A QA rubric is the set of criteria a support team uses to score a customer conversation, with a description of what each rating level looks like and how much each criterion counts. It turns “good service” into line items that two reviewers can grade the same way. The QA score is what you get when you apply the rubric to one conversation, and a QA scorecard is the form that carries both. Teams also call it a quality rubric or quality assurance rubric.
In short
- A QA rubric has five parts: criteria, a rating scale, level descriptions, weights and a short list of critical items.
- The level descriptions are what make it a rubric rather than a checklist, because they tell reviewers what “partly met” actually means.
- The QA score is the points earned divided by the points possible, and how you treat “not applicable” can move it by ten points.
- Most support teams do well with 5 to 8 criteria that can be checked from the transcript alone.
What a QA rubric contains
A QA rubric contains five parts: the criteria being judged, the rating scale for each one, a written description of every level on that scale, a weight that sets how much each criterion counts, and a few critical items that fail the conversation outright. Remove any one of them and reviewers start filling the gap with personal judgment.
The word comes from education. Carnegie Mellon’s teaching center defines a rubric as “a scoring tool that explicitly represents the performance expectations for an assignment” and notes that rubrics help ensure consistency across graders in courses with several of them. A support team with ten reviewers has exactly that problem.
| Part | What it answers | Example |
|---|---|---|
| Criterion | What are we judging? | “The answer was correct and complete” |
| Rating scale | How many outcomes are possible? | Met, partly met, not met, not applicable |
| Level descriptions | What does each outcome look like? | “Partly met: correct, but one needed detail is missing” |
| Weight | How much does it count? | 30 of 100 points |
| Critical item | What fails the whole conversation? | Account data shared before identity was verified |
Critical items are usually called an auto-fail: one miss sets the score to zero whatever happened elsewhere. In software, the rubric is the set of criteria and descriptions that QA scorecard software applies to each conversation.
QA rubric vs scorecard vs checklist
A QA rubric, a QA scorecard and a QA checklist describe the same activity at different levels of detail. The rubric is the grading logic, the scorecard is the form a reviewer fills in, and a checklist is a list of yes or no items with no level descriptions and usually no weights. Many teams use the words interchangeably.
| QA checklist | QA scorecard | QA rubric | |
|---|---|---|---|
| Main question | Did it happen? | What score did it get? | What does each score mean? |
| Rating | Yes or no | Yes or no, or a scale | Scale with a description per level |
| Weights | Rarely | Usually | Usually |
| Best for | Process and compliance steps | Day to day reviewing | Agreement between reviewers, disputes, automation |
In practice your scorecard is the rubric plus the form around it: the conversation link, the reviewer, the comments and the final score. If your scorecard has criteria and weights but no written description of each level, you have a scorecard that still needs its rubric.
An example QA rubric for chat and email support
Here is a complete QA rubric for written support on chat and email: six weighted criteria worth 100 points, a three-level scale with a description of each level, and two critical items. It is an illustration to adapt, not a standard. Change the criteria to match your products, policies and contact reasons.
| Criterion | Weight | Met (full points) | Partly met (half) | Not met (0) |
|---|---|---|---|---|
| Understood the real issue | 15 | Restates or acts on the actual problem, asks a clarifying question when the request is ambiguous | Answers the literal question but misses the underlying need | Answers a different question than the one asked |
| Correct and complete answer | 30 | Information is accurate and covers everything the customer needs to act | Accurate, but one needed detail is missing | Wrong, outdated or misleading information |
| Followed the required process | 20 | Every required step for this contact reason is done (refund policy, escalation path, tagging) | A minor step is skipped with no effect on the customer | A required step is skipped and the customer is affected |
| Clear next step and timing | 15 | Says what happens next, who does it and by when | Next step given, timing missing | Customer is left to guess what happens next |
| Tone fits the situation | 10 | Acknowledges frustration when present, plain language, no blame | Polite but generic or templated where it should not be | Dismissive, blaming or confusing |
| Notes good enough for the next person | 10 | Internal note or ticket fields let a colleague continue without rereading the thread | Notes exist but miss a key fact | No notes on a conversation that needs a handoff |
Critical items (score set to zero): account or personal data shared before identity was verified; a refund or credit promised outside policy.
