Call center quality assurance is the practice of reviewing recorded calls against a written standard, scoring them, and using those scores to coach agents, fix procedures and decide what to automate. Almost every program does it the same way: a reviewer listens to a few of each agent's calls and judges them.
That method has a name outside the contact center. It is sampling inspection, borrowed from manufacturing, and the research literature describes call center QA in exactly those terms: agent calls are listened to at random, live or recorded, and judged as requiring rework or not (Gans, Koole and Mandelbaum, Telephone Call Centers: Tutorial, Review, and Research Prospects, Manufacturing and Service Operations Management 5(2), 2003). Once you accept that framing, the statistics of sampling apply to your scores, and most of what a QA program believes about its agents does not survive them.
Below is a weighted scorecard with anchored scoring, the arithmetic for how many calls you have to score before a gap between two agents means anything, a calibration procedure, and what to do differently when your platform already transcribes and tags every call.
What is call center quality assurance?
Call center quality assurance is a repeatable review of recorded contacts against a standard your team has written down, scored on a scale, and fed back into coaching, process changes and routing. QA answers one question your dashboard cannot: on the calls that look fine in the report, did the caller actually get what they needed?
Gans, Koole and Mandelbaum separate the three kinds of quality a center tracks, and it is worth keeping them apart because they need different instruments. The first is accessibility: did callers get through, and how many gave up. That lives in your call center metrics, not in QA. The second is effectiveness, the contact center version of rework: was the caller's problem resolved, or did more work follow. The third is the content of the interaction itself, and that is where most scorecards spend all their weight.
The same authors note something uncomfortable about the middle one. Writing in 2003, they found no widespread, formalized schemes for managing service effectiveness. That is the gap a scorecard has to close, and it is why the weights in the next section put resolution and accuracy ahead of manner.
What should a call center QA scorecard measure?
Weight the scorecard toward whether the caller's reason for calling was handled correctly, and give manner the smallest share. The five categories below total 100 points, with resolution and accuracy carrying 60 of them. Copy it and change the weights, but keep that ratio: a scorecard where courtesy outweighs correctness will produce polite agents who send callers away unhelped.
| Category | What it checks | Weight |
|---|---|---|
| Resolution | The caller's reason for calling was handled, with nothing left for them to chase | 35 |
| Accuracy | Everything the agent told the caller was correct and current | 25 |
| Process | Verification, notes and disposition done the way the procedure says | 15 |
| Routing | The transfer or escalation was the right one, and the caller was told what happens next | 15 |
| Manner | Clear, courteous, and no unexplained silence | 10 |
A weight without an anchor is still a guess. Two reviewers will read "handled correctly" differently every time. Score each category on three points with the behavior written out, not on five or ten points where the middle is undefined.
| Category | 5 points | 3 points | 1 point |
|---|---|---|---|
| Resolution | Resolved on the call, nothing left open | Partly resolved, with a specific next step the caller agreed to | Not resolved, and no next step was set |
| Accuracy | Everything said was correct | Vague where it should have been specific, but nothing wrong | Told the caller something untrue or out of date |
| Process | Verification, notes and disposition all done | One of the three missing or thin | Account changed without verification, or no record of the call |
| Routing | Right destination, caller told what to expect | Right destination, caller left guessing | Wrong destination, or transferred to avoid the work |
| Manner | Clear and courteous throughout | One lapse in tone or an unexplained pause | Dismissive, or talked over the caller |
Convert to a percentage by scoring each category, dividing by 5, and multiplying by the weight. An agent who scores 3 on resolution and 5 everywhere else lands at 86, which has the right shape to it: one soft resolution should cost more than one flat greeting.
Auto-fails belong on their own list, not in the weights
Some behaviors should fail a call outright rather than lose it a few points, because averaging them into a weighted score hides them. Keep that list short, keep it binary, and review every call that trips it.
- An account was changed without verifying who was asking.
- The caller was told something that contradicts your own policy.
- The required disclosure your policy specifies was not given.
- The call ended while the caller was still asking a question.
- The agent was abusive or openly dismissive.
Five is about the limit. Once the auto-fail list grows past that, reviewers start negotiating with it, and a list people argue their way around measures nothing. What belongs on yours is a question for whoever owns your compliance obligations, not for the QA team alone.
How many calls do you have to score?
More than you are scoring now. A QA score is an estimate of a proportion from a sample, so it carries a confidence interval, and at the sample sizes most programs use that interval is wider than the differences the program is trying to act on.
