Last updated: September 9, 2026
What call center QA means
Call center quality assurance is the practice of reviewing recorded customer calls against a written standard, then using the results to coach agents and prove compliance. A QA program defines the scorecard, decides how many calls each agent gets reviewed, calibrates the reviewers, and reports scores to supervisors and clients.
That is the textbook version. The working version in most 30 seat centers is a supervisor who blocks off an hour, pulls two or three calls per agent, scores them in a spreadsheet somebody built in 2019, and emails the tab to whoever asked.
So an agent who took 1,400 calls last month just got judged on three of them, and those three were mostly the ones sitting near the top of the list when the supervisor sat down. Nothing about that is lazy. It is what an hour buys.
Two neighboring jobs get mistaken for QA. Workforce management forecasts volume and builds the on-call schedule that decides who sits on dispatch at 2am. Customer satisfaction surveying asks the caller what they thought afterward. QA reads what the agent did against a standard written down in advance, which is how a center holds a 4.7 star survey average and still fails a client audit on disclosures.
What a QA program actually contains
The scorecard
A scorecard is the written standard: criteria, a point value on each, and auto fails that zero the call regardless of the rest. Weighting is where the internal argument happens. A center that puts 40 points on the greeting and 10 on the callback number read back digit by digit has told its agents what it really cares about. Keep the auto fail list short and specific, something like a missed TPMO disclaimer or a Scope of Appointment nobody captured, because twelve auto fails means most bad calls score zero and you lose the ability to see which ones were worse. Borrowing a 100 point QA scorecard template beats arguing about weights from a blank page.
The sample
SQM Group found that roughly 60% of call centers evaluate five or more calls per agent per month, which sounds fine until you divide it into the agent's real volume and land under 3%. Almost nobody sizes this number on purpose. It gets set by the hours the supervisor has, then defended afterward as a methodology.
The reviewers
In a small center the reviewer is a supervisor working between escalations. Two reviewers landing 15 points apart on the same call is the normal state of an uncalibrated program, not evidence that one of them is bad at the job. Calibration variance tells you whether your scores mean anything, and a monthly session where three reviewers score the same two calls and argue about the gaps does more than another scorecard rewrite. Centers under 50 seats usually hand that job to whoever has the most tenure.
The feedback loop
An unreturned score changes nothing. Coaching is where the program earns its hours or quietly turns into a compliance artifact, and it lands better when the agent can hear the moment being scored instead of reading a number in an email. Build a path for disputes, because an agent who thinks a critical fail was wrong and has nowhere to say so will stop reading scores. Expect disputes to spike the first month a new scorecard goes live, which usually means the scores are being read.
The report
Reporting is what the program looks like to everyone who is not scoring calls. Supervisors want an agent scoreboard and a trend line. Owners want to know whether last quarter's coaching moved anything. Answering services and BPOs need a third layer, since they answer the phone for someone else's brand and each client wants quality on their own account script rather than a blended floor average.
Quality assurance and quality control are different jobs
Quality control inspects finished interactions and flags the ones that fell short. The assurance half changes what produced them, meaning the account script, the training, the criteria, and sometimes the hold policy causing the rudeness you keep scoring. Plenty of programs run QC under the QA name, which is survivable until the same criterion fails every month and nobody has touched the script.
How QA scores are calculated
Points earned over points available, expressed as a percentage, with weighted sections and an auto fail that overrides the arithmetic. Say a 100 point scorecard split across four sections, scored on one inbound call.
| Section | Points available | Points earned | Note |
|---|---|---|---|
| Opening and verification | 20 | 18 | Missed the account name on second verification |
| Call handling and script adherence | 40 | 32 | Skipped two probing questions before taking the dispatch |
| Compliance | 25 | 25 | Disclaimer read, callback number read back digit by digit |
| Closing and documentation | 15 | 11 | Ticket notes thin, no next step recorded |
| Total | 100 | 86 | 86% |
Had the agent skipped the disclaimer, the auto fail would set that same call to zero, which is why auto fails sit outside the weighted points instead of carrying a heavy penalty inside them.
SQM Group puts the industry benchmark average near 85%, treats 90 to 99% as good, and reports that only about 5% of agents reach 100. Compare inside your own program before comparing against that, because a scorecard with six auto fails and a scorecard with one are not measuring the same thing.
How many calls a QA program can actually review
Manual review covers somewhere between 1% and 3% of calls, with some sources putting the range at 2 to 5% (Verint, Gistly, ViciStack). Every vendor in this category quotes that statistic, ours included, so treat it as the setup rather than the case.
The case is what a small sample does to your data. Say an agent takes 1,400 calls a month.
