Customer service performance indicators describe different kinds of evidence. A satisfaction response reports a customer’s perception. A handling duration measures contact work. A quality review assesses an interaction against a rubric. Retention records what a customer cohort did over time.
These measures are useful together, but they should not be treated as interchangeable proof of individual performance or business impact. This guide explains eight core indicators, their formulas, the decisions they support, and the limits of attributing their results to a representative or service team.
All examples are illustrative, not company results or industry benchmarks.
Match the indicator to the level you can evaluate
| Level | Useful evidence | Interpretation limit |
|---|---|---|
| Interaction or representative | Relevant messages, reviewed behaviors, solution accuracy and clearly attributed work | One agent may not control the full case or customer outcome |
| Queue or team | Response timing, contact handling, eligible resolution cohorts, workload | Case mix, staffing, routing and dependencies affect comparisons |
| Task or customer journey | Effort feedback, repeated steps, validated completion and handoffs | Several channels or teams may contribute |
| Relationship or business | NPS, retained customer cohorts and customer reasons | Product, price, reliability and market conditions also matter |
Use individual evidence for specific coaching observations and team/business evidence to identify patterns for investigation. Avoid attaching a quarterly retention movement to one representative’s score.
For each indicator, state the population, source, reporting period, denominator, clock, exclusions, and missing-data policy. Preserve stable issue, ticket, contact, and customer IDs. Counts and coverage should travel with percentages.
1. Customer satisfaction score (CSAT)
What it indicates: how respondents felt about the named experience. Choose a clear question about a completed support interaction, task, product, or relationship; those scopes should not be mixed.
Five-point top-two CSAT = ratings of 4 or 5 ÷ all valid responses × 100. Neutral and dissatisfied responses remain in the denominator. Qualtrics explains this calculation and the limitations of respondent bias.
Show response counts and delivered-invitation coverage, along with the question and timing. Eighteen satisfied ratings among 20 valid responses give 90% CSAT. If 100 invitations were delivered, response coverage is 20%; the other 80 invitees’ satisfaction is unknown.
How to use it: inspect comments and linked cases to identify answer quality, expectations, product faults, handoffs, or policy concerns. Keep similar issue populations together before comparing teams. A customer score alone does not establish an agent error or long-term loyalty.
2. Net promoter score (NPS)
What it indicates: willingness to recommend the company, product, or service specified in a 0–10 question. The Bain definition groups respondents as promoters at 9–10, passives at 7–8, and detractors at 0–6.
NPS = percentage of promoters − percentage of detractors. Include all valid respondents in the denominator. Thirty promoters, ten passives, and ten detractors among 50 responses produce 60% − 20% = an NPS of 40. This is a score, not 40% satisfaction.
How to use it: ask for reasons, categorize themes, and involve the team that can address each one. Keep relationship surveys separate from surveys following support. Different scores between those populations do not isolate support’s contribution.
NPS can support relationship investigation, but it does not guarantee referrals, retention, or revenue growth. For mandatory internal services, first decide whether recommendation is a meaningful question; a differently worded usefulness measure should not be labeled NPS.
3. First contact resolution (FCR)
What it indicates: whether the first qualifying contact delivered a working resolution under an explicit validation policy.
Outcome FCR = eligible issues resolved during the first contact and validated through the chosen repeat window ÷ all eligible first-contact issues with a complete window × 100. Keep unresolved eligible cases in the denominator, and show pending validation separately.
Define a contact per channel, same-issue return matching, transfers, exclusions, and resolution evidence. A continuous interaction may involve an internal transfer without becoming a second customer contact under your definition; report no-transfer resolution separately if needed.
ICMI’s guidance recommends a designated repeat window, quality monitoring, and customer feedback. Allow customers time to test the solution. A single agent reply or solved status is not automatically validated FCR.
How to use it: review failure reasons by issue category: diagnosis, knowledge, permission, missing information, dependency, or necessary specialist work. Low FCR does not establish poor agent knowledge without that context. See the FCR improvement guide for targeted changes.
4. Average handle time (AHT)
What it indicates: the recorded work needed to handle a contact, useful for process analysis and capacity planning.
Voice AHT = total included talk, hold, and after-contact work time ÷ included handled calls. Amazon Connect documents these components, with task-specific pause behavior. Check provider fields and included records.
If 40 calls contain 200 talk minutes, 40 hold minutes, and 80 after-contact minutes, AHT is 320 ÷ 40 = 8 minutes. That says nothing by itself about whether the issues were fully resolved.
How to use it: compare similar issue types and inspect diagnosis, lookup, administration, and necessary documentation. For email or concurrent chat, define active work separately; calendar ticket duration and overlapping session time are not automatically labor time.
Pair handling changes with FCR, quality, and feedback. Shorter AHT may reflect simpler cases or less repetitive work, but it may also accompany incomplete service. Avoid ranking representatives across materially different case mixes.
