Call Center Metrics and KPIs: 32 Numbers That Actually Predict Performance
Most call center dashboards measure 40+ numbers and act on three. That gap is where money leaks.
The teams that run lean contact centers do not track more metrics. They track fewer, with sharper definitions, and connect them to the cost of every decision. A 2-minute reduction in Average Handle Time means something if it does not push repeat calls up by 8%. A 92% CSAT means something if the 8% who said "very dissatisfied" account for 60% of churn. Numbers without that context are scoreboard theater.
This guide covers the 32 call center metrics worth tracking in 2026, organized by what they actually decide: cost, customer outcomes, agent load, and AI performance. Every section has a formula, a real benchmark, and what changes when you put an AI voice agent on the front of your queue.
A call center metric is a measurement that maps one observable event in the contact center to a decision someone can make about staffing, training, technology, or process. If it cannot change a decision, it is reporting, not a metric.
The split between metrics and KPIs is simple in practice. Metrics are the raw measurements. KPIs are the small set tied directly to a business outcome the leadership team has agreed to move this quarter. A center can have 50 metrics and four KPIs. Most have 50 and 50.
Common mistake: treating every dashboard tile as equally important. If everything is a priority during the Monday review, nothing gets fixed. Pick four to six KPIs per role. Frontline supervisors monitor AHT and adherence. Operations directors monitor cost per call and FCR. CX leaders monitor CSAT, CES, and repeat call rate. Keep the rest as diagnostic metrics — looked at only when a KPI moves.
These measure how the call felt from the caller side. They are lagging indicators of everything else.
CSAT is the share of customers who rated a specific interaction positively, typically the top two of a five-point scale.
Formula: (Positive responses ÷ Total responses) × 100
Industry benchmarks land around 85% for financial services, 90% for retail, 82% for healthcare, 78% for telecom. The number alone is not very useful. The drivers are. Slice CSAT by call type, by agent tenure, by time of day, by issue category. The bottom decile of a CSAT distribution explains more than the average.
Pro tip: survey within 60 seconds of call end and keep it to one question. Response rates collapse past 90 seconds and past two questions. Aim for at least 25% response rate before you trust the score.
NPS measures whether customers would recommend you. It is a brand-level signal, not a call-level signal, so use it for trend lines across quarters, not real-time coaching.
Formula: % Promoters (9–10) − % Detractors (0–6)
The number itself is less informative than the Detractor comments. A drop of 5 points in NPS usually traces to a specific operational change — a process update, a policy shift, a queue change. Read every Detractor verbatim from the last 30 days before you spend a meeting debating tactics.
Pine Park Health raised scheduling NPS by 38% after putting an AI voice agent on patient calls. Their COO Mike Tadlock attributed the lift to ending phone tag for routine scheduling, which freed the human team to focus on the calls where they added the most value.
CES asks how hard a customer had to work to get their problem solved. Gartner's research showed CES predicts loyalty better than CSAT for service interactions specifically, because customers do not remember pleasant calls. They remember the painful ones.
Formula: Sum of effort ratings ÷ Number of responses
Anything above a 4 on a 7-point scale is a problem. The fix is rarely "train agents harder." It is usually: too many transfers, IVR routing that misses intent, knowledge bases agents cannot find what they need in, or callbacks that did not happen on time.
FCR is the share of issues solved on the first contact, without a callback or transfer. It is the single most predictive call center metric for both cost and CSAT.
Formula: Resolved on first contact ÷ Total contacts × 100
The reliable measurement is a 7-day or 14-day window — if the customer calls back about the same issue inside that window, the original call did not resolve it. Self-reported FCR from agents at end-of-call is inflated by about 15 to 20 percentage points compared with the windowed measurement. Use the windowed version or do not bother tracking it.
Median FCR sits around 70 to 75% across industries. Anything below 65% means your training, knowledge base, or routing is broken. Anything above 85% is suspicious — either the definition is loose, or simple calls are being counted disproportionately.
These are the speed metrics. They get over-monitored, which is fine if they are not over-corrected.
ASA is the average wait time before a live agent picks up, queue entry to answer. It excludes IVR navigation.
Formula: Total wait time across answered calls ÷ Total answered calls
The classic service-level target is 80% of calls answered within 20 seconds. ASA below 30 seconds is healthy for most B2C operations. Above 60 seconds and abandonment rates spike non-linearly.
ASA is a staffing decision, not a training decision. If it is high, you either need more agents during the peak hour or you need to deflect lower-value calls before they hit the queue. An AI answering service running on the inbound line picks up in under a second, which is why centers that deploy voice AI for tier-one calls see ASA drop without adding headcount.
