Call Center QA Scorecard Template + Examples

Quick answer

A call center QA scorecard is a form that scores each call against weighted criteria. A good one has six sections worth 100 points (opening, discovery, resolution, communication, closing, documentation), a separate list of critical items that fail the call on their own, and an anchor for every criterion describing what Met, Partly met and Not met sound like. Divide points earned by points available, excluding N/A items. The template below is ready to copy.

Most QA scorecard templates online are a list of headings: "greeting", "empathy", "resolution", each scored 1 to 5. Two reviewers using that list on the same call will give different scores, because nothing tells them what a 3 sounds like. Agents notice, dispute their results and stop trusting QA.

This page gives you a complete inbound customer-service scorecard with points, critical items and written anchors, a worked example of scoring one call, a short variant for AI voice agent calls, and the process around the form: calibration, how many calls to review, coaching, and when automated scoring of every call should replace sampling. If you run outbound campaigns, our outbound call quality assurance guide covers disclosures, opt-outs and outbound compliance items in more depth.

What Goes on a Call Center QA Scorecard?

A scorecard that works has four parts:

Weighted sections that follow the order of a real call, so the reviewer can score while listening.
Criteria that are observable. "Asked at least one question to narrow down the problem" can be checked against the recording. "Showed good attitude" cannot.
Critical items scored pass or fail, outside the points. A call where the agent read out a full card number is not a 92 percent call with one mistake.
Anchors: one sentence per criterion describing Met, Partly met and Not met. Anchors are what make two reviewers agree.

Keep it short. If scoring a 5-minute call takes 20 minutes, reviewers will cut corners or review fewer calls. Sixteen to twenty criteria is plenty for customer service.

The Inbound Customer Service QA Scorecard Template

Copy this into your QA tool or a spreadsheet and adjust the weights to your priorities. The weights below suit a general customer-service line where getting the answer right matters most. A billing team might move points from discovery to documentation; a technical support desk might give resolution even more.

Section	Criterion	Points
1. Opening (10)	Greets, gives company and own name within the first few seconds	5
1. Opening (10)	Takes the caller's name and uses it naturally	5
2. Discovery (20)	Lets the caller explain without interrupting	5
	Asks clarifying questions before proposing a fix	10
	Restates the issue and gets the caller to agree	5
3. Resolution (35)	Gives a correct and complete answer, in line with policy	15
	Takes ownership: does the work on the call, or names who will and when	10
	Resolves on this call, or transfers or escalates to the right place with context	10
4. Communication (15)	Acknowledges frustration or impact before moving to the fix	5
	Uses plain language, no internal jargon or codes	5
	Manages holds and silences: asks before holding, checks back, explains dead air	5
5. Closing (10)	Summarises what happens next, with a timescale	5
5. Closing (10)	Checks whether there is anything else, closes politely	5
6. Documentation (10)	Correct disposition or outcome code	5
6. Documentation (10)	Notes let the next person continue without calling the customer to ask again	5
Total		100

Notice what is not on the card: handle time. Speed belongs on a dashboard, not in a quality score. If agents lose points for long calls, they rush the resolution section, and your repeat calls go up. Our guide to reducing average handle time covers how to cut call length without hurting quality, and the call center KPI formulas guide shows where QA scores sit next to FCR, CSAT and AHT.

Critical items: pass or fail

Any one of these fails the call regardless of points. Report critical fails as their own number (for example, "2 critical fails in 40 calls this month"), because a 0 averaged into the score hides them.

Critical item	Fails when…
Identity verification	Account details were discussed or changed before the caller passed your verification steps
Accurate information	The agent stated a price, policy, date or entitlement that is wrong, or promised something they cannot authorise
Sensitive data handling	Card numbers, passwords or other sensitive data were read back, written in notes or asked for outside the approved process
Complaint and vulnerability handling	A formal complaint or a vulnerable caller was not logged or routed the way your policy requires
Required statements	A statement your rules require, such as a recording notice, was missed
Professional conduct	Rude, dismissive or abusive language, or hanging up on a caller without cause

Keep the list to what is genuinely serious. If ten things are critical, nothing is, and every agent fails something each month.

How to Score Calls: The Scale and the Anchors

Use two kinds of scoring:

Yes/no for facts: critical items, and criteria such as "correct disposition code".
Three levels for skills: Met (full points), Partly met (half points) and Not met (zero). Add N/A for criteria that did not apply, such as hold handling on a call with no hold.

Why not 1 to 5? A five-point scale looks more precise, but reviewers rarely agree on the difference between a 3 and a 4, and that disagreement is what kills trust in QA. Three levels with clear anchors give you more consistent data. If you need finer grading later, add it to one or two criteria, not all of them.

