Conductor can read calls and report what went wrong on each one and which problems repeat. Start from Call History, or ask in any Conductor conversation.
Conductor works with calls only. Chat sessions from Chat History cannot be attached or investigated yet.
Ask about selected calls#
Select calls#
In Call History, tick the calls you want. A bar shows "N selected".
Choose Ask Conductor about these calls#
The Conductor side panel opens with the calls attached (up to 10; with more selected, the first 10 are attached and a message says so) and a question ready: "What do these calls have in common, and what should the agent do differently?" (or "What happened on this call, and what should the agent do differently?" for one call).
Send#
Edit the question if you like and send it.
Ask about recent calls#
Without attached calls, ask in a Conductor conversation, for example "Why did my recent calls fail?" or "Look at the last 20 calls".
- Conductor looks at your most recent ended calls. A number from 1 to 50 in the message sets how many (default 10).
- If the message mentions failures ("fail", "unsuccessful", "bad", "negative", "error", "drop"), it only looks at calls that were unsuccessful or had negative sentiment.
What Conductor checks#
For each call, Conductor flags these issues:
| Issue | Rule |
|---|---|
| not successful | Post-call analysis marked the call unsuccessful. |
| negative sentiment | User sentiment was Negative. |
| tool failures | A tool result in the transcript contains an "error" or "errors" key, or "success": false. |
| slow responses | The call's average end-to-end latency was above 1,500 ms. |
ended with <reason> |
The disconnection reason starts with error. |
The reply says how many calls it looked at, how many had no problems, and the three most common issues. The result lists each call with its id, duration, issues and summary, and the issue counts as badges.
With an OpenAI key, Conductor also adds its own explanation before the findings, using the calls' status, duration, disconnection reason, sentiment and summary.
The investigation uses post-call analysis fields. Calls that have not been analysed (for example, still running) show fewer issues. Use Re-run analysis in the Call History selection bar to analyse calls again.
The Ask Conductor button#
The Ask Conductor button in the Call History header is different: it opens the agent edit panel, with the open call (or the 10 most recent calls in the current view) attached as context, including summaries and the start of each transcript. Use it when you already know the agent needs a change: "these callers keep asking about parking, add the parking details to the prompt". See Build and refine.
From findings to fixes#
- Open a flagged call from the result and read its transcript.
- Turn the scenario into a simulation test case, or ask Conductor to write tests with the agent and the calls attached.
- Fix the agent with the edit panel, then run the tests.