requests, verified bookings, accepted handoffs, and resolved work. Show reporting cutoffs and delayed data. A healthy worker or normally closed session is not a customer-success metric.
Detailed dashboards add integration and maintenance work. Begin with a reliable request-to-action table and an owned exception queue. Add charts when the definitions and joins are trustworthy, not before.
Test the dashboard using fictional partial outcomes, missing evidence, foreign-client identifiers, and failed log collection. Critical action state should remain recoverable even if an observability feed is delayed. Review completed-outcome dashboard design for the general operator model.
Questions about operations dashboards
Should operators see raw recordings by default?
Grant access by approved purpose and role. Many recovery tasks need identifiers and action evidence rather than full media.
Can the dashboard retry a failed action?
Only through an authorized workflow that distinguishes explicit failure from unknown outcome and checks for existing remote effects.
What belongs at the top of the page?
Unresolved consequential work, age, owner, and next action. Engineers can drill into session and resource evidence without hiding the business queue.