Here are anchors for the criteria reviewers argue about most. Write one for every line on your card.

Criterion	Met	Partly met	Not met
Clarifying questions	Asks at least one question that narrows the problem before suggesting a fix	Asks questions, but only after starting on a fix that turns out to be wrong	Assumes the problem and goes straight to a standard answer
Correct and complete answer	Answer matches the knowledge base or policy and covers every question the caller asked	Correct, but one of the caller's questions is left unanswered	Incorrect or incomplete in a way that will cause a repeat call (if it is a wrong fact, it is also a critical fail)
Ownership	Does the task on the call, or names who will do it and by when	Says someone will follow up, but gives no name, team or timescale	Tells the caller to call back, email or try another department
Acknowledgement	Names the impact in the caller's terms ("so you've been without service since Monday")	Uses a scripted phrase ("I understand your frustration") with no reference to the actual problem	Ignores clear frustration and moves straight to process
Next steps	States what happens next, who does it and when the caller will hear back	States what happens, but not when	Ends the call with no summary

Anchors should describe behaviour you can hear, not a feeling. If your agents follow call flows, link the anchors to your scripts so "Met" means the same thing in training and in QA. Our inbound call center scripts give worked openings, verification and closing lines you can score against.

A Worked Example: Scoring One Call

A customer calls about a double charge. The agent greets well and takes the name, lets the caller explain, but starts explaining billing cycles before asking which payment was duplicated. They confirm the refund correctly, say "billing will sort it" without a timescale, close without asking whether there is anything else, and leave notes that are accurate but thin. There was no hold.

Section	What happened	Points earned / available
Opening	Both met	10 / 10
Discovery	No interruptions (5), clarifying questions partly met (5 of 10), restated issue (5)	15 / 20
Resolution	Correct answer (15), ownership partly met: no timescale (5 of 10), resolved on the call (10)	30 / 35
Communication	Acknowledgement and plain language met; hold handling N/A	10 / 10
Closing	Next-steps summary met (5), no "anything else" (0)	5 / 10
Documentation	Disposition correct (5), notes partly met (2.5)	7.5 / 10
Total	Critical items: all passed	77.5 / 95 = 81.6%

The hold item was N/A, so its 5 points came off the total available: 77.5 divided by 95, not by 100. Had the agent read back the customer's card number, the call would be recorded as a critical fail, with the 81.6% kept only as a note for coaching.

The coaching point writes itself: two of the three lost marks are about the same habit, jumping ahead. One behaviour to work on ("ask which payment before explaining") is better than five.

Suggested score bands

These are starting points, not a benchmark. Set them after your first month of calibrated scores.

90% and above: strong. Use these calls as training examples.
80–89%: meets the standard. Coach one behaviour.
70–79%: below standard. Coaching session within the week.
Below 70%, or any critical fail: immediate review with the team lead.

QA Scorecard for AI Voice Agent Calls

If an AI voice agent answers part of your phone traffic, its calls need QA too. It will not forget to say its name, so most of the human scorecard does not apply. What goes wrong with automation is different: it can mishear and carry on, answer from general knowledge instead of your documents, quote a number nobody approved, say something is booked when it is not, or keep a caller in a loop when it should hand over.

Criterion	Type	What the reviewer checks
Stayed within its knowledge	Critical	Every fact (hours, policies, services) matches your documents. If it did not know, it said so or offered a person.
No invented prices	Critical	Any price spoken is on your price list, in the right currency and range.
Claimed only what was done	Critical	"Booked" or "cancelled" only if the calendar or system shows it. Otherwise it said a person will confirm.
Confirmed what it heard	20 points	Read back the service, date, time and name before acting, and corrected itself if the caller disagreed.
Handed off correctly	25 points	Offered a person when asked, when the request was out of scope, or after repeated misunderstanding, and the hand-off carried the details so the caller did not start again.
Answered every question	20 points	No question from the caller was skipped before the call ended.
Did not make the caller repeat	15 points	No question asked twice when the answer had already been given.
Correct outcome recorded	20 points	The result, callback request or follow-up message matches what was agreed on the call.

Weight hand-off heavily. A caller who gets stuck with an automated agent is the failure your customers remember. Our guide to AI voice agent to human handoff sets out the triggers and what context should travel with the call.

Score these two demo calls yourself

AI voice agent demo — simulated caller: the replies come from DialerBee's production voice agent, the callers are synthetic and the businesses are fictional.

