Call center metrics are the numbers a contact center uses to see whether callers are getting through, how long the work takes, what it costs, and whether the contact ended. The short list to run on is the share of callers who wait at all, service level, abandonment rate, average handle time, occupancy, cost per contact and repeat contact rate.
Every one of those can move for reasons that have nothing to do with how your team performed. Change the denominator, change the reporting interval, change your call mix, and the same half hour reads as a good one or a bad one. What follows is the formula for each metric, the specific arithmetic that makes it misleading, a worked example where three healthy looking numbers describe a queue that is losing callers, and a worksheet for picking five and leaving them alone.
What are call center metrics?
Call center metrics are operational measurements taken from your phone platform's call records, usually averaged over short intervals. The standard interval is 30 minutes, which is how ACD reports have summarized call-by-call data for decades (Gans, Koole and Mandelbaum, Telephone Call Centers: Tutorial, Review, and Research Prospects, Manufacturing and Service Operations Management 5(2), 2003).
A metric is any number you can read. A KPI is the small set you have agreed to be judged on. That distinction matters because the first list should be long and the second should be short. Centers that promote every available metric to KPI status end up with a dashboard nobody acts on, because in any given week something is always red.
One caveat before you trust any of them. The count your platform reports as calls offered is not the count of people who tried to reach you. Calls that got a busy signal never arrive at the platform at all, and in some industries a large share of arriving calls are self-served in the phone menu and never reach the queue (Gans, Koole and Mandelbaum, 2003). Your busiest intervals are the ones where that gap is widest.
The call center metrics and KPIs worth tracking
Each of these answers a different question. The third column is the one to argue about, because a metric that changes no decision does not belong on a dashboard.
| Metric | How to calculate it | The decision it changes |
|---|---|---|
| Share of callers who waited | 1 minus (calls answered immediately divided by calls offered) | Whether an interval is genuinely short-staffed |
| Service level | Calls answered within your target time divided by calls offered | How many people to schedule in the peak intervals |
| Average speed of answer | Total wait of answered calls divided by calls answered | How long the callers who got through actually held |
| Abandonment rate | 1 minus (calls answered divided by calls offered) | Whether the queue is losing demand you paid to generate |
| Average handle time | Total talk, hold and wrap time divided by calls answered | Capacity planning, and nothing else |
| Occupancy | Time on calls divided by (time on calls plus idle available time) | Whether the schedule is sustainable in that interval |
| Cost per contact | Fully loaded cost divided by contacts resolved | Sourcing and automation, one call reason at a time |
| Repeat contact rate | Contacts from the same person about the same thing inside a fixed window, divided by contacts | Where a procedure or a knowledge gap needs fixing |
| Transfer rate | Calls transferred divided by calls answered | Whether routing or the knowledge base is wrong |
Handle time, speed of answer, abandonment and occupancy are defined here the way the research literature defines them, which is also how most platforms calculate them. Handle time is talk plus hold plus wrap, where wrap is the work a rep does after the caller has hung up. Occupancy deliberately excludes time a rep was logged in but unavailable, which is why it is not the same as a utilization figure taken across a whole shift (Gans, Koole and Mandelbaum, 2003).
What makes each of these numbers misleading
None of the following are edge cases. They are the normal behavior of the arithmetic, and they are the reason two people can read the same report and disagree.
