QA Score Formula: Weighting and Auto Fails

Last updated: September 12, 2026

How a QA score is calculated

A QA score is points earned divided by points available on one call, shown as a percentage. An auto fail overrides that arithmetic and returns zero no matter what the rest of the call earned. Criteria that could not apply come out of the denominator, and an agent's score averages their scored calls over a period.

Division is the easy half. Say a card runs 100 points across 17 criteria and the agent earns 91, which reads 91% and gets argued about by nobody.

Now the same agent skips the recording disclosure on a call that required it. The card still shows 91 points earned and the score is zero, because that line sits on the critical fail list and one flag voids the call. Most of the confusion about QA scores lives in that override, where the number stops being arithmetic and becomes a rule somebody wrote down.

Weighted sections and where the weighting happens

Weight lives in the point values rather than in the division. A criterion worth 8 outranks one worth 4 because somebody decided its failure costs about twice as much, and the formula itself knows none of that. Sorting criteria by what their failure costs is the whole method, and it sits on how to build a call scoring rubric.

One hypothetical call against a 100 point card:

91 divided by 100 is 91%, with no section math required, because the weighting was spent when the points were assigned.

Plenty of programs run the other version, scoring each section out of 100 and multiplying by a section weight. Discovery at 72%, closing at 60%, the other four at 100%, weighted 15, 25, 20, 25, 10 and 5 percent, lands on 91% again. The two agree while the weights match the point spread, and they diverge as soon as a criterion cannot be scored, since one reweights inside its section and the other pulls those points off the whole card.

The auto fail sits outside the arithmetic

A critical fail is a flag, so a correct implementation returns a hard zero before the division ever runs. In a spreadsheet that is an IF wrapped around the division, covered on the Excel scorecard template.

Writing the auto fail as a large negative is the common mistake. Put minus 100 in a spare column and our 91 point call scores negative 9, which averages into the month as worse than a call that earned nothing at all. Double counting is the quieter version: the missed disclosure already cost 5 points on the compliance criterion, so deducting again says one thing twice and gives a disputing agent two cells to argue over.

Something real does get lost in the flat zero. A call handled well apart from one breach and a call that fell apart from the greeting both land on zero, so the ranking between them disappears. Counting flags separately as a critical fail rate recovers it.

Typical auto fails: a TPMO disclaimer that never arrived before plan benefits came up, a Scope of Appointment collected after the plan discussion, a callback number never read back digit by digit that turns out wrong, a dispatch decision that ignored the on-call schedule for that hour. Four to six items is our working range rather than a published standard, and everything on a short list is something you would escalate the same day.

Criteria that do not apply, and the denominator that moves

A call that never reached a billing question cannot be scored on the billing criterion. Mark it not applicable and take its points off both sides of the division.

Say that 8 point criterion comes off a 100 point card. The denominator is 92, the agent earns 85 of those, and the score is 92.4%. Grade the identical call against the full 100 and it reads 85%. Seven and a half points apart on one call, and the only difference is whether a reviewer zeroed a criterion or removed it.

The rule worth publishing in your rubric is narrow: not applicable when the call never created the opportunity, zero when the opportunity existed and the agent walked past it. That sounds loose and holds up fine, because it sends the reviewer back to the transcript for the moment instead of back to their own judgment.

Audit that occasionally, because a reviewer with nine calls left before lunch marks the criteria that take thinking as not applicable, and the average drifts up for a reason nobody logged.

Our own cap, editorial rather than published, puts a call missing more than a quarter of its points in an unscorable bucket instead of on the chart beside a full card. Machine scoring makes the applicability call checkable where a criterion's trigger is a named phrase or a countable event, and stays shaky where it turns on tone, which Aircall and StackAI both report.

Rolling one call up to an agent score

Two methods exist and most programs never notice they picked one: average the call percentages, or pool the points and divide once. Three scored calls at 85 out of 92, 78 out of 100, and 44 out of 50 on a short call where half the card did not apply. Averaging the percentages gives 92.4, 78.0 and 88.0, which comes to 86.1%. Pooling gives 207 earned over 242 available, or 85.5%.