Service level is the share of calls answered within a target time, usually expressed as "X% in Y seconds."
Formula: Calls answered within threshold ÷ Total calls offered × 100
The 80/20 standard is industry default but not industry optimum. For tier-one support, 80/30 is fine. For sales inbound, 90/15 protects pipeline. For emergency or roadside, 95/10. The right target depends on what abandoning the call costs the business.
Abandonment is the share of callers who hang up before reaching an agent. Most analytics platforms exclude calls dropped in the first five seconds — those are dial errors, not abandonments.
Formula: (Calls offered − Calls answered) ÷ Calls offered × 100
Below 5% is healthy. Above 10% means your queue is bleeding revenue or customer trust depending on the call type. The fix that compounds fastest is callback offers in the IVR — Sunshine Loans cut application abandonment to 5% after replacing their inbound queue with AI handling 700,000+ monthly contacts, which both removed the wait and made the callback unnecessary.
AHT is the most over-watched and under-understood metric in the industry.
Formula: (Talk time + Hold time + Wrap-up time) ÷ Total calls
The benchmark range is 5 to 9 minutes depending on industry. The trap is that pushing AHT down by 30 seconds often pushes repeat call rate up by 5 to 10 percentage points, which costs more than the time saved. Track AHT alongside FCR and repeat rate. Never alone.
When AHT is the wrong metric: for complex tier-two issues, longer calls correlate with higher resolution and higher CSAT. Holding tier-two agents to a tier-one AHT target trains them to rush, which is the most expensive coaching error a center can make.
Hold time is the slice of AHT where the caller is waiting mid-call. It is the leading indicator of knowledge base problems.
If your agents are putting customers on hold for 90+ seconds to look something up, the fault is not the agent. It is the system they are searching in. A streaming knowledge base that surfaces the answer in under a second eliminates most mid-call holds.
ACW is the time agents spend on documentation, ticket updates, and CRM entries after the call ends.
Formula: Total wrap-up time ÷ Total calls handled
Healthy ACW lands between 30 and 90 seconds. Past two minutes, you are either over-documenting or your CRM has too many required fields. The cheapest fix is automated post call analysis that transcribes, summarizes, and writes the disposition before the agent's next call connects. Centers that automated ACW report 15 to 25% capacity gains without changing headcount.
Active waiting calls is the real-time count of callers in queue. It is not a KPI. It is a supervisor's dashboard tile that triggers in-the-moment decisions: pull an agent off training, open a callback offer, escalate to the support overflow team.
Track the worst case, not just the average. A 95th-percentile hold time tells you whether your queue is failing the unlucky few even when your averages look fine. If the 95th percentile is 6 minutes and the average is 45 seconds, you have a routing problem, not a staffing problem.
Agent metrics fail when they reward gaming. The list below is filtered for metrics that resist gaming because they are tied to actual customer outcomes.
Utilization is the share of paid time agents spend on call-handling activity, including wrap-up.
Formula: (Handling time + ACW) ÷ Paid time × 100
The sustainable target is 75 to 85%. Above 90% and burnout shows up in CSAT within 60 days. The number is also misleading without a denominator audit: if your "paid time" excludes training and breaks correctly, 85% is healthy. If it includes them, 85% means agents are on the phone with no recovery time.
Adherence measures whether agents were available during their scheduled hours. Hard target: 90 to 95%. Anything below 85% means your WFM forecasts are wrong or your team has a discipline problem, and those are different fixes.
A productivity count, useful when broken out by call type. The Reddit threads from working agents make the point: a roadside emergency call cannot be benchmarked against a retirement plan call. If your team handles mixed call types, segment this metric by type or do not track it at all.
QA scores measure how closely agents follow process and quality standards during calls. The dirty secret of QA is that most programs sample 1 to 3% of calls, which means 97% of agent behavior is invisible to coaching. The teams that catch real issues — compliance skips, recurring de-escalation failures, knowledge gaps — review 100% of calls with AI-driven scoring and have humans deep-dive on the flagged ones.
Transfer rate is the share of calls handed off to another agent or department.
Formula: Transferred calls ÷ Calls handled × 100
The target is under 10% for established centers. Above 20% means routing logic is wrong, agent skills are too narrow, or the IVR is collecting the wrong intent upstream. An AI IVR that captures real intent with natural language eliminates a meaningful share of preventable transfers — that is the difference between routing on "press 1 for billing" versus routing on what the caller actually described.
