Interpret metrics, denominators, and evidence
Use Conversation Intelligence metrics without confusing unavailable data, coverage, or AI detections with facts.
3 min read
- Project
- Insights
- Analysis
Metrics and coverage appear in Insights for the current project, date range, and assistant scope. Check those controls before comparing values.
- Pending or failed analyses
- Historical envelopes missing the required field
- Conversations outside the active filters
Read it as: Use the displayed scope and supporting conversations to assess coverage; do not infer causation or a population-wide truth.
Core metrics
| Metric | Numerator | Denominator | Notes |
|---|---|---|---|
| Resolved rate | resolved + partially_resolved analyzed conversations | analyzed conversations with a resolution breakdown | Pending is shown separately. This is an AI classification, not a satisfaction score. |
| Knowledge-gap rate | successful analyses with one or more gaps | all successful analyses in scope | Gap-reason breakdowns may exclude historical records that lack the gap field. |
| Misunderstanding rate | conversations with one or more misunderstood_answers | analyzed conversations containing that field | A possible misunderstanding needs transcript review. |
| Budget/timeline rate | conversations with a non-empty mention list | analyzed conversations containing that specific field | Each field has its own eligible denominator. |
| Objections | conversations with an explicit objection | no displayed rate in the current panel | Review the underlying wording. |
| Alternatives | conversations mentioning any alternative | analyzed conversations containing the alternatives field | A named alternative is ranked only after at least two distinct conversations. |
For a synthetic period with 18 successfully analyzed conversations that include budget_mentions, and 3 with a non-empty list, the budget rate is 3 / 18 = 16.7%. If 5 older successful analyses lack that field, they are excluded—not assumed to contain no budget mention.
How to read patterns and evidence
Conversation insights aggregate AI-detected sentiment trend and shift, top topics, primary intents, language, and buying-intent distribution. Resolution quality also lists top unanswered topics. These are descriptive counts over the selected analyzed scope; they do not prove representative customer demographics, intent, conversion likelihood, or causal impact.
Signal panels show up to three recent evidence items. Evidence is a starting sample, not the full population. For alternatives, normalization removes presentation noise such as case and surrounding punctuation but does not perform fuzzy entity matching. Similar names can remain separate.
Limits to keep in mind
- Do not divide by all conversations unless the interface explicitly defines that denominator.
- Zero can mean no detections among eligible records;
N/Ameans there was no eligible denominator. Neither establishes recall or accuracy. - Percentages over a small eligible population are volatile. Avoid broad decisions from one or two conversations.
Last updated on