Call Center QA Metrics: Formulas and Benchmarks

Last updated: September 12, 2026

What call center QA metrics measure

Call center QA metrics measure the quality program as well as the calls it reviews. The working set is QA score, critical fail rate, coverage rate, calibration variance, and coaching completion. SQM Group puts the industry average QA score at 85%, treats 90 to 99% as good, and reports that only 5% of agents reach 100.

A QA average that climbs three months running is the most reassuring number on a contact center dashboard, and one of the easiest to produce while the floor gets worse.

Two ordinary things do it without anybody deciding to cheat. Reviewers working against a deadline reach for calls that score quickly, which pulls the sample toward agents who already pass. Scorecard drift does the rest: after a quarter that reads badly, the criterion everyone kept failing gets softened or demoted out of the auto fail column, and the same calls come back higher.

So the split worth drawing is between metrics that describe calls and metrics that describe the program doing the scoring. QA score and critical fail rate sit on the call side. Coverage rate, calibration variance, and coaching completion tell you whether the call side means anything, and a dashboard carrying only the first pair leaves an owner no way to tell a real 88 from a manufactured one. A program that has not written its standard down yet should start with what call center QA is.

The five metrics a small QA program should run

Five numbers cover a quality program at 20 to 80 seats, and each one has a standard way of going soft once somebody is measured on it.

The same table downloads as a CSV with the columns in that order.

QA score

Points earned over points available on one call, averaged across the calls scored for an agent or a team, with an auto fail taking the whole call to zero. That is the formula behind the 85% average and the 90 to 99% band SQM Group reports, and behind its finding that only about 5% of agents reach a perfect score.

The number also moves for reasons that have nothing to do with agents. A reviewer with six calls left and forty minutes to finish gives the benefit of the doubt on the judgment criteria, and that is a quiet upward push on the month's average.

Two programs both reporting 88 can be running standards that share almost nothing, so when the scorecard was inherited and nobody remembers who set the weights, a 100 point QA scorecard template is a cheaper place to start than defending the old one.

Critical fail rate

The share of scored calls carrying at least one auto fail. A missed recording disclosure on a call that required it, or a callback number in the message that does not match what the caller said, zeroes the call however well the rest of it went.

What breaks this metric is a long auto fail list. Push it to fourteen items and a third of your calls zero out, which leaves you unable to rank a sloppy call against one that put an account at risk. Four to six is our own working range rather than a published standard, and it holds because every item on a short list is something you would escalate the same day.

An auto fail is a same day event, which is why this one belongs on a weekly view rather than in a month end average.

Coverage rate

Scored calls divided by calls handled over the same window. Coverage decides how much weight the other four can carry, and most programs have never calculated it.

Say an agent takes 1,400 calls a month and a supervisor scores three of them. Coverage is roughly 0.2%, the average rests on three calls somebody chose, and that is normal rather than negligent. Manual QA reaching 1 to 3% of calls is the standing figure here, cited by Verint, Gistly, and ViciStack, and SQM Group reports that about 60% of centers evaluate five or more calls per agent per month.

Call Centre Helper puts a statistically defensible sample at about 65 evaluations per agent per month, at 90% confidence, for an agent handling 1,400 calls in that month. Change the volume or the confidence level and that sample size moves with it.

Run the arithmetic on 65 and it falls apart. At five minutes a call, one agent needs about five and a half hours of review a month, and a 20 seat floor needs near 110 hours before anybody has talked to an agent about what the scores showed. Nobody staffs that.

So treat 65 as a scale rather than a target. Measure your own coverage, publish it beside the average score, and refuse to act on a score whose coverage dropped that week. Automated scoring changes that arithmetic and brings a limit with it: rule based criteria hold up, and anything resting on tone gets shakier, which Aircall and StackAI both note. Getting coverage off 0.2% is worth doing, and it makes a vague criterion no easier to score.

The gaming here lives in the denominator. A program measuring against calls eligible for review, after the short calls and the transfers come out, can report 40% while scoring 6% of what the floor handled. Scoring every call instead of a sample changes the supervisor's week more than it changes the math.

Calibration variance

The spread between reviewers scoring the same call. Two supervisors score the same ten calls without seeing each other's sheets, and the variance is the gap between them.

No published benchmark for this survived our checking, so the baseline is yours to build: score the ten, keep the numbers, and compare next quarter's spread against them. When the gap is wide enough to move a call from a pass to a fail, the wording of the criterion is the problem rather than the reviewers.

A criterion like "showed empathy" produces variance no amount of calibration fixes, because two people can hear the same 40 seconds and disagree in good faith. Rewriting it as an observable behavior helps, and some disagreement stays. A program reporting zero variance is usually scoring out loud, where the second reviewer hears the first one's number before writing their own. Calibration variance in practice covers the session itself.

