100% Call QA versus Random Sampling

Random sampling is a compromise from an era when a person had to hear the whole call. You accepted that most interactions would never be seen, and you hoped the bad ones were evenly sprinkled through the sample. They are not. Complaints, new-hire calls and end-of-day fatigue cluster. A random 2% misses them on purpose.

What 100% should mean

It should not mean a robot files a score nobody reads. A useful version has two steps:

• A draft is produced for every interaction you choose to include — values, comments, and the fatal lines.
• A person signs the draft when the consequence is real: a fail on a compliance line, a new hire, a dispute, a customer who came back, or the agent’s scheduled sample.

That is Auto-QA with a human in the loop. The analyst’s job moves from “listen and type” to “check the evidence and decide”.

What you should still sample on purpose

Judgement lines stay human-led. Empathy, control of a messy conversation, and whether the explanation would make sense to a customer are exactly where a draft needs a reviewer. Keep a scheduled sample so every agent is heard by a person on a cadence, even if their AI draft looks clean. Clean drafts can be confidently wrong.

What you should stop sampling

Do not sample your way through legal disclosures, card data, or threat language if you can check those on the full set. A compliance score that only exists on the lucky draws is not a control. It is a story you tell the regulator.

How to explain it to agents

Say this: every call can be scored, a person signs anything that affects your record, and you can see the evidence. Hidden 100% scoring with no appeal is how you get a grievance, not a better floor. The dispute path is part of the design — see how agents should dispute a QA score.

Teams comparing tools should ask a blunt question: does the product draft the whole form for sign-off, or does it only highlight phrases? The comparison index is written around that difference.