Last updated: September 12, 2026
How many calls you have to score
This calculator answers one question: how many calls you have to score before the number that comes out is worth quoting. Give it your call volume for the month, a confidence level, a margin of error, and the minutes one review takes. It will not tell you which calls to pull.
The formula, so you can check the number
The calculator runs the standard sample size for a proportion, then corrects it for the fact that your month contains a finite number of calls. Two steps, both doable on paper.
The first step ignores your volume entirely and asks how much evidence the confidence level demands. Square the z score, multiply by 0.25, and divide by the margin of error written as a decimal.
| Confidence level | z score |
|---|---|
| 90 percent | 1.645 |
| 95 percent | 1.96 |
| 99 percent | 2.576 |
At 90 percent confidence with a 10 point margin, that comes out to 67.7 calls, which rounds up to 68. Notice that your volume has not entered the arithmetic yet.
The second step brings it in. Divide 67.7 by one plus 66.7 divided by your monthly calls. For an agent taking 1,400 calls a month the answer is 64.6, so 65 calls. Sampling from a small pool is easier than sampling from a huge one, which is why the correction always pushes the number down.
The 0.25 is the widest case, the pass rate that demands the most calls. Plug in a pass rate you already measured and the sample gets smaller, and we left that option out because an unmeasured pass rate is a guess that makes the sample look cheaper than it is.
What 65 calls per agent per month actually costs
Call Centre Helper published roughly the same figure, about 65 evaluations per agent per month at 90 percent confidence for an agent handling 1,400 calls. Treat that as unverified third party arithmetic that happens to match what the formula returns.
Scoring 65 calls at five minutes each is 5.4 hours per agent per month. Run twelve agents and you have 780 calls and 65 reviewer hours, which is a week and a half of somebody's time doing nothing else.
SQM Group puts 60 percent of call centers at five or more evaluations per agent per month. Five and 65 are not the same program. A team scoring five calls per agent is running a coaching ritual, which has its own value, and it is not producing a number anyone should put in a client report.
One correction before you quote the output anywhere. The formula sizes a sample for a proportion, meaning the share of calls that pass or the share carrying a critical fail. An average QA score is a mean and needs different arithmetic, so treat this sample as a close approximation when the mean is what you report.
Tightening the inputs costs more than people expect
Take the same agent at 1,400 calls and ask for 95 percent confidence with a 5 point margin. The sample jumps to 302 calls, 21.6 percent of the month, and 25.2 reviewer hours. At 99 percent and 5 points it is 451 calls, roughly a third of everything the agent said that month.
Most of that increase comes from the margin, not the confidence level. Holding confidence at 90 percent and moving the margin from 10 points to 5 takes the sample from 65 calls to 227. A 10 point margin on a 90 percent pass rate means the true rate sits somewhere between 80 and 100, which is a wide enough band that two agents can look identical when one of them is not.
Volume barely helps. A floor taking 38,000 calls a month needs 68 calls at 90 percent and 10 points, which is 0.2 percent of the month and 5.7 hours of review time at the same five minutes each.
That last number is where QA programs get fooled. A sample sized against 38,000 floor calls describes the floor and says nothing about one agent, one account, or the overnight shift, so an honest per agent program multiplies the sample by headcount.
Where a sample answers the question and where it cannot
Sampling is the right tool for measuring a rate. How the floor scores this month against last, whether new hires trail the veterans, whether two supervisors agree when you check calibration variance: those are rate questions, and a sample answers them at a cost you can plan around.
Compliance is a different problem wearing similar clothes. CMS requires third party marketing organizations to record their marketing, sales, and enrollment calls, and the TPMO disclaimer has to be delivered before any discussion of plan benefits. A sample that says 94 percent of calls carried the disclaimer has told you about the other 6 percent only that they exist somewhere in the month.
Nobody gets to explain to a regulator that the Scope of Appointment failure was outside the sample. Same story for a two party consent disclosure, a callback number read back wrong, or any criterion your scorecard treats as an auto fail. Those need every call checked, which is the argument for machine scoring rather than a bigger sample: rule based criteria score reliably against a transcript at any volume, and tone and empathy stay weaker, which Aircall and StackAI both note about automated QA.
So the split is by criterion rather than by program. Sample the judgment calls, cover the rules completely, and size the sample above for the part of the scorecard a human has to listen to.
Frequently asked questions
What is a good sample size for quality assurance?
Whatever the calculator returns for your volume at a confidence level you can defend, which for most call centers lands between 60 and 80 calls per agent per month. Anything under about 30 is too loose to rank agents against each other.
How many calls should QA review per agent each month?
Around 65 for an agent taking 1,400 calls. Most rooms score five, and that gap is the reason QA numbers get argued with, so staff what you can and put the margin in the report.
Can I size the sample for the whole call center instead of per agent?
You can, and the sample barely grows, which is the trap. A floor level sample supports floor level statements only, so anything you plan to say about an individual agent or a single client account needs its own sample for that population.
What margin of error should I use?
Start at 10 points and tighten only where a decision depends on it. Termination, bonus payouts, and client facing reports are worth 5 points and the reviewer hours that come with it, while a monthly coaching sweep rarely is.
Why does it assume a 50 percent pass rate?
Variance peaks at 50 percent, so the sample it asks for holds up no matter what the real rate turns out to be. A measured pass rate from a few months of scoring brings the number down, while a guessed one buys you a confidence interval you never earned.
The part the sample size cannot fix
Sixty five calls chosen badly is still a biased read. A reviewer who opens the first calls in the queue gets Monday mornings, short calls, and whichever accounts sort to the top, and the resulting 87 percent describes that slice rather than the month.
Randomization matters more than the count. Number the calls and pull them with a random number, or stratify across shifts, accounts, and call types first and sample inside each bucket, and the 65 calls you score start meaning something about the other 1,335.
Selection bias is also the cheapest defect in QA to fix. Your recording platform can almost certainly export a call list, so a team that cannot afford a bigger sample can still afford an unbiased one.
Scoring every call removes the question, which is what CallSniper does with the recordings you already keep. The paid pilot runs six months of your own audio for a fixed pilot price, so you can hold the score from the sample you have been drawing against the score from the whole month and see how far off it was.
Pair this with the QA scorecard template and the rest of the QA resources when you are rebuilding the program rather than just sizing the sample.
Sources
- Call Centre Helper on about 65 evaluations per agent per month at 90 percent confidence for 1,400 calls. Third party, unverified.
- SQM Group on 60 percent of call centers evaluating five or more calls per agent per month. Third party, unverified.
- Aircall and StackAI on rule based criteria scoring more reliably than tone. Third party, unverified.
- CMS requirements for third party marketing organizations, including call recording and the disclaimer before any discussion of plan benefits. Verified by search 2026-09-08, pending a recheck against the current CMS text.
- Sample size for a proportion with the finite population correction. Standard method, checkable by hand from the formula above.
- Call volumes, review minutes, headcounts, and the scores used to illustrate a margin. Editorial hypotheticals.