Customer service KPIs should show whether your service delivers the experience and commitments you intend. A short response time cannot answer whether the issue was resolved. A satisfaction score cannot tell you whether a contractual deadline was met.
This guide covers eight KPIs from three perspectives: customer feedback, service execution and continuing customer relationships. It then adds three operational measures for a helpdesk scorecard: resolution time, ticket volume and deflection, and agent utilization. Use them as a portfolio with complementary checks, rather than independent targets that every agent must maximize.
For a broader measurement list, see key metrics for your customer support team. For ownership and review rules once you choose a KPI, use the KPI specification below.
Eight customer service KPIs and the promise each tests
| KPI | Question it helps answer | Useful companion check |
|---|---|---|
| 1. CSAT | How satisfied were respondents with this experience? | Comments and response coverage |
| 2. NPS | How likely are respondents to recommend us? | Reasons, customer segment and relationship context |
| 3. First-contact resolution | Did the first contact resolve the eligible issue? | Repeat contacts and outcome validation |
| 4. Average first response time | How long did answered requests wait for a qualifying first reply? | Unanswered requests and response quality |
| 5. Customer effort | How easy or difficult was the specific task? | Journey friction and comments |
| 6. Customer retention | Which customers in the starting cohort remain? | Customer reasons and product/commercial changes |
| 7. Average handle time | How much measured contact work is required? | Issue complexity, quality and repeated work |
| 8. SLA compliance | Which applicable service targets were achieved? | Active deadlines, breaches and policy coverage |
1. Customer Satisfaction Score: satisfaction with an experience
CSAT asks about a defined interaction, product or service experience. Keep the wording and timing consistent enough to compare like experiences. An optional comment can explain what the rating refers to.
For a five-point satisfaction scale, count ratings 4 and 5 as satisfied. CSAT = satisfied valid responses ÷ all valid responses × 100. This top-two calculation is documented by Qualtrics.
Report the response count and invitation coverage. A rating may reflect the product, policy or outcome as well as the agent’s behavior. Review the conversation before assigning a coaching conclusion. Investigate negative feedback through the approved follow-up process, without offering benefits in exchange for a better score.
Use CSAT to check a service improvement at the relevant touchpoint; do not treat it as a census of everyone who received help. Report delivered invitations beside valid responses, segment by issue type before comparing teams and avoid rewarding agents for selectively sending surveys.
2. Net Promoter Score: recommendation sentiment
NPS uses a 0–10 likelihood-to-recommend question. Promoters score 9–10, passives 7–8 and detractors 0–6. NPS = percentage of promoters − percentage of detractors, with all valid responses, including passives, in the denominator. It ranges from −100 to +100 and is reported as a score, not a percent satisfied.
These scoring rules come from Bain’s Net Promoter System. Ask for the reason behind the rating, and state whether the survey concerns the whole relationship or a particular touchpoint.
A relationship NPS result is not an individual support-agent rating or proof of future growth. Product reliability, price and other experiences can affect the answer. Use it to identify issues for cross-functional review, and compare trends only with consistent audience and survey design.
A survey sent after support measures a different context from a periodic relationship survey. Do not subtract one score from the other and call the difference support’s causal contribution.
3. First Contact Resolution: a completed outcome
FCR measures eligible issues resolved during the first contact without another contact for the same unresolved need. Define what a contact means for each channel, how transfers count and how related conversations are linked.
For an outcome-based local definition, FCR = validated first-contact resolutions ÷ all eligible first contacts with a completed validation window × 100. Keep eligible unresolved cases in the denominator. Do not count a one-reply ticket or a closed status as automatic evidence of resolution.
ICMI’s FCR guidance discusses repeat-contact tracking, enough time for the customer to try the answer, quality review and customer feedback. Choose the validation window around the task, rather than applying an arbitrary delay to every issue.
Treat cases with an incomplete observation window as pending validation. If outcome evidence is missing, apply a documented rule and show the missing share rather than counting those cases as successes.
Group failed first contacts by knowledge, permissions, missing information, product defects and necessary specialist work before deciding on training, knowledge improvements or a change in agent authority. Some approvals remain necessary. Native one-touch ticket measures are useful proxies, not validated outcome FCR. Pair FCR with quality checks so the target rewards a verified outcome.
4. Average Response Time: the first qualifying reply
Specify whether you mean first response or subsequent response. For a local human-first-reply measure, start at customer request receipt and stop at the first qualifying public human reply. Define included channels and calendar or business hours.
