Observe

Metric definitions

Every metric in Analytics and Boards, exactly as it is calculated. This list is generated from the analytics engine itself, so it always matches what you see.

How results are decided

Successful — a call that did not fail and either passed at least half of its scorecard criteria, converted, or (with no scorecard) was analysed with no misunderstanding, no repetition and no negative sentiment. Calls not yet analysed count as unknown, not as failures.
Converted — the outcome is "booked" or "interested", or a scorecard criterion about booking, sales or sign-up passed.
Outcome — the lead disposition when a campaign analysed the call; otherwise "transferred", "booked" (a booking criterion passed), "failed", "short call" (under 15 seconds) or "completed".
Failure reasons — every reason a call fell short: a failed call or agent error, each failed scorecard criterion, the caller not being understood, the agent repeating itself, negative sentiment, a short call, frequent interruptions (3 or more), slow responses (over 3 seconds on average), a fallback reply, a failed tool call, or poor audio. One call can have several.

Calls

Total calls (calls, count) — Number of phone calls in the range.
Call duration (duration, time) — Length of the call from connect to hang-up.
Call minutes (talk_minutes, time) — Total connected time across calls.
Turns per call (turns, number) — Caller plus agent turns in the transcript.
Success rate (success_rate, percentage) — A call is successful when it did not fail and either passed at least half of its scorecard criteria, converted, or (with no scorecard) was analysed with no misunderstanding, no repetition and no negative sentiment. Calls not yet analysed count as unknown, not as failures. Higher is better. Not every call carries it — charts say how many do.
Conversion rate (conversion_rate, percentage) — A call converts when its outcome is 'booked' or 'interested', or a scorecard criterion about booking, sales or sign-up passed. Higher is better.
Conversions (conversions, count) — A call converts when its outcome is 'booked' or 'interested', or a scorecard criterion about booking, sales or sign-up passed. Higher is better.
Transfer rate (transfer_rate, percentage) — Share of calls the agent handed to a person.
Error rate (error_rate, percentage) — Share of calls that failed, or where the agent's model returned an error. Lower is better.
Evaluation score (eval_score, percentage) — Share of the agent's enabled scorecard criteria that the call passed (0–100%). Higher is better. Not every call carries it — charts say how many do.
Task completion rate (task_completion_rate, percentage) — Share of analysed calls that passed every scorecard criterion (or converted, when the agent has no scorecard). Higher is better. Not every call carries it — charts say how many do.
Negative sentiment (negative_sentiment_rate, percentage) — Share of analysed calls where the caller's sentiment was negative. Lower is better. Not every call carries it — charts say how many do.
Short-call rate (short_call_rate, percentage) — Share of calls under 15 seconds — usually hang-ups or wrong numbers. Lower is better.
Unique customers (unique_customers, count) — Distinct phone numbers that called or were called.
Repeat-caller rate (repeat_customer_rate, percentage) — Share of customers with more than one call in the range.
Total cost (cost, money) — Usage cost attributed to the call: telephony, speech-to-text, voice and model. Lower is better.
Cost per call (cost_per_call, money) — Average attributed cost of one call. Lower is better.
Cost per conversion (cost_per_conversion, money) — Total cost divided by the number of conversions. Lower is better.
Tokens used (tokens, tokens) — Model input plus output tokens attributed to the call.
Time to first response (first_response_ms, time) — From the call connecting to the first audio of the agent's greeting. Lower is better. Not every call carries it — charts say how many do.
Avg response latency (response_latency, time) — Time from the caller finishing a sentence to the first audio of the agent's reply, measured by the phone bridge. Tracked on calls from the release that added it; older calls show no value. Lower is better. Not every call carries it — charts say how many do.
p95 response latency (p95_response_latency, time) — The slowest 5% of the agent's replies within a call. Lower is better. Not every call carries it — charts say how many do.
Interruptions (interruptions, count) — Times the caller spoke over the agent and the agent stopped (barge-in). Lower is better. Not every call carries it — charts say how many do.
Interruption rate (interruption_rate, percentage) — Share of the agent's turns that the caller talked over. Lower is better. Not every call carries it — charts say how many do.
Caller talk time (caller_talk_ms, time) — Time the caller spent speaking, from the speech-to-text timings. Not every call carries it — charts say how many do.
Agent talk time (agent_talk_ms, time) — Time the agent's audio was playing to the caller. Not every call carries it — charts say how many do.
Silence (silence_ms, time) — Connected time when neither side was speaking. Lower is better. Not every call carries it — charts say how many do.
Tool-call success rate (tool_success_rate, percentage) — Share of the agent's tool calls (bookings, lookups…) that succeeded. Higher is better. Not every call carries it — charts say how many do.
Tool calls (tool_calls, count) — Tool calls the agent made. Not every call carries it — charts say how many do.
Fallback rate (fallback_rate, percentage) — Share of calls where the agent had to use a fallback reply because its normal pipeline failed. Lower is better. Not every call carries it — charts say how many do.
Speech recognition confidence (stt_confidence, percentage) — Average speech-to-text confidence on the caller's words. Higher is better. Not every call carries it — charts say how many do.

Usage & costs

Spend (spend, money) — Usage cost across every surface — calls, web chat, tests and tools. Lower is better.
Billable events (events, count) — Metered usage events.
Tokens (event_tokens, tokens) — Model input plus output tokens.
Failed events (event_error_rate, percentage) — Share of metered events that failed. Lower is better.

Model turns

Model turn time (turn_ms, time) — Server time for one agent reply: retrieval, memory, integrations and the model. Lower is better.
Model time (model_ms, time) — Time spent in the language model within a turn. Lower is better.
Knowledge lookup time (knowledge_ms, time) — Time spent searching the knowledge base within a turn. Lower is better.
Tool time (tool_ms, time) — Time spent running tools within a turn. Lower is better.
Model turns (turn_count, count) — Agent replies traced.

Groupings

Charts and board widgets can group by:

Date (date) — Calls, Usage & costs, Model turns
Agent (agent) — Calls, Usage & costs, Model turns
Call status (status) — Calls
Outcome (outcome) — Calls
Customer (customer) — Calls
Channel (channel) — Calls, Usage & costs, Model turns
Direction (direction) — Calls
Sentiment (sentiment) — Calls
End reason (endReason) — Calls
Hour of day (hourOfDay) — Calls, Usage & costs, Model turns
Weekday (weekday) — Calls, Usage & costs, Model turns
Failure reason (failureReason) — Calls
Result (result) — Calls
Provider (provider) — Usage & costs
Model (model) — Usage & costs
Usage type (eventType) — Usage & costs

Aggregations

Count — how many calls (or events) have the value.
Average, Sum, Minimum, Maximum — over the calls that have the value. Calls without it are left out, never counted as zero.
Rates (success rate, conversion rate…) are averages of yes/no, shown as percentages. Ratios such as cost per conversion and interruption rate divide two totals.

Next steps

← Analytics

Integrations & BYO keys →