| Metric | The failure mode | What to do about it |
|---|---|---|
| Average speed of answer | It counts only calls that were answered, so the waits of everyone who gave up are excluded. The worse your abandonment, the better ASA looks. | Never publish it without abandonment rate beside it. |
| Service level | The denominator is a choice. Measured against answered calls it flatters you, measured against calls offered it does not, and few teams know which one their platform uses. | Check the setting, write it down, and measure against calls offered. |
| Abandonment rate | It is partly a measure of your callers rather than your center. Patience varies by call reason, by whether an alternative exists, and by what you told them the wait would be. | Segment by call reason and track average time before abandoning alongside it. |
| Average handle time | A fall can mean reps got faster or that the call mix got easier. A rise can mean either in reverse. The average alone cannot tell you which. | Read it per call reason, never center-wide. |
| Occupancy | It is an interval statistic. A center can report occupancy in the nineties in every half hour while each person is on the phone half the day. | Pair it with the share of paid time reps were available to take calls. |
| Cost per contact | Divide by calls and automation looks cheap, because it takes the short ones. Divide by resolved contacts and the picture changes. | Use contacts resolved, per call reason. |
| Repeat contact rate | The window decides the answer. A 24 hour window and a 7 day window produce different numbers from identical data. | Fix the window once and never change it mid-quarter. |
The abandonment entry needs a note. Waiting statistics are objective, but abandonment is not purely yours: it carries the caller's own judgment about whether your service was worth the wait, which is why Koole and Mandelbaum describe abandonment and retrial measures as subjective in a way that waiting times are not (Queueing Models of Call Centers: An Introduction, Annals of Operations Research 113, 2002). The same paper reports a nearly linear relationship between the fraction of callers who abandon and the average wait, which is the useful half of the finding. If your waits move and abandonment does not, check your data before you celebrate.
The accessibility metric almost nobody tracks
The share of callers who have to wait at all is the root number underneath both service level and average speed of answer, since both are derived from it. It is also, as Gans, Koole and Mandelbaum observed, "almost never tracked by call center management" (2003).
Promote it to the top of your dashboard for a practical reason. Service level and ASA both compress a distribution into a single number, so they can sit inside target while a growing minority of callers has a bad experience. The share who waited compresses nothing. It answers one question: in this half hour, did people get straight through, or did they queue?
For a sense of what good looks like from real data rather than a vendor benchmark, consider the 12 call centers run by a large United States health insurer in the same study: about 40 percent of customers were delayed at all, with a 31 second average speed of answer, a 318 second average handle time, 2.8 percent abandonment, and 91 percent agent utilization (Gans, Koole and Mandelbaum, 2003). That is a center running hot on staffing while still answering most callers immediately. Those two facts are not in tension, and a dashboard without the share who waited would hide the good news.
Reading service level, ASA and abandonment together
Here is one half hour, with illustrative figures chosen to show the arithmetic. These are not benchmarks and not Telvana results.
600 calls are offered. 400 are answered immediately. 140 are answered after waiting an average of 60 seconds, of whom 30 got through inside 20 seconds. 60 callers hang up, after an average of 95 seconds on hold. So 430 calls were answered within 20 seconds, counting the 400 that never waited.
| Reading | Calculation | Result |
|---|---|---|
| Average speed of answer, as your platform reports it | (140 x 60) divided by 540 answered | 15.6 seconds |
| Average wait including the callers who gave up | ((140 x 60) + (60 x 95)) divided by 600 offered | 23.5 seconds |
| Service level against answered calls | 430 answered within 20 seconds divided by 540 | 79.6 percent |
| Service level against calls offered | 430 divided by 600 | 71.7 percent |
| Abandonment rate | 1 minus (540 divided by 600) | 10 percent |
| Share of callers who waited at all | (140 + 60) divided by 600 | 33.3 percent |
A 15 second average speed of answer is the figure that goes in the weekly report, and it is excellent. The same half hour lost one caller in ten and made a third of them queue. The two service level readings sit nearly 8 points apart on identical data, purely because of which denominator the platform was configured with.
None of these numbers is wrong. They answer different questions, and only the set of them describes the interval. If you publish one, publish the set.
Average handle time, occupancy and the limits of an average
Average handle time is a capacity input. It belongs in a staffing calculation and not in a performance conversation, because a rep who spends longer solving something properly produces a worse handle time and a better outcome. To find out whether your center is getting faster, read handle time per call reason and watch the mix separately.
