CONTENTS
How many calls do QA teams actually review?What a 5-call sample can catchThe sample size the math asks forSampling that works harder, if you're reviewing by handWhat the rules and your buyers expect you to keepScoring every call instead of sampling- Most QA teams review 3 to 5 calls per agent per month, and there's no industry standard for the number. On a busy outbound floor that's well under 1% of an agent's calls.
- If an agent skips a disclosure on 2% of calls, a 5-call random sample misses it about 90% of the time. You'd need roughly 149 random calls per agent per month to be 95% sure of catching it once.
- Floors that can't listen to that many calls get more out of targeted pulls (transfers, rejected leads, very short and very long calls) or out of scoring every call automatically and sending people only the failures.
Ask a call center whether it audits its recordings and you'll hear "we record 100% of calls." That answers a different question. Recording is a dialer setting; auditing is a person, or a system, actually checking what was said, and on most floors that happens to a handful of calls per agent each month while the rest sit in storage until a complaint drags one out.
Call center QA sample size is the number of recorded calls a quality team reviews per agent in a given period. The most common figure, per Call Centre Helper, is 3 to 5 calls per agent per month. This piece runs the numbers on what a sample that size can see, and what it would take to see more.
How many calls do QA teams actually review?
Fewer than most operators would guess, and the number has no rulebook behind it. Call Centre Helper puts the most common figure at 3 to 5 evaluated calls per agent per month, says plainly that "there are no industry guidelines" for the ratio of evaluators to agents or the number of calls, and suggests a 10-call baseline month for teams starting out.
The reason is time. A reviewer has to listen to the call, often more than once, then fill in a scorecard and write feedback. Supervisors who also run the floor do this between huddles and escalations, so the sample shrinks to whatever fits in the week.
Now put that against volume. Say an agent handles 40 recorded calls a day across 21 working days: that's 840 calls a month. Five reviewed calls is 0.6% of them. Scale it to a 50-seat floor and you have about 42,000 recorded calls a month with 250 of them heard by QA.
So the honest answer to "do teams audit their recordings or just store them" is: they audit a thin slice and store the rest.
What a 5-call sample can catch
A small sample is fine for coaching. It shows whether an agent sounds confident, follows the call flow and handles objections, because those habits show up on most calls. It breaks down on problems that happen occasionally, which is exactly where compliance risk lives: the skipped recording disclosure, the "yes" that wasn't really a yes, the agent who keeps pitching after "take me off your list."
The math is simple. If a problem happens on a share p of an agent's calls and you pull n calls at random, the chance you see none of them is (1 − p) to the power n.
| Calls reviewed per agent per month | Problem on 1% of calls | Problem on 2% of calls | Problem on 5% of calls |
|---|---|---|---|
| 3 | 97% chance you miss it | 94% | 86% |
| 5 | 95% | 90% | 77% |
| 10 | 90% | 82% | 60% |
| 20 | 82% | 67% | 36% |
| 50 | 61% | 36% | 8% |
Read the 5-call row. An agent who skips a required disclosure on 2 calls in every 100 (about 17 calls a month at 840) gets a clean QA month nine times out of ten. Their score looks perfect while the problem keeps happening on live calls, and the first time anyone hears one of those 17 calls may be when a buyer or a consumer's lawyer asks for it.
The table uses the simple formula. Sampling without replacement from a fixed 840 calls barely changes it: the exact chance of missing all 17 problem calls with 5 picks is 90.3%.
The sample size the math asks for
There are two different questions, and they need very different samples.
"Would I notice if this agent has a problem?" To be 95% sure of seeing at least one problem call, you need n where (1 − p)ⁿ is below 0.05:
- a problem on 5% of calls needs about 59 random calls
- on 2% of calls, about 149
- on 1% of calls, about 299
"What is this agent's fail rate?" Estimating a rate within plus or minus 5 percentage points at 95% confidence needs about 385 calls from a large population. With 840 calls in the month, the finite-population correction brings that down to about 264, which is close to a third of everything the agent said.
Nobody staffs QA for 149 calls per agent per month by hand. At 50 seats that's 7,450 calls, or several full-time reviewers doing nothing else. That's why sampling exists, and also why a clean sample says less than it seems to.