Call Centre Helper's sample size math puts a defensible read at roughly 65 evaluations per agent per month at 90% confidence. Most programs run five. The distance between those two numbers is the distance between a score you can act on and a score you can only quote.
| Manual sample | Every call scored | |
|---|---|---|
| Calls scored per agent per month | 5 (SQM Group) | 1,400 |
| Share of that agent's volume | under 3% | all of it |
| Evaluations for a 90% confidence read | 65 needed (Call Centre Helper) | met |
| A criterion the agent keeps missing | surfaces on a client audit | surfaces the week it starts |
There is a second problem with a small sample, and statistics has nothing to do with it. The calls a supervisor pulls are the calls a supervisor can find, which skews toward recent and already escalated. A quiet 40 second call where the agent took the wrong dispatch and the caller never complained does not surface at all. That failure mode gets its own page in why QA sampling misses the calls that matter.
Scoring every call fixes coverage and immediately creates a different problem, which is that the supervisor now has 1,400 scored calls and the same hour she had before. Full coverage on its own is a bigger spreadsheet. It pays for itself once scores sort by exception, so the four calls carrying a compliance fail rise to the top and the other 1,396 sit underneath as evidence you can hand a client.
Where automated scoring fits
Machine scoring is dependable on criteria that resolve against the transcript, like whether the disclaimer was read in full and where in the call it landed. Tone and empathy are weaker ground, and the published work agrees on that (Aircall, StackAI). A model can confirm the agent said the empathy phrase. Whether the agent meant it stays a human read, and any program scoring empathy by machine without calibrating against human scorers will drift somewhere strange inside a quarter.
A number on a dashboard is an assertion. Link it to the 14 second stretch of transcript that produced it and the supervisor can check it in seconds while the agent gets something concrete to argue with, which is why reviewing scored calls matters more than the score sitting next to it.
Cost is the other half of this. CallSniper's analysis runs on a model built by Trinsic, CPU based, hosted inside Trinsic's own cloud with no public LLM in the path, so scoring the 1,400th call costs about what scoring the first one did. Audio and transcripts stay in that cloud. Starting needs no phone system integration either, since recordings upload in bulk from whatever already stores them, which is the cheapest way to see how CallSniper handles a recorded call before anyone scopes integration work.
How to start a QA program in a week
- Pick one scorecard and freeze it. Four or five sections, 100 points, no more than three auto fails.
- Upload a month of recordings for one team from whatever system holds them today.
- Score a calibration set. Three reviewers, the same five calls, then a meeting about every gap wider than 10 points.
- Publish the first agent scoreboard with calibration variance next to it, so nobody mistakes a rough first month for a verdict.
- Book coaching on the two lowest criteria rather than the two lowest agents. A weak criterion is usually a script problem wearing an agent's name.
None of this touches the account script itself. A scorecard measures whether the agent followed the script, so if the script is missing a verification step, a clean QA program hands you a wall of 96% scores right up until a client audit finds the hole.
Frequently asked questions
What does a QA analyst do in a call center?
A QA analyst scores recorded calls against the program's scorecard and writes the feedback a supervisor delivers. The role also owns calibration sessions and the scorecard itself, including the argument about what each criterion is worth.
What is the difference between quality assurance and quality control in a call center?
Quality control inspects finished interactions and flags the ones that fell short. On the assurance side you change the scripts and the criteria so fewer interactions fall short to begin with.
What is a good QA score in a call center?
The benchmark average sits near 85%, with 90 to 99% treated as good and only about 5% of agents reaching 100 (SQM Group). Compare scores inside your own program first, because two scorecards with different auto fail counts produce different numbers for identical work.
How many calls does QA listen to?
Roughly 60% of call centers evaluate five or more calls per agent per month (SQM Group), which lands well under 3% of a busy agent's volume. A defensible read runs closer to 65 evaluations per agent per month at 90% confidence for an agent taking 1,400 calls (Call Centre Helper).
Do you need a phone system integration to start QA?
No. Recordings upload in bulk from whatever system already stores them, which lets a program start scoring before anyone scopes integration work. Integration is worth doing later, once the scorecard has settled.
What a scored call looks like on your own audio
The first thing on screen is a scored call with each deduction linked to the transcript line that caused it. A demo walks through that on your own recordings in about 30 minutes on Zoom, so bring a week you already have questions about, ideally one with a complaint in it. Six months of your historical audio run against your scorecard is the paid pilot, the longer version of the same meeting, at a fixed pilot price.
Sources
- SQM Group, benchmark QA scores and evaluation volume per agent. Third party, unverified by us.
- Call Centre Helper, QA sample size math and confidence intervals. Third party, unverified by us.
- Verint, Gistly, ViciStack, manual QA call coverage percentages. Third party, unverified by us. Used as setup only.
- Aircall, StackAI, reliability of automated scoring on rule based versus subjective criteria. Third party, unverified by us.
- Trinsic Technologies feature sheet and product calls, private CPU based model hosted in Trinsic's cloud, manual upload onboarding. Verified.