Skip to main content

Observability

Metrics

The numbers that tell you whether it is working.

What Metrics Means Here

Metrics are aggregates computed over completed calls. Per-call records are the raw material — see call logs and post-call analysis. This page covers the numbers you read to decide whether a Finn is working, and the ways those numbers mislead you.

The Four Headline Numbers

MetricDefinitionWhat a bad value means
Success rateCalls achieving the goal / calls that connected and completedFinn behaviour problem: prompt, guardrails, or closing
Completion rateCalls reaching a real human and getting through the conversation / calls attemptedList quality or calling hours. Not a Finn problem
Average durationMean length of completed callsUnder 20s, Finn getting hung up on or terminating early. Over 5min, Finn rambling or not closing
Cost per successful outcomeTotal cost / count of successful callsThe economic truth. Everything else is diagnostic

Success rate is defined by you, not by the platform. It is computed against the primary success field you defined in post-call analysis. If your primary success field is appointment_booked, success rate is the fraction of connected-and-completed calls where appointment_booked = true. Change that field definition and every historical success rate on the page changes meaning. See analysis fields.

Success rate and completion rate share no denominator. Completion rate counts attempts; success rate counts only calls that connected and completed. A campaign can post 90% success on 8% completion and still be a failure. Read them together.

Where Each View Lives

Metrics are read in the dashboard. There is no metrics aggregate endpoint.

PageScope
Dashboard (home)All deployments, high level
AnalyticsCross-deployment trends, time series, leaderboards
Deployment AnalyticsOne deployment, with per-call drill-in
Finn AnalyticsOne Finn across all its deployments
Call HistoryEvery call, searchable
Live DeploymentsReal-time counts and active calls

To get metrics into your own systems, use the per-deployment CSV or the PDF report (exports), or stream completed-call events to your endpoint (post-call webhook).

Distributions, Not Just Averages

Each post-call analysis field gets a distribution chart, typed by the field:

Field typeChart
BooleanBar chart of true/false
EnumStacked bar of categories
StringWord cloud plus a sample of five
NumberHistogram plus median

The string sample-of-five is a sample. Do not treat it as the range of answers.

Where These Numbers Go Wrong

Live numbers move. While a deployment is running, success rate is computed against calls completed so far. It fluctuates as more calls finish. Wait for the deployment to complete, or for at least 100 calls, before concluding anything.

Small denominators lie. Success rate on 12 calls is noise. The 100-call threshold is a floor for reading a number at all, not a guarantee of significance.

Aggregates are cached. The dashboard shows a Cached badge when the data is not freshly computed. Click Refresh to force a recompute.

DataCache TTL
Deployment list5 minutes
Per-deployment summary1 minute
Per-call recordsNever cached, always live

If a number you just changed something to affect has not moved, check the badge before assuming the change did nothing.

Drift is usually the list, not the Finn. Gradual degradation in success rate over weeks most often means the audience is going stale. The Finn config did not change. Check call outcomes and completion rate before rewriting prompts.

Editing a transcript edits the record. Correcting a transcript line updates the analytics record it feeds. Numbers computed before the edit will not match numbers computed after.

Reading a Bad Campaign

Work in this order. Each step rules out a cause before you spend effort on the next.

  1. Check completion rate first. Low completion with healthy success-among-completed is a list or calling-hours problem, and no prompt change fixes it.
  2. Check average duration. Very short means hangups early, so audit the welcome message.
  3. Check whether any post-call analysis field is flagged as extraction-failed disproportionately. That is a transcript or instruction problem, not an outcome problem.
  4. Sort the call table by sentiment ascending and listen to the worst five. Patterns usually show within three calls.
  5. Diff against a known-good Finn. Run the same audience on a baseline Finn for 50 calls and compare the same four numbers.

Do not try to read every call. A working campaign produces thousands. Read the headline numbers, then the ten best and ten worst calls.