Mean first response time = sum of completed qualifying first-response intervals ÷ number of those intervals. Publish unanswered counts and ages beside it, along with the median and a tail measure such as the 90th percentile. A completed-reply average leaves requests still waiting outside the average.
Check native tool behavior. Zendesk’s first reply metric starts at ticket creation and has channel and legacy automation exceptions. Your intended receipt-to-human definition may require a separate report.
Set a channel-specific expectation based on your service offer, demand and coverage. Automated receipts can set expectations, but they do not qualify for the local human-reply KPI. Review whether first replies advance the customer’s request, alongside speed. When the commitment slips, the queue or channel lead can inspect overdue requests by arrival time, priority, routing and staffing coverage; the response-time improvement guide covers queue changes.
5. Customer Effort Score: ease of the task
CES asks about the effort needed for a particular task, such as obtaining an answer or resolving an issue. Qualtrics describes this task-specific scope.
Choose and label the scale. For an illustrative local 1–7 ease survey, with 1 very difficult and 7 very easy, mean ease = sum of valid ratings ÷ number of valid ratings. Higher means easier. An effort scale with higher meaning harder has the opposite direction; neither should be silently relabeled.
Use comments to locate repeated explanations, channel switching, unclear instructions or approval friction. Verify the suspected bottleneck with case evidence. CES describes respondents’ perceived effort; it does not prove churn will fall after a particular change. Check whether the task was actually completed, because making it easy to end a conversation does not establish resolution.
6. Customer Retention Rate: the starting customers who remain
Define the customer entity and the meaning of retained: an active paid account, renewal or another appropriate business condition. For an end-of-period starting cohort, retention = customers from the starting cohort active at period end ÷ customers in that starting cohort × 100.
Track the same customers directly. Do not add newly acquired customers to the numerator. Document how reactivation and temporary pauses count. Subscriber and revenue retention differ: Stripe’s cohort documentation explains that revenue retention can exceed 100% because existing customers expand.
Review cancellation reasons and support history alongside usage, price, product and account changes. Association with support contact does not establish support caused churn or retention. Customers contacting support may already have more difficult conditions than those who do not.
7. Average Handle Time: measured work per contact
For a defined voice-contact cohort, AHT = (included talk time + hold time + after-contact work) ÷ included handled contacts. Keep the units and counted contact legs consistent. Amazon Connect documents these components and additional channel-specific rules.
Queue time before connection is not agent handle time. A ticket’s elapsed lifecycle is not the same as active work, and simultaneous chats make additive time tracking difficult. Use the platform’s documented method or label a separate work-time measure.
Use AHT to plan workload and identify unnecessary work. Compare similar issues, and review resolution and quality when it changes. A longer contact may reflect complexity or a thorough answer; a shorter one may be efficient or incomplete. There is no universal ideal duration for every service task.
8. SLA Compliance: performance against applicable targets
An SLA policy defines applicable service targets and their measurement rules. Establish the priority, policy, clock, start event, pause conditions and fulfillment event. Uptime and ticket-response promises are separate measures.
For a target-instance report, achievement = achieved targets ÷ (achieved + breached targets) × 100. Zendesk’s documented SLA target metrics use this denominator and distinguish active breached from active unbreached targets. State the date filter and cutoff. Show active unbreached targets separately because their outcomes are not yet known; check for missing or canceled target records.
A ticket may have several targets. Target achievement and the share of tickets meeting every required target are different KPIs. For the latter, define the ticket cohort and handle tickets with outstanding targets explicitly.
Zendesk’s SLA documentation also distinguishes native reply metrics from SLA behavior and documents pause and reopen rules. Validate the actual policy rather than assuming the label means elapsed customer time. Review approaching deadlines operationally, and do not alter an existing commitment merely to improve the reported score.
Three operational KPIs for a helpdesk scorecard
These measures describe the work itself rather than the customer’s view of it. They help with capacity and process decisions, and each needs a quality guardrail.
Average resolution time: the complete issue journey
Mean full resolution time = total creation-to-final-resolution duration for included resolved issues ÷ number of those issues. State the reporting period, clock and endpoint. First and latest resolution differ when a case reopens, and a paused SLA clock, business-hours duration and full calendar wait answer different questions; label them separately.
Examine long waits by stage and issue complexity, then clarify handoff and escalation responsibilities. Pair the KPI with open-case age, same-issue returns and solution accuracy. Closing a ticket early is not an improvement.
Ticket volume and deflection: demand versus successful self-service
Ticket volume = distinct eligible tickets created during the period. State whether several tickets for one issue remain separate. When growth changes the scale of comparison, normalize demand to a relevant population, such as contacts per 1,000 active accounts.
