Skip to main content
You work from the call details page, which you open from Call Logs or the Recent Red Flags table on the Overview dashboard.

​
What is a root cause analysis?

A root cause analysis goes past spotting that a call failed and works out why it failed. Once you know the cause, you can fix it so the problem doesn’t come back. Do one when:
  • A critical red flag is triggered, such as a hallucination or a missed disclosure.
  • A user reports a poor experience.
  • You notice the same issue in several calls in Call Logs.

​
Analyze a call

1

Open the call

In Call Logs or Recent Red Flags, click the call to open its details page.
2

Read the transcript

Read the conversation on the left to understand the context. Pay close attention to the turns leading up to the issue, and play the recording if the words alone don’t explain it.
3

Check the tabs

Use the tabs on the right:
  • Overview: the red flags, labels and sentiment for the call.
  • Evals: which evals failed, and the evidence each one cites.
  • Voice Metrics: response time and latency by stage. Look for slow turns or a long agent monologue. On a replay call, it also shows transcription accuracy. See Voice metrics.
  • Data Extraction: the values Tuner extracted, to check whether the agent captured the right details.
4

Open the trace when the problem is timing

If the call has a Traces tab, open it. The span tree shows the operations behind each turn, so you can tell whether a slow turn was the model, the synthesis or a tool call waiting on your backend. See OpenTelemetry Traces.
5

Form a hypothesis

Based on what you found, decide what caused the issue. For example: “The agent hallucinated because the prompt was ambiguous,” or “The agent skipped the disclosure because the instruction was missing from the prompt.”
6

Check it against other calls

Look for other calls with the same issue in Call Logs, using search and filters. If your hypothesis explains them too, you’ve likely found the cause.
A call's details page with the Overview, Evals, Voice Metrics, Data Extraction and Traces tabs, the transcript, and the red flags and labels on the call
Prove the fix. Add the call to your Dataset, then replay it after you change your agent. The caller says exactly the same thing on every replay, so if it now passes your evals, your fix worked.

​
When part of the analysis is missing

Tuner analyzes each call in several steps: outcome, intent, data extraction and evals. If one of the intent, data extraction or eval steps fails, the other steps still complete, and their results appear as usual. The results of the failed step are missing from the call. The Get call response lists which steps failed. Alerts are not checked for a call while any of its steps has failed.

​
Common root causes

​
Diagnose across all calls

Manual analysis is the right tool for one call. To find the cause of a pattern across many calls, such as why one intent keeps failing, use the MCP diagnosis. It reads your calls, red flags and eval results in bulk, finds the pattern and tells you what to fix. Start it from Diagnose Agent on the Overview dashboard or in Call Logs.

​
Next steps

Diagnose your agent

Analyze patterns across all your calls and get fix recommendations.

OpenTelemetry Traces

See which operation inside a slow turn took the time.