Call history, analytics & simulation
Call history
Section titled “Call history”Every call (phone, web, or test) produces a call record: transcript, duration, end reason, and — when the flow ran on the runtime’s v2 instrumentation — richer per-turn data: which node each turn belongs to, per-turn end-to-end latency, and a timeline of variable read/writes during the call.
Explore view
Section titled “Explore view”Opening a call’s Explore view shows:
- Path graph — a mini rendering of the flow that ran, with the actual path the call took highlighted (joined against the exact flow version’s node layout, so the path stays correct even if the flow has since been edited).
- Node messages — the transcript grouped by which node produced each turn.
- Variables timeline — every variable read or write during the call, in order.
- Latency strip — per-turn end-to-end latency.
Audio intelligence
Section titled “Audio intelligence”A call’s detail view also shows a post-call analysis pass when available: a summary, a topic, sentiment at the start and end of the call, whether the caller’s goal was reached, and any out-of-scope requests the caller made. If a call’s flow enabled the end-of-call survey, the survey rating and feedback appear here too.
Calls made before this instrumentation existed, or through paths that don’t populate every field, show a clear “not captured for this call” state rather than a misleading blank — Explore and audio intelligence degrade gracefully per field.
Analytics
Section titled “Analytics”The project Analytics page aggregates calls over a date range: call volume, duration, drop-off rate, and the Journeys tab, which aggregates the same per-call path data behind Explore into a flow-wide heatmap (edge thickness = traffic volume, node color = drop-off rate) plus a top-paths table and a drop-off ranking — the fastest way to see where callers abandon a flow. Survey results (response rate, average rating, feedback list) are also aggregated here when the survey is enabled.
Simulation
Section titled “Simulation”Before pointing real traffic at a flow, you can simulate it: a batch of AI-generated personas converse with the flow end-to-end, an LLM judge scores each conversation against the flow’s intent, and a heatmap (the same shared component as Analytics → Journeys) shows where simulated callers diverged or dropped off.
Launch a simulation batch from a flow’s Test stage using a preset:
| Preset | Use for |
|---|---|
| Quick | A fast sanity check with a small persona count. |
| Standard | The default size/turn-count for regular pre-release checks. |
| Deep | The most thorough pass — includes a “turbo” reasoning mode for the simulated personas. |
Raw persona-count/turn/timeout knobs beyond the presets are available to organization administrators only.
After a batch completes, its findings — the judge’s report and the specific failing paths — can be handed directly to the flow editor’s AI assistant (“Improve flow with AI”) to iterate on the flow with that context already loaded.