Deflection needs more care because tickets that would have occurred without self-service are not directly observed. A before/after drop can reflect seasonality, product changes, fewer customers or difficulty reaching support. Where feasible, use an observed measure: validated self-service completion rate = eligible attempts with sufficient success evidence and no same-issue assisted contact during the follow-up window ÷ all eligible attempts with a complete window × 100.
The KCS guidance explains why searches, views or sessions cannot be mapped directly to resolved issues. Vendor labels differ too: Intercom’s Fin definitions count customers who leave before an answer or after negative feedback as deflected, so a high figure does not necessarily mean customers succeeded.
Illustrative example: a mature self-service cohort has 50 eligible attempts, with 20 validated successes, 10 known failures and 20 unknown outcomes. Counting only validated successes gives 40% completion, with 40% unknown. Calculating 20 ÷ 30 would hide the unknown share, and neither figure proves how many tickets would otherwise have occurred. Keep a reachable human handoff while you investigate failed searches and unfinished tasks.
Agent utilization and productivity: capacity without shortcuts
Utilization is a time ratio; productivity is output relative to a stated input. For a staffing view, direct-work utilization = included direct support-work minutes ÷ scheduled on-duty minutes × 100. Declare whether documentation and after-contact work count as direct work, and whether breaks, meetings, training and absence stay in the denominator. The HDI utilization discussion shows why that definition matters; its historical ranges are not universal targets.
Productivity = qualifying resolved issues ÷ included staffed hours is one possible team measure. Define credit for collaborative work, exclude duplicate closures and show case mix. Do not target 100% utilization or compare individuals handling materially different work.
A worked example: a reply promise with two guardrails
This example is illustrative. A team reviews one week’s first-reply SLA target instances using a documented creation-date cohort and Friday cutoff. There are 80 achieved, 20 breached and 15 active unbreached targets. Achievement is (80 ÷ 100) × 100 = 80%; the 15 active unbreached targets remain a separate pending-outcome count. Any breached active targets belong in the breached count.
The mean completed human first-reply interval is 20 minutes across 60 requests, but 12 other requests remain unanswered. The mean and SLA result do not share a denominator or necessarily the same reply qualification. The team therefore checks outstanding ages and samples replies for usefulness before interpreting “fast service.”
A proposed routing change is tested against three questions: does it improve the relevant reply promise, does it preserve useful answers, and does it avoid leaving harder cases behind? No improvement is assumed until comparable post-change evidence is available.
Specify and own each KPI
A metric describes something. A KPI is a selected measure of progress toward an important objective, so it needs an accountable owner, a defined population, a target or review threshold and an action policy. Keep the executive scorecard small and leave other measures available for diagnosis.
| Field | What to record |
|---|---|
| Objective and owner | Customer or operational need; person accountable for review and action |
| Unit and source | Issue, ticket, contact, customer, respondent or agent-hour; source system and fields |
| Formula and population | Numerator, denominator, eligibility, exclusions, duplicates and missing-evidence policy |
| Time rules | Reporting period, time zone, clock, cohort and required observation window |
| Target and rationale | Service commitment or locally justified target; baseline and definition version |
| Guardrails | Quality, customer outcomes, coverage, backlog or capacity checks |
| Review and action | Cadence, minimum evidence for a decision, investigation owner and escalation route |
Where samples are too small or source coverage is incomplete, mark the result for review instead of showing a reassuring green indicator. Store counts with percentages, and separate a missed service commitment from a data-quality problem.
Set targets without making KPIs compete
Start with the service you actually offer, relevant commitments and a representative baseline. Segment channels, issue complexity and operating hours. Investigate external benchmarks only when definitions and audiences match; an unexplained industry percentage is not a defensible target.
Choose a small active portfolio. For example, a response commitment might be paired with unanswered age and reviewed response quality; a first-contact outcome with repeat-contact checks; a work-efficiency objective with resolution quality. Relationship feedback and retention provide context for leadership without becoming an agent’s personal quota.
Assign a specific change when evidence identifies a gap: routing, knowledge, staffing coverage, authorization or product behavior. Agree on the outcome check and review date. For each missed KPI, distinguish a data error, a temporary event, a recurring process issue and a capacity constraint.
Review queue exceptions at the cadence your service promise needs, process trends weekly and relationship measures when their sampling supports interpretation. Keep a change log for definitions, calendars, automation and survey policies. A balanced KPI portfolio should make service promises clearer and improvement decisions testable.