5. Customer effort score (CES)
What it indicates: perceived effort or ease for a named customer task. Qualtrics describes CES as complementary experience evidence, which benefits from follow-up feedback.
Publish the question, scale anchors, and aggregation. For a local scale from 1, very difficult, to 7, very easy, mean ease score = sum of valid ratings ÷ valid responses. Higher means easier under those anchors. An effort-intensity question may reverse the direction.
For ratings of 7, 6, 5, and 2, the mean is 5. The two lower ratings deserve investigation even though the average looks less concerning.
How to use it: inspect repeated information requests, unclear steps, failed self-service, transfers, and handoffs. Ask which part could have been easier. An easy process does not establish an accurate solution, and CES is not a guaranteed forecast of churn.
6. Average response time
What it indicates: how long customers waited for the defined reply. First response and subsequent response timing answer different questions, so label them separately.
Mean first response = sum of eligible request-to-first-qualifying-reply durations ÷ requests with that reply. State the start event, qualifying reply, channel, clock, and reporting period. Keep unanswered requests and their age visible.
For a separately chosen ongoing-reply measure, define a customer turn and the next qualifying service reply. Then calculate sum of eligible turn-to-next-reply durations ÷ completed eligible reply cycles. Group consecutive customer messages consistently and show turns still awaiting a reply. This is a local reporting definition, not a claim that every platform uses it.
Zendesk’s documentation shows differences between native reply fields, clocks, channel behavior, and SLA rules. Automatic acknowledgment can set expectations without satisfying a local first-human-response policy.
How to use it: inspect routing, queue coverage, arrival patterns, priority, and ownership. Show mean, median, tail, and maximum where useful, with unanswered work separate. Speed is not evidence that a reply was relevant or that an issue was resolved. Use the response-time guide for queue improvements.
7. Customer retention rate
What it indicates: whether customers from a defined starting cohort remain active at a chosen endpoint. Define the customer unit, active status, period, and reactivation policy.
End-activity retention = starting-cohort customers still active at period end ÷ customers active at period start × 100. This counts starting customers who reactivate before the endpoint; uninterrupted retention requires a separate continuous-status rule.
Suppose 100 customers start the period and 90 of those IDs remain active at the end. Retention is 90%. Twenty new customers join, but only 15 remain active at the endpoint, so total active customers end at 105. Subtracting all 20 acquisitions from 105 would incorrectly give 85%; direct cohort matching avoids that mistake.
Customer-count retention differs from revenue retention. Stripe’s cohort documentation separates retained subscribers and retained revenue, which can exceed 100% with expansion.
How to use it: investigate relevant customer cohorts, reasons, product experience, pricing, billing, and service history. Retention is a wider business outcome; correlation with support contacts does not prove service caused churn or renewal. Use the loyalty measurement guide for relationship context.
8. Quality assurance score
What it indicates: how reviewed work met a documented evaluation rubric. It is reviewer-assessed evidence, not a customer satisfaction score or a measurement of every interaction by default.
A local point-based review formula is awarded applicable points ÷ maximum applicable points × 100. For weighted criteria, publish the weights and normalization. Exclude genuinely inapplicable criteria from both parts, distinguish missing evidence from N/A, and report defined critical failures separately or under an explicit override policy.
Zendesk documents weighted review calculations, N/A exclusions, and critical-group/whole-review behavior. Provider settings matter; the local points convention here is not every tool’s formula.
A reviewed interaction with 44 awarded points out of 50 applicable points scores 88%. That one score does not establish team quality. State the number of reviewed cases, selection method, rubric version, reviewer, and evaluation scope.
How to use it: inspect criterion findings for accuracy, diagnosis, clear communication, and next steps. Zendesk distinguishes case-level and message-level attribution; assign feedback to the relevant work when several agents contributed.
Calibrate reviewers on the same cases and compare evidence, as described in Zendesk’s calibration guidance. A targeted complaint sample is useful for investigation but not automatically representative. The QA program guide covers review selection and coaching in more detail.
Turn indicators into a defensible performance review
Read the evidence together without collapsing it into a causal story. If response time improves but validated FCR falls, inspect whether replies became less complete or whether case complexity changed. If quality scores are high but effort feedback is poor, investigate the journey between interactions as well as the rubric’s scope.
Check data coverage first, then case mix, then source records and customer reasons. Select a specific action and a suitable evaluation period. Training or a new tool may be a plausible response, but completing rollout is not evidence of improved customer outcomes.
Use comparable cohorts and preserve definition versions. Allow repeat and retention windows to complete. For individual reviews, rely on clearly attributed behaviors and adequate relevant evidence, with a correction route for factual errors.
For a practical meeting artifact, use the filled customer service scorecard. For targets, ownership, and action policies, continue with the support KPI guide. These indicators provide evidence for decisions when their scope and limits remain visible.
