Customer support quality assurance is an ongoing program for reviewing conversations, improving guidance, and helping agents deliver accurate, useful service. It combines clear standards, review coverage, consistent scoring, coaching, and follow-up on recurring problems.
A fast response is not necessarily a correct one. A friendly conversation can still contain unsafe advice, and an escalation can be the right decision. A useful QA program examines what happened and why, instead of treating one score as a complete verdict on an agent.
This guide explains how to build a support QA program, with an illustrative scorecard, sampling plan, and review worksheet. For documenting a particular activity, see quality assurance procedures; for comparing performance with a relevant reference, see customer service benchmarking.
What support QA should review
Review both the interaction and the process supporting it. Did the agent understand the request, give accurate advice, protect information, and own the next step? Did the agent have the documentation, access, and authority needed to do that work?
ASQ distinguishes QA from quality control: QA builds confidence through planned activities, while QC focuses on fulfilling requirements through operational checks. In support, the program and preventive changes are QA work; reviewing an individual answer is one of its quality checks. Teams often use “support QA” for the combined practice.
Include relevant channels, roles, languages, and case types. A rubric built only for a short chat may not fit an asynchronous engineering investigation. Distinguish what the agent controlled from product failures and policy decisions they could not change.
1. Define observable quality standards
Translate broad goals such as “be helpful” into behavior a reviewer can identify. For example: confirm the affected environment before giving version-specific steps, explain the expected result, or name the owner of an escalation.
Involve experienced agents, reviewers, and the people who own important policies. Include examples of acceptable behavior and common exceptions. A customer who has already supplied a detail should not be penalized for receiving a reply that does not ask for it again.
Define serious failure categories separately, such as disclosing private information or recommending an unauthorized consequential action. Follow your approved response process when a review uncovers one; do not let a high tone score average it away.
2. Build a scorecard with evidence
The following is an illustrative starting point, not a universal standard. Use criteria appropriate to your service, and record the evidence behind each rating.
| Criterion | What the reviewer checks | Example evidence |
|---|---|---|
| Understanding | The answer addresses the actual request and relevant context | Accurate summary and use of previous attempts |
| Accuracy | Advice matches current documentation and the customer’s environment | Relevant reference, verified steps, clearly stated uncertainty |
| Clarity | The customer can understand the next action and expected result | Ordered steps, prerequisites, useful language |
| Care and tone | The reply acknowledges relevant impact and stays respectful | Specific acknowledgement without blame or empty promises |
| Ownership | The next action, owner, and update commitment are clear | Complete handoff or follow-up plan |
| Required safeguards | Applicable information, access, and approval rules were followed | Approved verification and no unnecessary disclosure |
A simple scale can be 2 = met, 1 = partly met, 0 = not met, and N/A = genuinely inapplicable. Define what “partly met” means for each criterion. Use N/A for a criterion that does not apply, not for a failure or missing evidence.
For an unweighted score, divide earned points by the maximum points for applicable criteria, then multiply by 100. If five criteria apply, their maximum is 10 points. An illustrative score of 8 points is 80%. Keep serious failure flags and the reasons visible beside that summary.
If you use weights, document them and how exclusions affect the denominator. Do not compare scores from different rubric versions as though the calculation stayed the same.
3. Plan review coverage without hiding selection bias
Define the population and period you are reviewing. Include routine sampling as well as targeted reviews of important cases. These serve different purposes.
| Review stream | Purpose | Reporting caution |
|---|---|---|
| Routine selected sample | Understand ordinary service across relevant groups | Document selection and coverage; do not assume it represents unreviewed groups |
| Targeted high-risk cases | Investigate serious errors, complaints, or unusual events | Do not mix the failure rate with routine samples without explaining the selection |
| New-agent or changed-workflow reviews | Check learning and adoption | Report the specific cohort and case types |
| Calibration cases | Align reviewer interpretation | Keep practice ratings distinct from agent performance records |
Review complete context, not just a convenient snippet. Multi-agent conversations need clear attribution: distinguish a specific person’s message from an inherited case outcome. Zendesk QA’s reviewer guidance, for example, explains different attribution for whole-conversation and specific-message reviews. Check the behavior of your own tool.
There is no universal number of reviews per agent that makes every finding reliable. Choose coverage based on the question, risks, volume, variation, and available review capacity. Report counts and gaps alongside scores, especially before making individual comparisons.