Sampling that works harder, if you're reviewing by hand
If manual review is what you have, pick calls on purpose instead of at random. A random pull spreads attention evenly across thousands of routine calls; a targeted pull points it at the calls most likely to cost you money.
Pull these first:
- Every transfer a buyer rejected or disputed. These calls already have a dollar figure attached, and the recording is your side of the argument.
- Calls under about 30 seconds and calls far longer than your average. Very short calls hide skipped disclosures and hang-ups after a pitch; very long ones hide agents freelancing off script.
- A new agent's first two weeks, at a much higher rate. Turnover means a constant stream of people who haven't absorbed the script yet.
- Calls where the transcript contains "stop", "remove", "don't call" or "not interested" followed by more pitching. In VICIdial you can find candidates by disposition and call length before anyone listens.
Then calibrate. Have two reviewers score the same five calls each week and compare; if their scores drift apart, the scorecard wording is vague. Our free call center QA scorecard template keeps critical checks, like a missing consent "yes", outside the weighted average so a call can't pass on points while failing on compliance.
What the rules and your buyers expect you to keep
The FTC's Telemarketing Sales Rule asks telemarketers and sellers to keep records for 5 years. Since the 2024 amendments, 16 CFR 310.5 covers a detailed record for each call: the seller, the numbers, the date, time and duration, the scripts used and the outcome, including transfers. Our Telemarketing Sales Rule checklist has the full list.
Buyers ask a sharper question than regulators do. When a buyer rejects a transfer as unqualified, they want that specific call, and they want to believe it. A recording nobody reviewed and a file you could have edited put you on the back foot, which is the problem we cover in why a call recording isn't evidence.
Scoring every call instead of sampling
The other way out of the sample-size problem is to stop choosing. Software can score every recorded call against your own checklist and send reviewers only the calls that failed, which turns 42,000 calls into a short list a supervisor can get through before lunch.
Klariqo does this on VICIdial. You connect the dialer with read-only access (setup takes about 10 minutes), write your checklist (disclosures, script lines, consent language, banned phrases) and every recorded call gets scored, human agents and AI agents alike. A failed check quotes the exact line from the transcript, so the reviewer reads the problem before deciding whether to listen. Each call also becomes a Klariqo Call Record, signed and given a government-grade timestamp from DigiCert, so when a buyer asks about a call you send a record they can check themselves.
Call QA is $0.05 per call with the record included, and your first 10,000 recorded calls get their record free. See the scoring in action on our call QA page, or read how it fits a VICIdial floor in QA 100% of your VICIdial calls.
Connect VICIdial and score every call →
Last updated 2026-11-10 by Ansh Deb, Founder & CEO of Klariqo. Sources: Call Centre Helper, QA staff-to-agent ratio · 16 CFR 310.5, Telemarketing Sales Rule recordkeeping. Sample-size figures are our own arithmetic, using the formulas shown. General information, not legal advice.
Questions people ask
How many calls should QA review per agent per month?
The most common figure is 3 to 5, according to Call Centre Helper, and there's no industry standard. That's enough to coach habits that show up on most calls. To be 95% sure of catching a problem that happens on 2% of an agent's calls, you'd need about 149 random calls a month.
What is a good QA sample size for a call center?
It depends on the question. Spotting an occasional problem at 95% confidence takes roughly 59 calls (if it happens on 5% of calls) to 299 calls (if it happens on 1%). Estimating an agent's fail rate within plus or minus 5 points takes about 264 calls out of 840.
Do call centers actually listen to their recordings?
Most review a small sample, typically a few calls per agent each month, and keep the rest for disputes and record-keeping. On a floor where an agent handles 840 recorded calls a month, 5 reviewed calls is 0.6%.
Is random sampling or targeted sampling better for call QA?
For compliance, targeted. Pull rejected transfers, very short and very long calls, new agents' calls and calls where the customer asked to stop. Random sampling is better for measuring an agent's general quality.
How long do telemarketers have to keep call records?
Under the FTC's Telemarketing Sales Rule, 16 CFR 310.5, most required records must be kept for 5 years. State rules and your buyer contracts may add their own requirements.
Can software review 100% of calls?
Yes. Klariqo scores every recorded call on VICIdial against your own checklist for $0.05 per call, quotes the exact line behind each failed check, and signs each call into a record your buyer can verify.