Two things make this a rubric rather than a list. Each level has a description a reviewer can point to in the transcript, and “notes” can be marked not applicable when no handoff is needed. For ready-made versions by channel, see our QA scorecard templates.
How a QA score is calculated from the rubric
A QA score is the points a conversation earned divided by the points it could have earned, shown as a percentage. In most setups, criteria marked not applicable come out of both numbers, and a critical item sets the score to zero. Averaging those scores across conversations gives an agent’s or team’s quality score.
Take one email scored with the rubric above:
| Criterion | Weight | Rating | Points |
|---|---|---|---|
| Understood the real issue | 15 | Met | 15 |
| Correct and complete answer | 30 | Partly met | 15 |
| Followed the required process | 20 | Met | 20 |
| Clear next step and timing | 15 | Not met | 0 |
| Tone fits the situation | 10 | Met | 10 |
| Notes for the next person | 10 | Not applicable | removed |
The conversation earned 60 points out of a possible 90, so the score is 67%. Look at what the “not applicable” rule does. If the team counted it as a fail, the score would be 60 out of 100, or 60%. If it counted as a pass, 70 out of 100, or 70%. Same conversation, a ten point spread, purely from a rule nobody wrote down.
That is why the N/A rule belongs in the rubric itself. The weights matter just as much: a slip in tone costs 5 points here while a wrong answer costs up to 30. Our guide to weighted QA scorecards shows how to set those numbers from evidence, and the internal quality score post covers how the per-conversation scores roll up into one team number.
Analytic or holistic: choosing a QA rubric format
There are two rubric formats. An analytic rubric scores each criterion separately and then combines them, which is what almost every support QA rubric does. A holistic rubric gives one overall rating against a single description of quality. Analytic takes longer to build but tells the agent exactly which part of the conversation to fix.
The University of Oklahoma’s assessment office sums up the tradeoff: an analytic rubric’s “criterion can be weighted to reflect the relative importance of each dimension”, while with a holistic rubric “criteria cannot be weighted”. The same page notes that analytic rubrics take more time to develop and apply.
For support teams the choice is usually easy:
- Use analytic for agent QA, coaching and anything tied to a target or a bonus. Agents need to know which criterion cost them points.
- Use holistic for quick health checks, such as a manager rating a batch of conversations “good, needs work, poor” to spot a problem before building a full rubric.
Which quality assurance criteria belong in a rubric
A criterion belongs in a QA rubric when it passes four tests: a reviewer can see it in the conversation, it checks one thing, the agent controls it, and it applies to the contact type being scored. Criteria that fail these tests produce the scores agents dispute most, because two fair reviewers can reach different verdicts.
- Observable. “Showed empathy” is a feeling. “Acknowledged the customer’s frustration before answering” is something you can quote.
- One check per criterion. “Greeted the customer and verified their identity” is two criteria. Split anything joined by “and”.
- In the agent’s control. A slow tool or a policy the agent cannot change should not cost the agent points. Track it as a process issue instead.
- Fits the contact reason. Rapport building makes sense on a complaint and very little on a password reset. Scoring a customer who only wants to make a payment on whether the agent built rapport tells you nothing about that conversation.
On numbers, keep the list short: every extra criterion adds review time and another place for reviewers to disagree, so most teams do well with five to eight. If you plan to automate scoring, the bar is higher still; our guide on how to write criteria an AI can score rewrites common vague criteria into checkable ones.
How to build a QA rubric in six steps
Building a QA rubric takes six steps: pick the outcomes it should protect, list the behaviors that drive them, write each as an observable criterion, describe every rating level, set weights and critical items, then test it on real conversations with several reviewers before agents see a single score.