Turning handle time into a staffing answer takes the queueing math rather than a ratio, and our Erlang C calculator for call center staffing does that part: enter your busiest hour's volume, your handle time and a service level target, and it returns how many people you need on the phones and on the schedule. Erlang C is an approximation, and the page is honest about where it breaks down, chiefly that it assumes nobody abandons, which is exactly the behavior the previous section was about.
Occupancy is where measurement most often turns into a management mistake. Two dated reference points from the literature are useful here. A quality-driven operation in Koole and Mandelbaum's data ran at roughly 65 percent utilization with a one second average speed of answer, while efficiency-driven centers in the same paper ran very close to 100 percent, with speed of answer measured in minutes (2002). Those are two deliberate designs rather than a good center and a bad one. Decide which you are running and staff to it, instead of discovering it from a report.
What occupancy cannot tell you is how any individual's day went. In the worked example Gans, Koole and Mandelbaum use, a center reports 95 percent occupancy in every single half hour while each rep is scheduled on the phone for only half the day (2003). Both facts hold at once, so a supervisor who manages people to an interval statistic is working from the wrong number.
Cost per contact and repeat contacts: the denominator problem
Cost per contact decides sourcing and automation, and it is the easiest metric here to get wrong, because the denominator is usually calls.
Divide fully loaded cost by calls handled and anything that takes the short calls looks cheap. Divide by contacts resolved, per call reason, and you find out what the work costs. The gap between those two figures is the work that came back.
On the cost side, labor is most of it. Salaries accounted for 60 to 70 percent of a call center's total operating costs in Koole and Mandelbaum's review (2002), which is why staffing decisions dominate the budget and why nearly every cost lever runs through the schedule.
Repeat contact rate is the metric I would track in place of first call resolution, for one reason: you can measure it. First call resolution usually depends on a rep marking their own call resolved, or on a survey that a small and unrepresentative share of callers answers. Repeat contacts sit in your call records without anyone being asked. Pick a window, count contacts from the same person about the same thing inside it, and hold the window fixed. If you also run first call resolution, keep both and expect them to disagree.
Comparing what a call type costs in house against a provider or an AI agent is a separate exercise with its own arithmetic, and call center outsourcing against AI agents works through how each is priced, with a scorecard for deciding one call type at a time.
How to pick five call center KPIs and leave them alone
Five is enough, and the small number is doing real work: a metric set that changes every month cannot show a trend. Fill this in for your own center, then publish the same five weekly for a quarter before touching the list.
| Slot | The question it answers | Metric | Segment it by | Read it with |
|---|---|---|---|---|
| Access | Did callers get through? | Share of callers who waited at all | Half hour, day of week | Abandonment rate |
| Delay | How long did the ones who waited, wait? | Service level against calls offered | Half hour | Average speed of answer |
| Loss | Who gave up, and after how long? | Abandonment rate plus average time before abandoning | Half hour, call reason | Share who waited |
| Cost | What does this work cost? | Cost per resolved contact | Call reason | Volume by call reason |
| Completion | Did the contact end? | Repeat contact rate in a fixed window | Call reason | Transfer rate |
Two rules separate a dashboard from a report nobody opens. Every metric gets segmented by call reason, because an unsegmented average cannot survive a change in your call mix. And every metric is published next to the one that catches its failure mode, which is what the fourth and fifth columns are for.
If a metric in your current set does not fit a slot above, ask what decision it changes. Some have good answers, like schedule adherence while you are fixing a scheduling problem. Others have been on the report since before anyone currently in the room arrived.
What measurement does to the people being measured
Continuous electronic performance measurement is itself a working condition, and it is part of why this job exhausts people faster than most. In a study of 150 call center employees at a commercial bank, emotional exhaustion was by far the strongest predictor of job satisfaction, with a standardized beta of minus 0.644 in a model explaining 54 percent of the variance, while workload on its own lost its predictive power once burnout entered the model (Keser and Yılmaz, Workload, Burnout, and Job Satisfaction Among Call Center Employees, Journal of Social Policy Conferences 66 to 67, 2014).