The table below is the 95 percent confidence interval around an observed pass rate, calculated with the Wilson method the NIST and SEMATECH e-Handbook of Statistical Methods recommends for proportions. These are not benchmarks and not Telvana figures. They are what the formula returns for a given number of scored calls.
| Calls scored per agent | Calls passed | Observed rate | 95 percent interval | Width |
|---|---|---|---|---|
| 4 | 3 | 75.0 percent | 30.1 to 95.4 percent | 65 points |
| 5 | 4 | 80.0 percent | 37.6 to 96.4 percent | 59 points |
| 10 | 9 | 90.0 percent | 59.6 to 98.2 percent | 39 points |
| 20 | 18 | 90.0 percent | 69.9 to 97.2 percent | 27 points |
| 50 | 45 | 90.0 percent | 78.6 to 95.7 percent | 17 points |
| 100 | 90 | 90.0 percent | 82.6 to 94.5 percent | 12 points |
| 200 | 180 | 90.0 percent | 85.1 to 93.4 percent | 8 points |
Read the first row again. A reviewer scores four of an agent's calls, three pass, and the honest statement is that this agent's true pass rate is somewhere between 30 and 95 percent. That is not a quality score. It is a coin flip with a decimal point on it.
Programs usually express their effort as a share of volume, which makes the sample sound larger than it is. An agent taking 400 calls a month who gets four of them scored has been sampled at 1 percent, and the sample is four calls. The percentage is the reassuring number and the count is the one that governs the statistics.
When is a gap between two agents real?
At around 200 scored calls per agent. If one agent is observed at 90 percent and another at 80 percent, their confidence intervals keep overlapping until the sample reaches roughly 183 calls each, and a formal two-proportion test needs about 199 per agent to detect that 10 point gap at 95 percent significance with 80 percent power. The method for that calculation is in the e-Handbook's section on sample sizes for testing proportions.
The overlap, from the same Wilson formula:
| Calls scored per agent | Agent A observed 90 percent | Agent B observed 80 percent | Verdict |
|---|---|---|---|
| 20 | 69.9 to 97.2 percent | 58.4 to 91.9 percent | Indistinguishable |
| 50 | 78.6 to 95.7 percent | 67.0 to 88.8 percent | Indistinguishable |
| 200 | 85.1 to 93.4 percent | 73.9 to 85.0 percent | Just separated |
Three things follow, and none of them require more reviewers.
Stop ranking agents on monthly QA scores. At realistic sample sizes a leaderboard mostly reorders noise, and the people at the bottom this month are disproportionately the people who had an unlucky four calls. The e-Handbook makes the general point plainly: there is no correct sample size without first stating what difference you want to detect and what risk of being wrong you will accept (Sample sizes required).
Pool scores where pooling is legitimate. A team of 15 agents scored at 10 calls each gives you 150 calls about the team, and 135 passes out of 150 puts the team's rate between 84.2 and 93.8 percent. That is a usable number about a procedure even though each agent's own 10 calls told you almost nothing. Most of what QA should fix is a procedure, a knowledge gap or a routing rule, and all three live at the team level.
Use QA findings as a pointer, not a verdict. A low score is a reason to go look at a specific call with a specific agent. It is not evidence about their month.
Why two reviewers score the same call differently
Because a scorecard is a measuring instrument, and instruments vary by operator. Metrology has measured this for decades and splits it in two: repeatability is the variation when the same operator measures the same item twice, and reproducibility is the variation between different operators measuring the same item. Both belong in any claim about measurement quality.
The standard treatment is the gauge repeatability and reproducibility study (NIST and SEMATECH e-Handbook, Gauge R and R studies), and the handbook lists differences among operators as one of its named sources of bias, alongside drift and resolution (Analysis of bias). A QA scorecard has all three problems. Reviewers disagree with each other, they disagree with themselves across a long shift, and a three point scale cannot resolve a difference finer than two points.
Scoring is still worth doing. What it needs first is a measurement of the reviewers, taken before anyone treats a score as a fact about an agent.
How to run a calibration session
Score the same calls independently, then compare. The procedure below takes about an hour a month and is the cheapest thing in this article.
- Pick five calls that are not easy. Include one auto-fail candidate, one where resolution is arguable, and one short call that went fine.
- Have every reviewer, including whoever owns the program, score all five alone, with no discussion and no shared document.
- Collect the scores and look at the spread per category before anyone explains themselves. The category with the widest spread is your worst anchor, not your worst reviewer.
- Rewrite that anchor on the spot. If two experienced people read the same behavior differently, the wording is at fault.
- Repeat on the same five calls a month later with no one rereading their old scores. Reviewers who no longer agree with themselves tell you the scale is too fine.
Do this before you publish a scorecard, not after someone's score is challenged. A standard that has never been calibrated will be defended by whoever wrote it and distrusted by everyone else.
What changes when every call is already transcribed
The sampling constraint disappears, and the useful question changes from how many calls to which calls. If your platform produces a transcript, a summary and an outcome tag for every contact, you no longer have to pull calls at random to get a representative view. You can read the whole population for the signals, and spend reviewer time only on calls the data has already flagged.