Salon booking in Arabic

AI voice agent demo — simulated caller · Jordanian Arabic · 49 seconds

Clinic questions and a callback

AI voice agent demo — simulated caller · English · 70 seconds

In the salon call, the agent reads back the service, day and time before closing ("so you'd like to book a haircut with a blow-dry for next Thursday at 4 in the afternoon?"), then says the team will check the booking and confirm it. It does not claim the slot is booked, which passes "claimed only what was done". In the clinic call, a reviewer would check the opening hours and the quoted price range against the clinic's own documents and price list, and confirm that "call me back on this number" was saved as the caller's number in the callback request. Transcripts for both are on the AI voice agent page.

How to Run a QA Calibration Session

Calibration is how you make sure a score means the same thing whoever gives it. Without it, an agent's result depends on which reviewer they drew.

Pick 3 to 5 calls: one clearly good, one clearly poor, and the rest borderline. Borderline calls are where disagreement shows up.
Score alone first. Every reviewer, and ideally a team lead, scores the calls independently before the meeting. Scoring together turns calibration into following the loudest voice.
Compare item by item, not just the totals. Two reviewers can both give 80% while disagreeing on half the criteria.
Discuss each disagreement against the anchor. Ask: did the anchor allow both readings? If yes, rewrite the anchor, not the reviewer.
Record the decision with the call as a reference example ("this is Partly met for ownership, because…"), so new reviewers can learn from it.
Repeat monthly, and weekly for the first month of a new scorecard or a new reviewer.

Measuring inter-rater agreement

Two simple measures are enough:

Score range: each reviewer's total for a call should be within about 5 points of the group's average. Reviewers who are consistently high or low need a one-to-one.
Item agreement: the share of individual criteria where two reviewers gave the same rating.

Raw agreement can flatter you, because on criteria that almost everyone meets, reviewers agree by chance. Cohen's kappa corrects for that. A worked example: two reviewers make 40 yes/no decisions (4 calls × 10 items). Both say yes on 26, both say no on 8, and they split on 6. Raw agreement is 34 of 40, or 85%. Reviewer A said yes 75% of the time and reviewer B 70%, so the agreement you would expect by chance is 0.75 × 0.70 + 0.25 × 0.30 = 60%. Kappa = (85% − 60%) ÷ (100% − 60%) = 0.625. That 85% headline hides a real gap, and the items they split on are the anchors to rewrite.

In DialerBee's QA scorecards, several reviewers can score the same call independently in the platform, and the system compares their scores to show the variance between reviewers.

How Many Calls Should QA Review per Agent per Month?

There is no industry standard. When Call Centre Helper asked practitioners (opened October 2026), answers ranged from 4 to 8 calls a month, through 10 and 15, up to 50 to 60 calls per agent per month. The right number depends on what you want the sample to tell you.

Two pieces of maths help. First, small samples miss habits. If an agent skips identity verification on 1 call in 10, the chance that a random sample contains no such call is:

Calls reviewed per month	Chance of missing a 1-in-10 habit entirely	Margin on the monthly average (illustrative)
4	66%	± 9.8 points
8	43%	± 6.9 points
20	12%	± 4.4 points
50	0.5%	± 2.8 points

The first column is 0.9 raised to the number of calls. The second is a 95% margin assuming an agent's call scores vary with a standard deviation of 10 points; your spread may differ. At 4 calls, an agent's 82% could really be anything from about 72% to 92%. That is not precise enough to rank agents or tie scores to pay.

A practical plan for human sampling:

Established agents: 6 to 10 random calls a month for a fair trend, plus targeted calls.
New agents and agents on an improvement plan: 2 to 4 a week for the first month.
Targeted calls on top: long calls, escalations, complaints, repeat callers and negative-sentiment calls. Keep these separate from the random sample in reports, or they drag scores down unfairly.

How Do You Link QA Scores to Coaching?

A score that arrives two weeks later with no conversation changes nothing. A loop that works:

Coach within a few days, while the agent still remembers the call.
Pick one or two behaviours, taken from the criteria the agent lost points on most across several calls, not from one bad call.
Play the moment, not the whole call. Jump to the timestamp and let the agent hear it.
Agree the new behaviour in a sentence the agent can use on the next call.
Re-score after two weeks on the same criteria. If the score has not moved, change the coaching or the anchor.
Look across the team. If most agents lose points on the same item, it is a training, script or process problem, not an agent problem.

In DialerBee, reviewers score with inline playback, annotate specific moments on the recording's timeline and read the transcript alongside the audio. Coaching notes are attached to the scored call, agents see their scores and notes in a self-review dashboard and acknowledge the feedback, and scores are tracked over time by agent, team, campaign and individual criterion. QA data exports to PDF and CSV, with scheduled reports for team leads.

Automated AI Scoring of Every Call vs Human Sampling

Human sampling reviews a small slice of calls. Automated scoring can cover all of them. Neither replaces the other.

