Short answer: There isn't one number. How many calls you QA depends on the question you're trying to answer. Center-level trends need a few hundred to a couple thousand calls a month. Fair agent coaching needs dozens of calls per agent. Catching a rare compliance miss can take hundreds of calls per agent, which is why compliance items often get a much larger sample.

I get asked this on almost every demo. "How many calls should we be scoring?"

My answer is always the same... what are you trying to find out?

Most QA programs pick a number like 5 or 8 calls per agent because that's what the team can get to. That's a staffing number. It's not a sampling plan. We've already covered the statistics behind QA sampling and evaluations per agent. This post is the practical part: how to build a plan around the decisions you actually make.

Start With the Question, Not the Quota

Each call you score should help answer one of four questions:

  1. How is the center doing overall? (trend reporting, client scorecards)
  2. How is this agent doing? (coaching, reviews)
  3. Are we missing something serious? (compliance, rare but costly failures)
  4. Is this new thing working? (new hire, new program, new script)

Each question needs a different sample. When you mix them into one quota, you end up with too many calls for some questions and way too few for others.

Question 1: How Is the Center Doing?

This is the easy one. For a center-wide score, you need a few hundred to a couple thousand randomly selected calls a month, and that number barely moves as volume grows.

Here's the math for a few example volumes. It uses the standard sample size formula for a proportion, with a finite population correction (NIST explains the method here). These are worked examples, not benchmarks.

Example monthly calls 95% confidence, ±5% 95% confidence, ±3% 99% confidence, ±3%
10,000 370 965 1,557
20,000 377 1,014 1,688
40,000 381 1,040 1,762

Notice the pattern. Doubling your volume barely changes the number of calls you need. Center-level QA is cheap to get right.

The catch: the calls have to be picked at random and spread across agents, shifts and call types. "Grab a few from this morning" isn't random.

Spread the sample like your volume

Random is step one. Proportional is step two. If your evening shift takes 30% of your calls, about 30% of your sample should come from the evening shift. Same for queues, call types and agents.

Here's an example. Say you plan to score 400 calls this month, and 25% of your volume is billing calls. Then about 100 of those 400 should be billing calls. If you let the sample drift toward whatever's easy to grab, your center score ends up describing the easy calls, not your center.

For BPOs: every program is its own center

If you run multiple client programs, each one needs its own sample. One big blended number tells a client nothing about their program. Plan each program separately, and put the sample size right on the client QA report.

Question 2: How Is This Agent Doing?

This is where most programs fall apart.

To say something about one agent with any confidence, you need a lot more calls than most teams score. Example: an agent handles 1,000 calls a month. To estimate their pass rate on a question within ±10 points at 90% confidence, you'd need to score about 64 of their calls. At 95% confidence, about 88.

Compare that to the 4 to 8 calls a month that a lot of centers score per agent. With 5 calls, one bad call is 20% of the sample. That agent's score can swing hard based on which 5 calls got picked. Agents know that. It's why they argue about every score.

What to do about it

Question 3: Are We Missing Something Serious?

This is the question nobody does the math on, and it's the one that hurts you.

Say an agent skips a required disclosure on 1 out of every 100 calls. How many calls do you have to score to have a good shot at catching it even once?

The math is simple: the chance you miss it on each call you score is (0.99) raised to the number of calls. Here's what that works out to:

If the problem happens on... Calls to score for a 90% chance of seeing it once For a 95% chance
5 of every 100 calls 45 59
1 of every 100 calls 230 299
1 of every 200 calls 460 598

Now flip it. If you score 8 calls a month for that agent and the problem shows up on 1 of every 100 calls, you have about an 8% chance of seeing it that month.

So you're not finding it. Your customers are. Or a regulator is.

That's the real case for a much larger sample on compliance items. You don't need a huge sample to know your average. You need it to catch the rare stuff. That's especially true in collections. Here's what a collections team should score on compliance.

A practical split

You don't have to treat every question on your form the same way:

Question 4: Is This New Thing Working?

New hires, new programs and script changes need more coverage up front. You're looking for problems before habits set in.

A simple rule I like: score heavier for the first few weeks of anything new, then step down to your normal plan once the scores settle. You want to find out on week one that a new hire isn't verifying callers, not at their 60-day review. (Here's how I'd run QA through nesting.)

Put It on One Page

Your sampling plan doesn't need to be fancy. Write down:

  1. Your questions. Which of the four are you answering, and for whom?
  2. Your sample for each. Center, per agent, compliance items, new hires.
  3. How calls get picked. Random, spread across agents, shifts and call types. Not hand-picked.
  4. What happens with the results. Who coaches, when, and how you check that it worked.

Then compare that plan to your QA capacity. If the plan says you need hundreds of calls per agent on compliance items and your team can score 8, that gap is your real QA problem.

Where OttoQA Fits

Most teams can't close that gap with people. The listening alone eats the day.

That's where OttoQA comes in. OttoQA scores calls against your own QA forms, so the listening stops coming out of your team's day. That makes it realistic to run both plans: a much larger sample for compliance items, a smaller one for everything else.

No seat licenses, no IT, and setup takes about 2 weeks. There's a $1,000 one-time setup fee, and you can start with a 200 to 500 call proof of concept: we score 200 to 500 of your calls against your form.

FAQ

How many calls should you QA per agent per month?

Enough to make a fair call about that agent. For an agent handling about 1,000 calls a month, that's roughly 64 calls to estimate a pass rate within ±10 points at 90% confidence. Most centers score far fewer, so agent-level scores swing a lot.

How many calls should you score?

It depends on what you need to know. For center trends, a random sample of a few hundred to a couple thousand calls a month is enough. Fair agent coaching takes dozens of calls per agent, and catching rare compliance misses can take hundreds of calls per agent.

Does the sample size grow with call volume?

Barely. In the examples above, going from 10,000 to 40,000 calls a month moves the 95% confidence, ±5% sample from 370 to 381 calls. What drives the number is your confidence level and margin of error.

Should we hand-pick bad calls to review?

Review them, yes. But keep them out of your sample. Hand-picked calls skew your scores. Run a random sample for measurement, and handle complaint calls and escalations as a separate review.

Can a BPO use one sample across all clients?

No. Each client program needs its own sample and its own form. A blended number doesn't tell any one client how their program is doing.

About the author

Tom Laird is the founder of OttoQA and CEO of Expivia, with about 30 years in contact centers. He hosts the Advice from a Call Center Geek podcast. Connect with Tom on LinkedIn.