Quality assurance in most contact centres reviews between one and three percent of interactions, several days after they happened, against a scorecard filled in by a person who was not there. Nobody designed that. It is what you get when reviewing is expensive and volume is large.
What sampling actually measures
At two percent, most agents are assessed on a handful of conversations a month. That is a sample far too small to separate a genuinely weak performer from an unlucky one, which everyone on the floor knows, which is why QA scores are argued with rather than acted on. You are not measuring quality; you are measuring which conversations got picked.
It also systematically misses the thing you most want to catch. Rare failures — the compliance slip, the promise nobody could keep — are rare, and a two percent sample is almost designed not to find them.
What changes at a hundred percent
Scoring every closed interaction automatically against your own criteria changes the character of the function. Reviewers stop hunting for cases and start adjudicating the ones flagged. Coaching stops being an argument about whether the sample was fair. And the rare failure becomes findable, because you looked at all of them.
Three conditions, or it backfires
- Your scorecard, not a default one. Categories, attributes and weights have to match how your operation actually defines good, or you have automated the wrong judgement at scale.
- Humans keep the final say. AI generates the draft assessment and highlights the risk; a reviewer confirms, adjusts or overrides. Automated scoring that closes its own loop will be resented and eventually ignored.
- Agents can dispute. A formal route to challenge a score is what makes the whole thing legitimate on the floor. Without it you have surveillance, not quality.
The second-order benefit
Once every interaction is scored, quality data becomes operational data. Patterns by queue, by product, by branch and by time of day stop being anecdotes. The most useful output of full-coverage QA is usually not a better agent ranking — it is the discovery that a particular process, not a particular person, is generating the complaints.