The operational reading is about where the exhaustion sits. It falls between your workload and your attrition, so a metric set built to pressure individuals rather than to find broken procedures costs you in turnover, and turnover is the most expensive thing in a contact center. A metric that points at a call reason gives a supervisor something to fix on Monday. A metric that points at a person mostly gives them something to defend.
Which call center metrics move when AI agents take part of the volume
The measurement problem arrives before the results do. When an AI agent starts taking a slice of your volume, the denominator of nearly every metric above changes at the same time, so a before-and-after comparison on center-wide averages tells you almost nothing.
Three things make that comparison readable. Freeze a pre-period and keep its per-call-reason breakdown. Name the one number you expect to move, and where you will read it, before you route anything. And keep handle time segmented, because the calls leaving your queue are the short ones, which pushes the human average up without anyone slowing down.
Two new metrics also appear, and they are easier to define now than to reconstruct later: the share of contacts the agent completed without a transfer, and the transfer rate with its reasons. The second is the more useful in the first month, because every transfer marks a gap in a procedure or a knowledge base that someone can go and fix.
On the Telvana platform, every call is recorded, transcribed, summarized and tagged by outcome, which is what makes per-call-reason reporting practical instead of a spreadsheet project: you count outcomes rather than calls. Teams running several programs or client accounts give each one its own workspace, with its own agents, numbers and reporting. OakTech Systems runs calling programs for its clients that way, with a separate workspace per client program and more than 20 AI agents in production. Across five production customers, Telvana has handled more than 195,000 calls in the last 12 months.
For which metrics move and which do not, what AI agents handle in a call center covers the inbound side and the three different products sold under that name. Outbound programs run on a different metric set built around reach rather than response, and running an outbound call center has that table along with the federal pacing rules. Where the queue itself is the problem, hold queue relief and after hours and overflow are the two patterns most centers start with.
Frequently asked questions
What are the most important call center metrics?
The share of callers who waited at all, service level against calls offered, abandonment rate with average time before abandoning, cost per resolved contact, and repeat contact rate in a fixed window. That set covers access, delay, loss, cost and completion, and each one catches a failure mode in another.
What is ASA in call center metrics?
ASA is average speed of answer: the total time answered calls spent on hold, divided by the number of calls answered. Its important limitation is that it excludes the waits of callers who hung up before being answered, so a center with rising abandonment can show a falling ASA. Always read it with abandonment rate.
What is the difference between a call center metric and a KPI?
A metric is any number your platform can report. A KPI is the short set you have agreed to be judged and to make decisions on. Most centers have dozens of metrics available and should promote about five to KPI status, because a list that changes every month cannot show a trend.
What is a good service level for a call center?
There is no derived optimum. In the research literature the service level target is a management choice that is then traded off against staffing cost, and target answer times are typically set at 20 or 30 seconds (Gans, Koole and Mandelbaum, 2003). The widely quoted 80 percent in 20 seconds is a convention rather than a figure anyone calculated for your center. Pick a target, record which denominator you measure it against, and hold both steady.
How many call center KPIs should we track?
Five, published weekly and left alone for a quarter. Track as many secondary metrics as you like for diagnosis, but keep them off the scorecard. The test for promoting one is whether a change in it would change a decision you make.
How often should we review call center metrics?
Watch the interval level numbers daily for staffing, review the five KPIs weekly for trend, and review the KPI list itself quarterly at most. Reporting intervals of 30 minutes are the operational standard and the right resolution for staffing decisions. They are the wrong resolution for judging whether anything is improving.
Start with the numbers you already have
Before changing anything, pull last month's call records, sort by reason for contact, and work out your five metrics for the top three reasons separately. Most centers find at least one reason where the center-wide average has been hiding a problem, and that reason is usually where the first useful change is.
To see which of those call reasons an AI agent could take, book a demo and bring your top call types and the numbers you run on. We will show you which ones Telvana can handle, let you hear it take one, and walk through what would move in your reporting and what would not.