Analytics

Totals, trends and breakdowns over your calls, how each number is counted, and the drop-off funnel for flow agents.

Analytics in the dashboard answers "how are my agents doing?" for any date range: how many calls, how many succeeded, what they cost, how quickly the agent answered. Anyone in your organization can read them, and they are computed from the same calls the call list shows, when you ask, so they always agree with it.

Filter
Date range	A date or a date and time. An end date includes that whole day. Default: the 30 days up to now. At most 366 days.
Time zone	The time zone days, weeks and months are cut in. Default UTC. A call at 22:30 on the 14th in Chicago is on the 14th, not the 15th.
Group by	Hour (up to 14 days), day, week (Monday to Sunday) or month.
Agent, direction	Narrow to one agent, or to inbound or outbound.

The page shows totals, a chart with one point per bucket that has calls, and breakdowns (how calls ended, direction, and your ten busiest agents).

How each number is counted

Calls is every call that started in the range, whatever its status. Failed is calls whose status is failed. A call still in progress is neither completed nor failed.
Minutes is the sum of call durations divided by 60. Cost is the sum of each call's call_cost, in units of $0.0001 (cost_usd has it in dollars).
Success rate is calls where the post-call analysis said the call was successful, out of calls where it recorded a verdict at all. A call with no verdict (analysis not run, failed, or removed by the agent's data-retention setting) is not in the denominator, and evaluated tells you how many are.
Sentiment is the analysis's user sentiment. Calls without one count as unknown.
Response time is, for each call, the average time from the caller finishing to the agent's first audio, over its turns. p50, p90 and p99 are taken across calls: p90 is the response time of a call slower than 90% of the others. Each stage (llm, tts, endpointing) is the median across calls of that call's own median and 90th percentile for the stage.

A bucket with nothing to average (no verdicts, no timed calls) has a null rate rather than a zero, so a quiet hour is not shown as a failing one.

Where flow calls go

For an agent that runs a conversation flow, the flow funnel shows, for every node, how many calls reached it and how many ended on it. A call is counted once per node however many times it looped through. Node names come from the agent's current flow; a node that has since been removed is still listed, marked in_current_flow: false.

The path is stored when a call ends, so calls that ended before this feature was added have none and are not counted. The path holds only node and edge ids, nothing the caller or agent said.