4. Calibrate reviewers on the same cases
Have reviewers independently assess the same conversations with the same scorecard. Compare the ratings and evidence before discussing what each criterion means. The purpose is shared interpretation, not simply forcing scores to match.
Zendesk’s calibration documentation describes reviewers rating the same batch and comparing scores and comments. In its calibration workflow, those scores do not affect the workspace’s Internal Quality Score. A manual process can also keep training cases separate from formal review records.
Record disagreements, revise unclear definitions, and add examples. Repeat calibration after meaningful rubric changes and when differences appear between reviewers. Make the outcome available to the team so agents understand how they are evaluated.
Illustrative disagreement: One reviewer marks an escalation as poor problem-solving; another treats it as correct because the action required specialist authority. The rubric should distinguish a justified escalation from an avoidable handoff.
5. Turn reviews into specific coaching
Give feedback about an observable action, its consequence, and an alternative. Link the conversation evidence and the relevant guidance. “Be more empathetic” is difficult to practice; “acknowledge the blocked task before requesting another screenshot” is concrete.
Ask the agent what context affected the case. Missing access, unclear policy, and conflicting documentation may require a process fix rather than coaching alone. Provide a route for challenging an inaccurate review and record the final decision.
Agree one useful practice action and a follow-up point. Use sanitized examples for exercises and check later cases for application. The customer service training manual can help organize guidance and practice.
Recognize careful diagnosis, accurate escalation, and useful knowledge contributions. Avoid rewards based only on speed, volume, or a score that hides serious failures.
6. Connect QA findings to customer outcomes
Use operational and customer measures to complement reviews, not to replace them. A satisfaction survey captures respondents’ views; a rubric assesses your defined standards; a time metric measures the interval between specified events. They answer different questions.
- Repeat contacts or reopenings: Investigate whether the original result was incomplete, misunderstood, or affected by a new issue.
- Response and resolution time: Review delays alongside accuracy and case complexity.
- CSAT: Report the scale, response count, and response rate; do not assume all customers responded.
- Escalation patterns: Distinguish appropriate specialist involvement from missing frontline guidance.
- Recurring rubric failures: Check whether the cause is training, policy, access, or product behavior.
Our CSAT calculation guide explains consistent denominators. Use the operational efficiency guide to pair speed and resource measures with quality.
A rising QA score after coaching does not establish causation if the cases or rubric changed. Compare similar groups and record other influences. Keep category-level findings visible so a summary score does not hide a growing accuracy problem.
7. Choose tools around the review workflow
A spreadsheet and approved access to conversations may support a small program. Dedicated tools can add assignments, scorecards, calibration, feedback, and reporting. Choose based on the work you need to perform rather than an assumed maturity level.
Check integration coverage, attribution, permissions, retention controls, exports, and support for disputes. If you review recordings, use approved access and exclude unnecessary private information. Screendesk provides recording and helpdesk video workflows that may supply visual case evidence; it is not a substitute for the rubric or review decision.
AI can help identify conversations to inspect or propose ratings. It can also miss context or apply a criterion incorrectly. Validate outputs against reviewed cases, investigate disagreements, and keep a human route for consequential feedback. Automation does not inherently remove bias or guarantee objective scores.
A support QA review worksheet
Conversation and date: Agent / message or whole-case scope: Channel, language, and case type: Review stream and selection rule: Scorecard version: Evidence available and missing: Criterion / rating / evidence: N/A reason, if applicable: Serious failure flag and immediate response: Customer outcome and unresolved issue: Coaching action or process correction: Owner and follow-up date: Agent response / disputed finding: Final review decision:
8. Close the loop on systemic problems
Summarize recurring findings for the team that can act on them. Include the category, examples, scope, customer impact, and proposed investigation. Protect private case details when sharing evidence.
Assign an owner and review whether the correction addresses the original problem. Update the rubric, documentation, and training when a product or policy changes. The continuous improvement examples offer ways to structure a focused experiment.
Common questions about support quality assurance
Is a high CSAT score enough to show support quality?
No. It is useful customer feedback, but depends on who responds and what they are rating. Review accuracy, required safeguards, and case outcomes as well.
How often should reviewers calibrate?
Use a cadence that suits review volume and variation, and repeat calibration after important rubric changes or persistent disagreement. Document what changed rather than holding meetings without updating guidance.
Should AI scores determine agent performance?
Validate their accuracy and context before relying on them. Keep evidence, reviewer oversight, and a challenge route available, especially when feedback affects consequential decisions.