- Start from outcomes. Write down the three or four things a good conversation must achieve for your business: a correct answer, a resolved issue, a compliant process, a customer who knows what happens next.
- List the behaviors behind them. Read 20 to 30 recent conversations, good and bad, and note what agents did that made the difference.
- Write the criteria. Turn each behavior into a single, observable statement using the four tests above.
- Describe each level. For every criterion, write what met, partly met and not met look like, with a short example from a real transcript.
- Set weights and critical items. Give the most points to what hurts customers most. Keep critical items to the two or three failures that are truly non-negotiable.
- Test before rollout. Have three or four reviewers score the same ten conversations independently and compare. Where they disagree, the description is unclear, not the reviewer.
Step 6 is the one teams skip, and it is where most rubrics fail. If two reviewers can give the same ticket very different scores because “showed empathy” means something different to each of them, agents stop trusting the number. Regular QA calibration keeps that gap closed after launch. Very different contact types, such as billing questions, bug escalations and refund requests, may also need separate rubrics because they do not share criteria.
When a formal QA rubric is not for you
A full weighted rubric with level descriptions is not worth building for every team. If one or two people handle a few dozen tickets a week, a manager reading conversations and giving direct feedback covers the same ground with less overhead. A short yes or no checklist is enough until disagreements about quality start to appear.
A rubric also will not fix a process problem. If agents fail the same criterion across the board, the cause is usually a policy, a tool or training, and scoring it harder only makes the scores look worse. Build the rubric when you have several reviewers, agents asking why they lost points, or a plan to score conversations automatically.
Apply one rubric to every conversation
A rubric only pays off when it is applied the same way to every conversation, and that is where manual review runs out. A reviewer reading a small sample cannot score consistently at volume, and each reviewer drifts a little over time. Automated scoring applies one written standard to all of them, with nobody tired or rushed.
Kaizo takes your existing criteria rather than imposing a template, and scores 100% of your support conversations against them, whether a person or an AI agent answered, with the evidence linked to each score. When you change a criterion’s description, the change applies to every future evaluation, so nobody has to re-brief a team of reviewers. At SteelSeries, AI summaries cut QA time in half.
See it in a live demo on realistic sample data: Kaizo scores 100% of support conversations against your own rubric and turns the findings into coaching. Let’s talk.
Frequently asked questions
What does QA rubric mean?
QA rubric means the scoring guide a quality assurance team uses to grade customer conversations. It lists the criteria, the rating scale, what each level looks like and how much each criterion counts. “Quality assurance rubric” and “quality scorecard” are common names for the same thing.
What is a QA score?
A QA score, sometimes called a QA rating, is the result of applying the rubric to one conversation: points earned divided by points possible, as a percentage. Criteria marked not applicable are usually removed from both numbers, and a critical item sets the score to zero. Team and agent quality scores are averages of these per-conversation scores.
How many criteria should a QA rubric have?
Most support teams do well with 5 to 8 criteria, each checkable from the transcript. Fewer than that rarely covers accuracy, process and communication together, and much longer lists are slow to review and hard to score consistently. Teams with very different contact types often keep a short rubric per type rather than one long one.
Should a QA rubric use a yes or no scale or a 1 to 5 scale?
Use yes or no for steps that either happened or did not, such as identity verification or a required disclosure. Use a three-level scale (met, partly met, not met) for behaviors with real degrees, such as answer completeness. Five-point scales are only worth it when each point has its own written description.
How often should you update a QA rubric?
Review the rubric when products, policies or contact reasons change, and at least a few times a year. Change one criterion at a time and note the date, because every change shifts the scores and makes trends before and after the change harder to compare.
Can a QA rubric be used for AI agents?
Yes. The same rubric can score conversations handled by an AI agent and by people, which keeps quality comparable across both. The criteria need to be observable in the transcript, which is good practice for human review anyway.







