Last updated: September 12, 2026
What counts as a good QA score
SQM Group puts the industry average QA score at 85% and treats 90 to 99% as good, with only about 5% of agents reaching 100. Those figures are third party and unverified. A score means something only against the scorecard that produced it, so a 92 and a 78 can describe the same agent on the same day.
That second part is the one people skip. Ten criteria, generous partial credit, nothing sitting in the auto fail column, and a competent agent lands in the low 90s and stays there all year. Thirty criteria, half of them binary, four auto fails, and the same agent comes in around 78 while the card does exactly what it was built to do. The number moved and the calls did not.
So the benchmark is worth knowing because your client or your board will quote it at you. It is worth almost nothing as a target.
Why the rubric decides the number
Four things about a scorecard set the range your floor can land in before anybody picks up a phone. Criteria count, weighting, partial credit, and the length of the auto fail list.
Partial credit is the quiet one. Say you inherit a card with 10 criteria worth 10 points each and every criterion allows a half mark. A reviewer who is not sure gives 5, because 5 is defensible and 0 starts an argument. Across 40 calls that one habit holds a team in the high 80s no matter what happened on the phone.
The arithmetic, including how an auto fail zeroes a call and how weights become a percentage, lives in the QA score formula.
When a score looks wrong the obvious repair is to tighten the card. About half the time the card is fine and two reviewers are nine points apart on the same call, which no amount of rewording fixes. Run a calibration session before you touch a single criterion.
A pass mark and an average are different numbers
A pass mark is the line one call has to clear to be acceptable. An average is what a month of calls comes to after the bad ones get absorbed by the good ones. Setting 90 as "the target" without saying which one you mean produces a fight at every review meeting.
The two also drive different behavior. A published pass mark makes an agent care about their worst call, since one call under the line is visible. Publish the average instead and they care about volume, because a good Friday covers a bad Tuesday.
For most small programs the pass mark is the more useful of the two, and it is the one almost nobody writes down.
What the distribution shows that the average hides
Say two teams of 12 both report 88 for the month. Everybody on the first team sits between 84 and 92. The second has nine agents at 96 and three at 64.
| Shape behind the 88 | What it means | What to do about it |
|---|---|---|
| Twelve agents clustered between 84 and 92 | Nobody is failing and nobody is excelling, which usually points at the card or the script rather than the people | Check which criteria never separate anybody, then look at the script |
| Nine agents at 96 and three at 64 | Three people are handling calls badly right now, and the average is hiding them behind the other nine | Coach the three by name this week, off their own flagged calls |
Same 88. Completely different week for a supervisor.
Pull the median and the bottom quartile alongside the mean and the shape stops hiding. The mean is the least useful number on the dashboard and the one that ends up in the client report anyway.
We went looking for a published benchmark on healthy distribution shape and did not find one, so the target has to come from your own history.
When a very high average means the card is too easy
If most of your floor sits at 96 to 99, weigh that against SQM Group reporting that only about 5% of agents reach a perfect score. A quarter of your team at 100 is a statement about the scorecard more than the team.
There is a cheap test. Pull every criterion and look at how often it was failed in the last 90 days. Any criterion with a 100% pass rate across a full quarter is either a solved problem or something an agent cannot fail, and either way it adds nothing to the score. Rewrite it, or move the weight to criteria where calls genuinely differ. A scorecard built from scratch is often less work than defending an inherited one.
One honest limit on scoring in general: machine scoring holds up well on binary rule based criteria and gets shakier on tone and empathy, which Aircall and StackAI both note about automated QA. A card weighted heavily toward empathy will produce more arguable scores regardless of who or what does the scoring. In CallSniper every scored line carries a transcript timestamp and a confidence score, so a supervisor can spend their review time on the flags the model was least sure about.
Auto fails bend the distribution out of shape
An auto fail takes the whole call to zero. A call that would have scored 94 misses the recording disclosure in the opening seconds and books a 0, and that 0 goes into the same average as everything else.
What comes out is two populations mixed together: an ordinary spread in the 80s and 90s, plus a spike sitting at zero. Averaging across both gives a number that describes neither, and the month reads as a slump when it was four broken calls.
The length of that list matters as much as the calls do. A program with 12 auto fails reports a lower average than an identical program with 4, on the same floor and the same calls. Keep the critical fail rate as its own line, which is one of the QA metrics worth running next to the score itself.
Watch your own number move instead of chasing 90
The usable version of all this is short.
- Freeze the rubric for at least a quarter. Every edit resets the baseline, and the months before it stop being comparable.
- Track median and bottom quartile per agent next to the mean, and read all three together.
- Compare this month to your own last month. A comparison against another company's number is meaningless unless you also have their card and their reviewers.
- Stamp the card version on every report, so a jump can be traced to a rubric change instead of being celebrated.
None of that fixes a bad script, and a scorecard never will. When every agent fails the same criterion, the problem is upstream of the floor.
If you want to see where your own distribution actually sits before changing anything, a first pass on your own audio scores a batch of your recent recordings against the card you already use and hands back the spread rather than one number, for a fixed pilot price.
Frequently asked questions
What is a good QA score in a call center?
SQM Group reports an industry average of 85% and describes 90 to 99% as good, with about 5% of agents reaching 100. Treat it as context rather than a target, since the number depends on how strict your scorecard is.
Is an 85% QA score good?
By SQM Group's own band it is average rather than good. Whether 85 is strong for your program depends on how strict the card behind it is.
What should we set as our pass mark for a single call?
No published standard held up when we went looking. Take the median of a month of scored calls on your current card, set the line from there, and leave it alone long enough to mean something.
Why did our QA average drop after we rewrote the scorecard?
Because the scorecard changed, most likely. Scores are only comparable within a card version, so a rewrite starts a new baseline and the first month under it is not a performance result.
Can an AI QA score be compared to a human QA score?
Only on the same rubric, and the two line up more closely on rule based criteria than on tone. The practical check is to have a supervisor score a set of calls blind, then compare criterion by criterion against the machine result.
Sources
- SQM Group on the 85% industry average, the 90 to 99% good band, and about 5% of agents reaching 100. Third party, unverified.
- 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 two team distribution example. Illustration built for this page, not a published benchmark.