Monitor Analysis results

Analysis runs after a conversation ends when its frozen agent version enables at least one evaluation criterion, data point, or sentiment setting. The latest result appears in Conversation Overview; detailed runtime evidence appears in Activity.

Interpret result fields and states#

A completed Analysis can contain a summary, aggregate result, criterion results with rationale, typed data collection values with rationale, sentiment, model, version ID, and the conversation segments read. The aggregate result is Success if all configured criteria succeed, Failure if any criterion fails, and Unknown when no criterion determines that aggregate.

The lifecycle is pending, processing, done, skipped, or failed. Pending is waiting for a claim. Processing is owned by a worker. Done has a validated result. Skipped means there was no transcript or no active Analysis output. Failed means model execution or result validation failed; the conversation itself remains available for inspection.

Do not confuse a failed Analysis with a failed conversation. Check the conversation status and end reason independently, then inspect analysis_completed or analysis_failed under Activity.

Rerun an ended conversation#

Use Run analysis or Run analysis again only for an ended conversation. The action resets a done, skipped, or failed state and attempts Analysis again. An already processing run is not duplicated. API-key authentication can read conversation detail but cannot trigger this dashboard mutation; the Analysis rerun API requires its supported session authorization.

A rerun uses the original frozen agent-version Analysis configuration, not the latest draft or latest published version. It can therefore remain skipped if the original configuration had no active outputs. It is billable only when it reaches the model; an empty transcript or inactive Analysis setting skips without a model request or added cost.

Investigate failure and delivery#

For skipped, confirm the conversation has at least one completed turn and the frozen version enables a criterion, data point, or sentiment. For failed, inspect the safe analysis_failed event and nearby llm_request; provider response text is not exposed. A malformed model response can fail after usage has already occurred.

After successful Analysis, an enabled post-call webhook is attempted. Webhook failure does not change done into failed; delivery has its own status and post_call_webhook_failed event. Review daily impact under Usage and quota, because Analysis model cost is included in the conversation total and workspace allowance.