Coaching completion

Scored calls that flagged something and produced a documented coaching conversation, divided by scored calls that flagged something. Whatever counts as documented is fine, as long as it is the same thing every month.

High coverage with low completion is a reporting exercise. It shows up as a dashboard full of scored calls, a supervisor who can name the problem accounts, and agent scores that have not moved in two quarters. Scoring is measurement, and the conversation after it is the part that changes a call, which is why what happens after the score gets its own page.

Metrics that belong to operations rather than QA

Average handle time, first call resolution, service level, and abandon rate come off the phone platform rather than out of somebody reading a call. This page leaves them there, along with the argument about what a good handle time is.

One caution about the boundary. Handle time inside a QA scorecard punishes the agent who reads a callback number back digit by digit, which is the behavior the capture criterion exists to get.

Reporting a client account separately from the floor

An answering service or a BPO scoring against one blended average is describing a floor that nobody buys. The operator picks up in the client's business name, works from that account's script, and follows the on call schedule in force at that hour, so the standard changes with the account even though the person does not.

Say 80% of your volume scores 92 and one account at 20% scores 71. The blended number is about 88, which reads as a healthy program right up until the account at 71 gives notice.

CallSniper reports dashboards by agent and by client account, and organization branding puts that report under your logo when you send it out. Scoring per client account also makes a quality claim verifiable, since you hand the client calls taken in their name against the script they approved.

The three you watch on a Monday

Most programs track more numbers than anyone acts on, and the tell is a dashboard nobody opens between board meetings. Three of the five earn a weekly look because each one ends in a decision somebody makes that week.

Coverage rate decides whether the rest of the dashboard gets read at all. When it drops, the scores moved because the sample moved, and nothing else on the page is worth acting on until coverage comes back.

Critical fail rate decides what gets pulled today. A single auto fail on a medical account is a phone call to that client rather than a line in a report somebody reads on the 5th.

The third is QA score by agent, read as movement rather than as a level, and it picks who gets the coaching slot this week. An agent who fell nine points since the scorecard changed is a different conversation than one who has sat at 74 all year.

The other two run monthly. Calibration variance gets rerun whenever the scorecard wording changes, since that is when reviewers quietly diverge, and coaching completion answers a question that does not move week to week.

A call review scores a batch of your own recent recordings against your existing scorecard and hands back these numbers on your own audio, straight out of whatever your platform exports, so nothing waits on a telephony integration.

Frequently asked questions

What is a good QA score in a call center?

SQM Group puts the industry average at 85% and treats 90 to 99% as good, with only about 5% of agents reaching 100. Take those as orientation rather than a target, since a score is only as meaningful as the scorecard behind it. Publish calibration variance beside the average or the number is unreadable.

How do you calculate a QA score?

Divide points earned by points available on the call, then average across the calls scored for that agent or team. An auto fail overrides the arithmetic and takes the call to zero, which is the point of having auto fails at all.

What is a critical fail?

A criterion serious enough that failing it voids the score for the whole call. Two common ones: a recording disclosure that never got read on a call requiring it, and a callback number that was never read back and turns out to be wrong.

How many calls should QA review per agent per month?

Call Centre Helper puts a defensible sample at about 65 evaluations per agent per month at 90% confidence for an agent taking 1,400 calls, and both assumptions move the answer. Practice sits far below that, with SQM Group reporting that around 60% of centers evaluate five or more calls per agent per month. Measure your own coverage rate before arguing about the target.

What is the difference between QA metrics and call center KPIs?

QA metrics come from somebody applying a written standard to a recorded call. Contact center KPIs such as average handle time, service level, and first call resolution come off the phone platform and measure the operation. Both belong on a report to an owner, and mixing them into one scorecard makes the scorecard worse.

Sources

• SQM Group on the 85% average QA score, the 90 to 99% band, and 5% of agents at 100. Third party, unverified.
• SQM Group on roughly 60% of centers evaluating five or more calls per agent per month. Third party, unverified.
• Call Centre Helper on about 65 evaluations per agent per month at 90% confidence, 1,400 calls. Third party, unverified.
• Verint, Gistly, and ViciStack on manual QA reaching 1 to 3% of calls. Third party, unverified, setup only.
• Aircall and StackAI on rule based criteria scoring more reliably than tone. Third party, unverified.
• Trinsic feature sheet and call notes, 2026-08-27 and 2026-09-03. Verified internally.
• The auto fail range, the five minute review estimate, and the blended account example. Editorial judgment.

Related industriesAll industries