AES surveys the agent on how hard their job felt — too few centers track this. Without it, burnout is invisible until attrition spikes. The Reddit operator threads pasted here make the point: agents game metrics like calls per hour because the metrics are not measuring what the work actually costs them. AES is the only number that surfaces this.
Turnover rate is the share of agents leaving over a defined period. The contact center industry runs at 30 to 45% annual turnover, which is the single largest hidden cost in the operating model. Each agent departure costs roughly $4,000 to $10,000 in recruiting, onboarding, and lost productivity. Cut turnover by 10 percentage points and you have funded most contact center tech projects.
These are the metrics finance cares about and operations directors translate into staffing decisions.
CPC is the all-in cost of handling one customer contact, including agent labor, technology, supervision, and overhead.
Formula: Total operating cost ÷ Total calls handled
Benchmarks vary wildly. A human-handled tier-one support call costs between $3 and $8 in most industries. An AI-handled call on a modern voice platform runs between $0.10 and $0.30. SWTCH cut support costs by more than 50% after deploying AI customer support on Retell AI, with their CEO noting that the system "answers calls in seconds, handles urgent EV support at scale, cuts support costs by over 50%, and significantly improves our SaaS margins."
Repeat rate is the share of customers who call back about the same issue inside a defined window (typically 7 or 14 days).
Formula: Calls about a repeat issue ÷ Total calls × 100
Below 10% is healthy. Above 20% means FCR is being measured wrong or your knowledge base is incomplete. Repeat calls are the cleanest signal of root-cause failures — every recurring topic in the repeat-call pile is either a product issue, a policy issue, or a documentation issue. Fix those upstream and the metric collapses.
CPC stops being useful when calls, chats, and emails sit on the same dashboard, because their costs differ. Measure cost per contact by channel, then look at total cost to resolve — sometimes a $1.50 chat plus a $4 follow-up call costs more than handling the original problem on a $5 phone call.
Arrival rate is the volume of inbound contacts per unit time, used for workforce management forecasting. The pattern matters more than the average — daily peaks, day-of-week effects, and seasonal swings drive staffing far more than the rolling 30-day mean.
Containment is the share of contacts resolved entirely within one channel without escalation. Self-service and AI voice both push containment up. Genesys-style measurement uses this as the primary indicator of whether digital channels are actually working — a chat channel with 30% containment is a feeder for the phone queue, not a deflection tool.
Age of query is the average elapsed time for unresolved issues, useful in centers with case-based work alongside call handling. Anything past 72 hours typically requires a callback or proactive outreach, which is its own staffing line.
The aggregate version of agent utilization, used in operating model decisions. Sustained sub-70% utilization means the center is overstaffed for current volume; sustained 85%+ utilization means quality and retention are at risk.
The metrics that mattered in a fully human center are necessary but not sufficient for a center where AI handles part of the volume. The four below are specific to evaluating AI handling.
Containment is the share of AI-handled calls resolved without human transfer. Above 60% is good for tier-one inbound. Above 80% means either the use case is well-fit or you are scoping the deflection too narrowly. Medical Data Systems runs at a 70% containment rate — their AI handles 100% of inbound calls with only 30% needing human transfer, generating roughly $280,000 monthly in collections.
Accuracy is the rate at which the AI correctly identifies what the caller wants. Below 90% and your CSAT will suffer regardless of how good the rest of the system is. The leading indicator of accuracy problems is high false-positive transfer rates — the AI routing to humans on cases it should have handled.
Escalation is the share of AI conversations handed off to a human. Track the reason codes, not just the rate. Some escalations are by design (complex billing disputes, accounts the AI should not touch). Others are failure modes (the AI did not understand, the customer demanded a human, a tool call failed). The two require different fixes.
AI handle time is the analog to human AHT for AI-handled calls. With sub-800ms response latency on a modern voice platform, AI handle times typically run 30 to 50% shorter than human handle times for matched call types — because the AI does not need hold time to look things up, and wrap-up is automatic.
Most call center metrics were designed for a fully human operating model. They still work, but four of them shift in ways that matter.
AHT distribution flattens: Human centers see wide variance in handle time because agent skill differs. AI-handled calls cluster tightly around a mean — same handling on every call, no fatigue, no inconsistency. Variance shifts from agent-to-agent to call-type-to-call-type.
Wrap-up time drops to zero on the AI portion: Documentation happens automatically through structured outputs from the call. Human agents still wrap up the calls they take, but the share of total wrap-up time falls in proportion to AI containment.
