How Many Calls Should QA Review?

“How many calls should we monitor?” usually gets a inherited answer: four a month, or two a week, or 2% of volume. Those numbers were set when listening was expensive. They are not a measurement plan.

What a small sample can and cannot say

A handful of calls can catch a disastrous fatal error if you are lucky. It cannot tell you whether an agent is an 82 or an 91. The range is too wide. Ranking people, or putting them on a plan, from four observations is how QA loses the floor.

For a yes/no line (passed the disclosure or not), a rough worst-case sample at 95% confidence is:

• about 100 evaluations to be within ±10 points
• about 380 evaluations to be within ±5 points

That is the campaign, not each agent. Per agent, you will rarely afford ±5 from manual listening alone. Be honest about that in the report. Say “directionally lower on openings” rather than “ranked 14th”.

A practical split

Use three layers, not one quota.

1. Census on the lines that can hurt you. Every call is eligible for the fatal and compliance checks you can automate or draft with AI Auto-QA. A missed disclosure should not depend on being drawn.
2. A fixed human sample per agent for the judgement lines (clarity, resolution, control of the call). Enough that a bad week is visible — often one or two a week, scored by a person, not eight rushed at month end.
3. A targeted pull for complaints, repeat contacts, long handle times, and new hires in their first month. This is where random sampling is the wrong tool.

Do not confuse coverage with listening time

100% QA versus sampling is a workflow question. An AI draft of every interaction, with a person signing the risky and the sampled ones, is how teams get the census layer without hiring a QA analyst per team.

Write the number down: evaluations per agent per month, who scores them, and which lines are census. If you cannot say it in a sentence, the programme does not have a sample design yet.

Model the analyst time you get back when drafts replace blank forms — QA ROI calculator.