On Telvana, every call an agent handles is recorded and transcribed, with a summary, sentiment, outcome tags and analytics by agent and day. Tagging is automatic across four outcomes: transferred, booked, voicemail and needs follow-up. Transferred calls carry their transcript and summary in the dashboard, so a reviewer can see why each handoff happened without replaying the audio. A workspace per team or client program keeps that reporting separate, which matters for an outsourcer running QA for several clients at once under one account. What none of that does is score the calls for you. The standard, the weights and the judgment stay yours.
Spend review time this way once you can see every call instead of a sample.
| Signal in the call record | What it usually means | What to actually review |
|---|---|---|
| Transferred after a long attempt | The agent tried and could not resolve it | Whether the knowledge or the routing was the gap, not the agent |
| Second contact from the same number inside a week | The first contact did not resolve | The first call, which nobody ever pulls |
| Handle time in the top tenth for that call reason | An edge case, or a procedure that does not fit the work | The procedure |
| Handle time in the bottom tenth for that call reason | Possibly closed before the caller was done | The disposition and the notes |
| Negative sentiment with no transfer and no follow-up | The caller gave up on being helped | The moment the call turned |
| Tagged needs follow-up with no follow-up logged | The work fell out of the process after the call ended | The handoff, not the call |
The second row is worth more than the rest of the table. A repeat contact means the first call did not resolve, so that first call is the failure, and it is the one nobody reviews, because random sampling has no way to find it. Linking contacts to a phone number is all it takes to pull it.
Reading handle time as a signal needs one caution. A long call is not a bad call and a short one is not a good one, so that column says review the procedure instead of the agent. The failure modes of that particular number are in the guide to average handle time.
How to tie QA scores to the rest of your numbers
Join QA to resolution measures, not to speed measures. Scoring quality against handle time creates pressure to end calls, and the two numbers then move together for a reason that has nothing to do with quality. The safe joins are repeat contact rate, transfer rate and abandonment, because each one is a consequence a QA finding can plausibly explain.
A worked example, with illustrative figures and no benchmarks in it. A team's QA scores look flat month over month, while repeat contacts on billing questions rise. Random sampling will not catch this: billing is one reason among many, and four calls per agent are unlikely to include the failing ones. Pulling the first call behind every repeat contact in that reason does catch it, and the finding is usually a procedure that sends people away with the wrong expectation. The scorecard then tells you what to fix in it.
If part of your volume has moved to AI agents, score those calls on the same scorecard instead of a separate one. The categories hold, the auto-fails hold, and an agent that books an appointment incorrectly fails accuracy exactly the way a person does. The relevant difference is that a change to an AI agent's instructions applies to every future call, so a QA finding there compounds in a way that coaching never does. Before any of it reaches a caller, our team tests the agent on your call types and you approve it, which is a QA step in its own right. For the pre-launch version of this work, see how to test conversational AI before customers use it, and for where agents fit in a center that already runs, see the guide to adding AI agents to a call center. The rest of this cluster is in the metrics and operations guides.
Frequently asked questions
What is QA in a call center?
QA is the structured review of recorded contacts against a written standard, scored and fed back into coaching and process changes. It covers whether the caller's problem was resolved, whether what the agent said was correct, whether the procedure was followed, and how the interaction sounded.
How many calls should you score per agent?
Enough to support the decision you intend to make. For coaching a specific behavior, a handful of calls chosen for a reason is fine. For claiming that one agent performs differently from another at a 10 point gap, the arithmetic puts it near 200 calls per agent, which is why most programs should pool scores at the team level instead.
What should be on a call center QA scorecard?
Resolution, accuracy, process adherence, routing and manner, weighted so resolution and accuracy carry most of the score, with written anchors for each score point. Compliance behaviors belong on a separate binary auto-fail list rather than inside the weights.
What is a good call center QA score?
There is no published number worth copying, because a score only means something against the scorecard and the calibration that produced it. A 90 on a loose scorecard with uncalibrated reviewers tells you less than a 70 on a tight one. Compare your scores to your own scores, over periods long enough to carry a usable sample.
How often should you calibrate QA reviewers?
Monthly, and always before a scorecard is used for a decision about pay or performance. Rerun the same calls periodically so you can see whether reviewers still agree with themselves as well as with each other.
Do you still need QA when an AI agent takes some of the calls?
Yes, and it gets easier, because every one of those calls has a transcript and an outcome tag instead of a sampling frame. Score them on the same scorecard your people are scored on, and treat findings as instruction changes, not coaching notes.
See which of your calls would score well
Book a demo and bring the call reasons your QA program keeps flagging. We will show you which of them a Telvana agent can take, let you hear it handle one, and walk through the call record it produces: the recording, transcript, summary and outcome tag you would be reviewing.