CSAT splits by handler: Track AI-handled and human-handled CSAT separately. They will look different — sometimes AI scores higher (no hold time, no transfers), sometimes lower (limited problem types). The aggregate score hides both.
Cost per call collapses on the contained portion: The math changes from "how do I reduce CPC by 10%" to "how do I grow AI containment from 40% to 70%." Same goal, completely different operational lever.
GiftHealth saw 4x operational efficiency gains after deploying AI voice agents for healthcare pharmacy operations. The number worth noting is not the multiplier but the mechanism: the same team handled four times the volume because the AI absorbed the predictable calls while humans took the complex ones.
If you are building or rebuilding a call center reporting layer in 2026, this is the small set of KPIs worth wall-mounting. Everything else is a diagnostic.
| Role | Primary KPIs | Diagnostic Metrics |
|---|---|---|
| Frontline supervisor | Adherence, AHT (by type), QA score | Hold time, ACW, calls per hour |
| Operations director | FCR, Service level, CPC | Abandonment, repeat rate, utilization |
| CX leader | CSAT, CES, NPS | Repeat rate, transfer rate, AES |
| Finance | Cost per call, CPC by channel, Containment | Headcount utilization, turnover cost |
| AI/Automation lead | Bot containment, Intent accuracy, AI CSAT | Escalation rate, AI AHT |
Eighteen metrics across five roles. If your dashboard has more than 25 visible tiles, you are tracking things nobody is acting on.
Three moves, in order, fix most contact center reporting problems.
First, define every metric in writing with the exact formula and the data source. Most centers have three different definitions of FCR floating between teams. Pick one, document it, retire the others.
Second, segment the four critical metrics — CSAT, AHT, FCR, repeat rate — by call type. The aggregate numbers are misleading because tier-one and tier-two have completely different operating norms. You cannot manage a mixed-handle queue with mixed-handle metrics.
Third, audit the data feeding the dashboard. ACD systems, CRM logs, and QA platforms often disagree on what counts as a "call" or a "resolution." Until those reconcile, the dashboard is showing you fiction.
Once those three are clean, the upgrade most centers should consider is deflecting tier-one volume to an AI voice agent that handles the predictable calls end-to-end. The economics work because of the cost gap between AI handling and human handling, but the operational value is bigger: AI handling is consistent, it does not burn out, and it generates clean structured data for every interaction that feeds back into the QA and analytics layer automatically.
Teams running call center automation on Retell AI typically see containment land between 50% and 80% on tier-one inbound within the first 60 days, with deployment timelines measured in days rather than months. The platform handles 30+ million calls per month across more than 3,000 businesses, which is the production benchmark worth comparing against when evaluating any voice AI infrastructure.
First Call Resolution, when measured with a 7- or 14-day repeat-call window. FCR correlates with CSAT, CPC, and repeat call rate simultaneously, so improving it moves three KPIs at once. Self-reported FCR is much less useful — track it from the data, not from agent end-of-call surveys.
The healthy range is 80 to 90% depending on industry. Retail and consumer services run higher (85–92%), healthcare and financial services run lower (78–85%), telecom and utilities lower still (75–82%). The number matters less than the trend and the bottom-decile drivers.
Talk time is just the agent-caller conversation. AHT includes talk time plus hold time during the call plus wrap-up time after the call ends. A center with 4-minute talk time, 1-minute average hold, and 2-minute wrap-up has a 7-minute AHT.
No. Targets above 95% are unsustainable and indicate over-control. 90 to 93% is healthy and accounts for normal human variability — breaks, restroom needs, brief tech issues. Below 85% is the actionable threshold.
They split the metric stack: AI-handled and human-handled volume need to be tracked separately. AHT on AI calls runs shorter, wrap-up drops to near zero, CSAT depends on the call type, and cost per call collapses on the contained portion. Aggregate metrics hide both the gains and the failure modes.
Metrics are measurements. KPIs are the small subset of metrics tied directly to a business outcome you have agreed to move. Most centers have 40+ metrics and should have 4 to 8 KPIs per role. The rest are diagnostics — useful when a KPI moves, ignored when it does not.
They measure different things. CSAT measures specific interactions. NPS measures brand loyalty across all touchpoints. If your call center is the primary customer touchpoint, CSAT is enough at the call level and NPS is your quarterly trend. If your customers interact with your brand through many channels, NPS catches issues CSAT will miss.
See how much your business could save by switching to AI-powered voice agents.
Total Human Agent Cost
AI Agent Cost
Estimated Savings
A Demo Phone Number From Retell Clinic Office
Start building smarter conversations today.
