Calls and analytics
Read real calls turn by turn, watch trends across them, and use both to make the agent more reliable.
| Surface | What it shows |
|---|---|
| Calls | One row per call, opening into the turn-by-turn conversation, the recording and the per-call scores. |
| Overview | Trends across calls on a dashboard you arrange. |
Start with a bad call under Calls for why, and Overview for how often.
When a call appears
Calls land within a minute or two of finishing. The recording has to upload and the analyser has to run, so a recent call may read as still processing.
Each call stores the turn-by-turn transcript, tool calls and analysis (summary, sentiment, outcome, duration). When recording is enabled, it stores the audio too. Sensitive details are redacted before storage -- see Privacy.
The call list
The Calls tab lists this agent's calls, newest first. Each row carries the headline signals:
- Customer (name or caller number), Channel (voice or text) and the call's date.
- Metrics as a rolled-up pass/fail tally, plus the three headline scores as their own columns: Task (did the agent complete the task), Hallucination (did it make something up) and Latency p50 (the median response time in milliseconds).
- Sentiment and Outcome (completed, abandoned, transferred or failed).
Filtering, saved views and export
The toolbar gives you Search, Filters (by sentiment, outcome, channel, source and date range), saved views and Export CSV.
A saved view stores the current filter set so you can return to it. The list defaults to every source (production, test and eval calls together); the Source filter narrows it.
Reading a call turn by turn
Click a metric pill to open that call's scores, anywhere else on the row for the full call.
The conversation sits on the left and an inspector panel on the right. The header shows channel, time, caller, conversation length and outcome.
Tool calls in the transcript
When the agent called a tool, a chip sits above the agent's reply, coloured green when the tool succeeded and red when it failed. Tap it to inspect the tool's arguments and what it returned.
This is usually enough to answer "why did it say that" -- the agent's prompt, its tools and what the caller said are all on the page.
The recording
For voice calls with recording enabled, an audio player sits above the conversation with a waveform and separate lanes for agent and caller. The player's clock shows the recording's own length, which can differ from the conversation length in the header.
If a call just ended, the recording shows "Recording processing" until the upload finishes. Text chats have no recording. Recording is governed by your Privacy settings.
Recording disclosure
Whether callers are told the call is recorded is part of your privacy settings, not something you set per call. Set the spoken disclosure in Privacy.
Per-call scores
The inspector panel leads with the verdict: sentiment, failure reason, and one-line summaries of each failing metric (linking to full reasoning). A key signals strip shows latency, talk ratio and tool-call success.
The metric grid
The per-call metric grid is grouped into:
- Judge metrics, scored by an LLM judge against your rubric, each showing a pass or fail with the reasoning behind it.
- Quality checks, deterministic pass/fail measures such as dead air, tool-call success and repetition.
- Audio quality, numeric readings taken from the recording (for example pitch and speaking-rate consistency).
Scoring on demand
If a call was never scored, Evaluate metrics runs the scoring engine on demand. Leave a thumbs-down vote on a verdict you disagree with to feed back into tuning. Scoring settings are covered in Evals. A compare control puts another recent call side by side.
Using call records in practice
- Find a failure. Open a bad call and read it back: a misread intent, a tool that returned an error (the red chip), a clumsy hand-off. The recording often makes a tone or timing problem obvious that the text alone hides.
- Spot a pattern. Filter to failed or transferred calls and read a handful. The same stumble across several calls is a prompt or tool fix waiting to happen.
- Confirm a fix. After you publish a change, watch new calls. The compare view lines up a call from before the change against one from after.
The improvement loop
- Find the weakest behaviour in a bad call or a sagging trend on the Overview.
- Write an eval for it so the failure is pinned.
- Fix the cause: tighten the prompt or a tool description (tools first, prompt last), then publish.
- Watch the new calls. Confirm the fix holds and the trend moves.