Transcripts and analytics

Once your agent is taking calls, this is where you find out how it is doing. Every call is transcribed, analysed and scored, so you can read exactly what happened on one call and measure whether the agent is getting better across many.

Three surfaces in the agent's sidebar cover post-call work:

• Overview is the analytics grid: trends across calls (volume, duration, outcomes, latency) on a dashboard you arrange.
• Transcripts is the call list: one row per call, opening into the full turn-by-turn conversation, the recording, and the per-call scores.
• Insights clusters where calls fail into recurring patterns and writes them up for you.

Start with a single bad call in Transcripts when you want to know why. Move to Overview and Insights when you want to know how often and what pattern.

Where to look after calls start

Calls land within a minute or two of finishing. The agent has to hang up, the recording has to upload and the analyser has to run, so a call you just made reads as still processing for a short window before it settles.

Each call is saved as a session with its turn-by-turn messages, any tool calls the agent made, an analysis (summary, sentiment, outcome, duration) and, where recording is enabled, a link to the audio. Sensitive details are redacted before anything is stored: card numbers, tax file numbers and other configured categories are masked on the way in, on both the transcript and any custom metadata. See Privacy for what gets redacted and how to configure it.

Transcripts

The Transcripts tab lists this agent's calls, newest first, one row per call. Each row carries the headline signals so you can scan without opening anything:

• 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).

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, for example "failed calls this week" or "transfers". The list defaults to showing every source (production, test and eval calls together); the Source filter narrows it. Export sends the current filter set to a CSV download.

Click a metric pill to jump straight into that call's scores; click anywhere else on the row to open the full call.

Reading a call turn by turn

Opening a call shows the conversation on the left and an inspector panel on the right.

The conversation reads in order, exactly as it happened: each caller turn and each agent turn as a chat bubble. When the agent called a tool, that shows as a chip above the agent's reply, coloured green when the tool succeeded and red when it failed, and you can tap a chip to inspect the tool's arguments and what it returned. This is usually enough to answer "why did it say that": the agent's behaviour follows from its prompt, its tools and what the caller said, and all three are on the page.

The header above the conversation carries the channel, time, who called, the analysed conversation length and the outcome.

The recording

For voice calls with recording enabled, an audio player sits above the conversation with a waveform and separate lanes for the agent and the caller, so you can hear the call while reading along. The player's clock shows the recording's own length, which can differ from the analysed conversation length in the header.

If a call just ended, the recording reads as "Recording processing" for a short window while the audio uploads, then either appears or settles into "No recording available". Text chats have no recording. Recording is governed by your privacy settings; when it is switched off, no audio is stored.

Per-call scores

The inspector panel leads with the verdict: the analyser's sentiment and any failure reason, the one-line summaries of each failing metric (each links down to its full reasoning), and a key signals strip with the deterministic numbers operators scan first (latency, talk ratio, tool-call success).

Below that is the per-call metric grid, 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).

If a call was never scored, the panel shows "Not scored yet" with an Evaluate metrics action that runs the scoring engine on demand. You can re-run scoring at any time, and you can leave a thumbs-down vote on a metric verdict you disagree with, which feeds back into tuning that metric. What is scored, and how, is configured on the agent's Scoring page and covered in Evals.

There is also a compare control: pick another recent call and view the two side by side to see what changed between them.

Using transcripts in practice

Transcripts are the input to everything else:

• Find a failure. When a call goes wrong, open it and read it back to see exactly where: 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 change the agent and publish, watch new calls. The compare view is useful here: line up a call from before the change against one from after.

Overview (analytics)

The Overview tab is a dashboard grid of charts, scoped to this agent. It answers the across-many-calls questions a transcript can't: how busy is the agent, how long are calls running, how often do they succeed, and how fast is it responding.

Each widget is a chart built from one data field (duration, success, outcome, sentiment, latency, topic, call-ended reason, and others), bucketed by a group (day, week, hour, agent, outcome, sentiment, topic, and others) and reduced by an aggregation (count, average, sum, median p50, p95). Chart types include line, bar, stacked bar, pie, KPI tiles and lists. You can add a widget yourself or apply a starter dashboard that ships a themed set in one click, for example:

• Operations: call volume over time, by hour and by agent, plus average duration and top topics.
• Quality and outcomes: success rate and its trend, outcome mix, sentiment, drop-off points and review coverage.
• Performance and latency: median (p50) and p95 latency with average call duration over time.
• Executive at-a-glance: the headline numbers (total calls, average duration, median latency) with a weekly trend.

A source filter on the control bar switches the whole grid between production, test, eval and all calls, so trends from real callers stay separate from your own test runs.

The quality metrics that decide whether calls feel right (response latency and time-to-first-audio, talking over each other, and how often fillers cover a wait) are measured by the runtime on every call. Latency surfaces directly as a chartable field here (the Performance and latency dashboard plots its p50 and p95), and the full breakdown, including the per-stage latency waterfall, is on each call's detail and the evals scoring views.

Insights

The Insights tab (in beta) takes the question "where does this agent fail?" and answers it without you reading every call. It clusters this agent's failing calls into recurring failure modes, scoped to one agent, in two views:

• Failure summary: a written-up read on the past 7 days of calls, naming the top failure modes and what to change. Start here.
• Failures by metric: one card per metric, with the past 24 hours of failing calls grouped into the patterns behind each metric.

Insights are generated on demand: press Generate (or Deep Research for the 7-day write-up) and the job runs, then the clusters appear. Only metrics on your workspace rubric appear here, so what shows up depends on what you have chosen to measure. A Fix with AI control hands the top failure modes to the agent's AI assistant so it can draft prompt and tool changes that address them.

The improvement loop

Transcripts, Overview and Insights all feed the same loop, and they map onto the agent's prime directive of being reliable first:

1. Find the weakest behaviour. Read a bad call in Transcripts, or let Insights cluster the recurring failures for you, or spot a sagging trend on the Overview.
2. Write an eval for it so the failure is pinned and you can prove the fix. New behaviour starts as a failing test.
3. Fix the cause: tighten the prompt or a tool description (tools first, prompt last), then publish.
4. Watch the new calls. Confirm the fixed behaviour holds in fresh transcripts and that the trend on the Overview moves the right way.

Reliability is built one fixed failure at a time. Read transcripts to understand a single call; watch the Overview and Insights to know whether the agent is improving across all of them.

Related

• Evals: the rubric, scoring and the test suites the per-call metrics come from.
• Voice and fillers: latency, turn-taking and the fillers that mask a wait.
• Prompt and Tools: the two things you usually change after reading a bad call.
• Privacy: redaction, recording and the spoken disclosure.
• Agents: the agent and how publishing works.