	Human sampling	Automated scoring of every call
Coverage	A few calls per agent per month	Every call
Consistency	Depends on calibration and reviewer	Applies the same rules every time
Judgement	Strong on context: was the exception justified, was the tone right for this caller	Weaker on nuance and edge cases; needs checking against human scores
Finding rare problems	Often misses them (see the table above)	Can surface them for a person to review
Cost per extra call	Reviewer time for each call	Low once set up
Best used for	Calibration, disputed scores, coaching, critical-fail decisions	Coverage, trends and choosing which calls a person listens to

The model that works: let automated scoring cover every call and flag the ones worth a listen, and have people score a smaller set in full. Each month, compare automated and human scores on the same calls, the way you calibrate two reviewers. If they diverge on an item, fix the criterion before you trust the trend.

DialerBee scores every call automatically, so quality review covers all calls instead of a sample. Alongside that, auto-flag rules send calls to the QA review queue by duration, disposition or sentiment score, and reviewers can still pick random or targeted calls. Scorecard templates have custom sections, questions and weights, and can differ by campaign or team, so a billing line and a sales team do not share one form.

For AI voice agent calls, each call leaves a result with the recording, the transcript and the outcome, which is what a reviewer needs to score it. Before an agent goes live, you can test it in a simulator, and once it is live a supervisor can listen in and take over a call.

Common QA Scorecard Mistakes

Too many criteria. Thirty items means slow reviews and small samples. Cut to what you will actually coach.
Vague criteria with no anchors. "Professional tone" scored 1 to 5 invites disputes.
Averaging critical fails into the score. Report them separately.
Scoring speed as quality. Keep handle time out of the form.
Mixing targeted and random calls in one average. Targeted calls are picked because something went wrong.
Never changing the form. Review the scorecard every quarter. Drop items everyone always meets; add what calibration keeps arguing about.

Frequently Asked Questions

What should a call center QA scorecard include?

Five or six weighted sections that follow the call: opening, understanding the issue, resolution, communication, closing and documentation. Add a short list of critical items, such as identity checks and accurate information, that fail the call on their own. Give every criterion an anchor that says what Met, Partly met and Not met sound like.

How do you score a call on a QA scorecard?

Listen to the whole call once, then score each criterion against its anchor. Award full, half or zero points, mark criteria that did not apply as N/A and remove their points from the total. Divide points earned by points available. If any critical item failed, the call is a fail whatever the percentage.

What is a good QA score for a call center?

There is no universal benchmark, because the number depends entirely on how strict your anchors are. Many teams treat 85 to 90 percent as the target, with any critical fail counted separately. Watch the trend for each agent and each criterion rather than chasing one figure.

Should I use a 1 to 5 scale or yes/no?

Use yes/no for facts, such as whether the agent verified identity, and a three-level scale of Met, Partly met and Not met for skills. Five-point scales feel more precise, but reviewers rarely agree on the difference between a 3 and a 4, which makes calibration harder.

How many calls should QA review per agent per month?

Practitioners report anything from 4 to over 50 a month. Four to eight calls gives a rough picture but misses habits that happen on one call in ten. Review more for new or struggling agents, and use automated scoring of every call to decide which calls a person listens to.

How often should QA calibration sessions be held?

Monthly is a sensible default, and weekly while a new scorecard or a new reviewer is bedding in. Each reviewer scores the same calls alone, then the group compares results item by item and rewrites any anchor that allowed two readings.

Can AI score every call instead of human reviewers?

AI scoring can cover every call, which sampling never does, and it is consistent. It is weaker on judgement, such as whether an exception to policy was the right call. The practical model is AI scoring for coverage and flagging, people for calibration, disputed scores and coaching.

How do you QA an AI voice agent's calls?

Use a shorter scorecard focused on what can go wrong with automation: did it confirm what it heard, answer only from your documents, quote only your prices, claim an action only when it was done, and hand over to a person when it should. Treat invented facts and invented prices as critical fails.

The Bottom Line

A QA scorecard is only as good as the agreement between the people using it. Start with the template above, write anchors for every criterion, keep critical fails separate, calibrate monthly, and review enough calls that the score means something. Use automated scoring to cover every call and point people at the ones that matter. If you want scorecards, recordings, transcripts, coaching notes and automated scoring of every call in the same system as your queues and AI voice agent, see DialerBee QA scorecards and the full call center software, book a demo, or start a 14-day free trial with no credit card. Plans and prices are on the pricing page, and the AI receptionist page shows what the voice agent handles on inbound calls.

Call Center QA Scorecard Template: A Ready-to-Use Form with Weights, Critical Items and Scoring Anchors