Track customer outcomes, operational effort, and model-derived signals separately, with consistent definitions and coverage shown beside every metric. Start with the decision the dashboard must support. A daily supervisor view, a coaching plan, and a product team's root-cause review need different levels of detail.
Level AI describes survey and conversation insights with inferred satisfaction. AWS documents separate talk, interruption, and non-talk conversation metrics. These capabilities are inputs to a measurement design, not proof that a single score explains customer experience.
Which categories of evidence belong together?
Use customer feedback to understand what customers report, operational records to understand what happened, and model-derived signals to prioritize investigation. Give each category its own label. For example, a predicted satisfaction value should not be presented as though the customer submitted a survey.
Define your unit of analysis before joining data. A contact, a case, and an entire customer journey are different units. One issue may create multiple calls or chats; treating each as an independent success can hide a recurring resolution problem.
What should a practical metric dictionary contain?
| Metric family | Definition to record | Coverage to display |
|---|---|---|
| Customer feedback | Question, scale, timing, and calculation | Eligible contacts and submitted responses |
| Resolution | What counts as closed and repeat work | Cases with a reliable outcome |
| Effort and time | Included work and timer boundaries | Contacts with complete timing records |
| Quality | Rubric version and critical-error rules | Reviewed or scored eligible contacts |
| Inferred signals | Model, label, and validation approach | Analyzed contacts and unsupported cases |
Assign an owner to the dictionary and retain prior versions. Annotate a chart when the survey, scoring rubric, model, or data integration changes. Otherwise a technical change can look like a sudden customer-experience improvement or decline.
How do you turn a dashboard into an action?
For each metric, write the decision, the responsible team, and the evidence needed before acting. An increase in a contact reason might trigger a review of representative conversations and product incidents. It should not automatically become a conclusion about agent behavior.
Compare like populations and show counts. Separate new agents from experienced agents when onboarding is the question, and separate channels when their timing works differently. Keep inaccessible or unanalyzed contacts visible as a coverage gap rather than silently dropping them.
Which follow-up questions should a pilot answer?
Ask whether the intended work changed, whether customers reached resolution, and whether the effect persists across relevant segments. Inspect contradictory evidence: improved inferred sentiment with more reopened cases deserves investigation, not a favorable headline.
Use the handle-time and satisfaction guide for a balanced pilot scorecard. For turning themes into a research workflow, see evaluating Voice of the Customer platforms.
How should CSAT, NPS and inferred satisfaction appear on the same dashboard?
Qualtrics defines a common five-point CSAT calculation as the percentage of responses rated four or five. NPS uses a recommendation question: subtract the percentage rating zero to six from the percentage rating nine or ten. Keep the question wording, invitation population and survey timing beside each result.
| Illustrative result | Calculation and coverage | Decision it can support |
|---|---|---|
| Survey CSAT: 80% | 80 satisfied responses of 100 received; 500 eligible invitations, so 20% response coverage | Investigate comments and service outcomes for the respondents |
| NPS: +30 | 50 promoters minus 20 detractors among 100 respondents; 30 passives remain in the denominator | Investigate reasons for recommendation within the measured population |
| Model-inferred satisfaction | Report model version, output scale, eligible/processed counts and validated examples separately | Prioritize conversations for review; never fill missing surveys with predictions |
These invented numbers explain arithmetic, not observed customer outcomes. To act on a low score, inspect representative conversations, distinguish product defects from agent behavior, assign the responsible team and test one change. Measure whether that mechanism changed the intended task outcome and retain survey feedback as an independent check. Analytics software can identify evidence for action; its presence does not establish a CSAT or NPS gain.
What are the limits of this guide?
Sources were checked October 20, 2026. This is an editorial measurement framework, not a validated industry benchmark or a tested product comparison. Platform calculations, channel support, joins between contacts and cases, and model coverage need local confirmation. Aggregated trends cannot establish why an individual customer felt a certain way; review the underlying evidence and preserve uncertainty when the available records do not resolve it.