What should a support QA scorecard include?

Most scorecards cover four areas: correctness (was the answer right), process (were required steps followed), communication (tone, clarity, empathy), and outcome (was the issue resolved without a repeat contact). Keep criteria observable in the conversation itself; "showed ownership" is coachable only if you define what it looks like in a transcript.

Is reviewing a sample of conversations good enough?

It is better than nothing, and it was the only option for a long time. But sampling reliably catches only frequent behaviors. Rare-but-serious failures, compliance misses, and errors confined to one conversation type tend to slip through. If a mistake shows up in one conversation out of two hundred, a 2 percent sample will almost never surface it quickly.

Does support QA replace CSAT?

No, they answer different questions. CSAT tells you how customers felt, from the minority who respond. QA tells you what actually happened in the conversation, for every conversation if coverage is full. Teams get the most from using them together: QA explains the "why" behind CSAT movement.

What Is Support QA (Quality Assurance)?

A definition of support QA: how it is measured, why it matters, and a worked example. Full disclosure: unitQ publishes this guide and builds QA into its quality platform.

Support QA is the practice of evaluating customer support conversations against a defined quality standard, usually a scorecard that covers accuracy, tone, process compliance, and whether the customer’s problem actually got resolved. The goal is twofold: coach individual agents with specific, evidence-based feedback, and give leadership a trustworthy read on service quality that does not depend on which customers happened to answer a survey.

How support QA is measured

The core instrument is the scorecard. A typical one breaks a conversation into criteria: did the agent identify the real issue, was the information correct, did they follow required process steps (identity verification, disclosures, escalation rules), was the tone appropriate, and did the resolution stick. Each criterion gets a rating, criteria carry weights, and the weighted result becomes the conversation’s QA score.

Traditionally, human reviewers scored a small sample of conversations by hand, often just a few percent of total volume, because manual review is slow. Reviewers also run calibration sessions, scoring the same conversations independently and comparing results, so that a given score means the same thing regardless of who assigned it.

The newer model is AI-based QA, where a model applies the scorecard to every conversation rather than a sample. Humans shift from scoring everything to auditing the AI’s grades and handling the judgment calls. That change matters more than it sounds: with full coverage, QA stops being a spot check and becomes a continuous quality signal you can trend, segment, and alert on.

Why it matters

Sampling misses things. When you review two conversations per agent per week, a recurring mistake can run for months before anyone reviews an instance of it. Full-coverage QA surfaces patterns, an outdated policy being quoted, a workflow step that agents consistently skip, a category of issue that never gets resolved on first contact.

QA also protects agents, not just customers. A fair, consistent scorecard turns coaching from vague impressions into specific moments: here is the conversation, here is the criterion, here is what great looks like. Agents trust scores they can inspect.

Finally, support conversations are one of the richest sources of product signal a company has. QA review is often where teams first notice that a “support quality” problem is actually a product defect generating avoidable contacts. Support QA and product feedback analysis work best when they share a pipeline, which is the design behind tools like unitQ Support.

A worked example

A fintech support team manually reviews about 2 percent of conversations. Scores look fine for a quarter. When they move to AI scoring across all conversations, one finding jumps out: in a small but steady share of refund conversations, agents quote a policy that changed months earlier. No sampled review had caught it, because the error appeared in only a narrow conversation type. The fix is a macro update and a coaching note, and the miss-rate on that criterion drops within weeks. The lesson is not that reviewers were careless; it is that sampling arithmetic was against them.