Six tenths of a point, which is nothing most months and widens as the denominators move around. Pooling is more defensible on paper, since it weights a call by how much of the card that call exercised. We would still run the mean, because an agent handed their row can recompute it from the calls in front of them.

Auto fails land hardest right here. Four scored calls at 92, 88, 90 and one flagged call at zero average to 67.5, which describes a policy rather than a month of work, so the critical fail count sits in its own column on the per agent rollup.

Rounding and display

Round once, at display, and keep the full value in storage, because rounding each call before averaging shifts the agent's number in a way nobody sees until it lands next to a threshold. Three calls at 84.5 each. Round first, half away from zero, and you get 85, 85, 85 for an average of 85.0, where leaving them alone gives 84.5, and at a pass mark of 85 that decides the month.

Tools disagree on that tie. Excel's ROUND takes 84.5 to 85, and a Python service using the built in round takes it to 84, because that one rounds to the even digit. When a dashboard and a spreadsheet differ by a point on one call, this is usually why.

Display decides the rest. A conditional format reddening anything under 85 flags a call showing 85 on screen, because the stored value is 84.6, and the supervisor holding that screenshot has a fair complaint.

When two QA scores compare, and when they do not

Comparison holds when both scores came off the same rubric version, from reviewers whose spread has been measured, on calls of the same type. Break one and the numbers still line up neatly in a column, which is the dangerous part.

Rubric version hides best. Move a weight mid quarter and every score written before the change stops comparing to every score after it, while the chart keeps drawing one line through both. Stamp the version on each scored call and let changes land on a period boundary.

Reviewer spread takes an afternoon to measure. Two supervisors score the same ten calls separately, and if the gap runs 6 points, an agent sitting 4 points ahead of another is noise wearing a number (running the calibration covers the session).

Call type gets caught least often. A 45 second call answered under a client account script and a six minute billing dispute exercise different opportunities, so an average blending them describes a floor nobody manages.

A dashboard filter spanning two rubric versions produces a clean chart and a confident wrong conclusion, with nothing on the screen to warn anybody. None of this arithmetic can tell you whether the scorecard measures something real either, so a precise 91 on a card of vague criteria is an exact number about nothing.

A call review runs your existing scorecard across a batch of your own recordings and hands the scores back with the rubric version, the reviewer, and the call type attached to each one, for a fixed pilot price.

Frequently asked questions

What is the QA score formula?

Points earned divided by points available on the call, shown as a percentage, with the critical fail check running first and returning zero when a flag is set. Criteria the call never created an opportunity for come out of both sides.

How do you calculate a QA score percentage in a call center?

Add the points awarded across every scored criterion, divide by the points available on that call, and multiply by 100. Holding the card at 100 points makes those two figures identical.

Does an auto fail zero the score or deduct points?

Zero. A deduction big enough to matter double counts the criterion that already lost points, and it falls apart as soon as two flags land on one card.

How do you handle criteria that do not apply to a call?

Remove the criterion from the denominator rather than scoring it zero, since a zero says the agent had the chance and missed it. Publish the distinction in the rubric, because the same call scores several points apart depending on which one a reviewer picks.

What is a good QA score?

SQM Group puts the industry average at 85% and treats 90 to 99% as good, with only about 5% of agents reaching 100. Read those as orientation rather than a target, since a score means as much as the scorecard behind it and the coverage under it, covered on the QA metrics worth running.

Sources

• SQM Group on the 85% average, the 90 to 99% band, and 5% of agents at 100. Third party, unverified.
• Aircall and StackAI on rule based criteria scoring more reliably than tone. Third party, unverified.
• CMS rules behind the TPMO disclaimer and Scope of Appointment examples. Regulatory, pending confirmation.
• Excel and Python rounding behavior on a tie. Product behavior, pending confirmation.
• Trinsic feature sheet and call notes, 2026-08-27 and 2026-09-03. Verified internally.
• The four to six auto fail range, the quarter of a card cap, and the preference for a mean agent score. Editorial judgment.

Related industriesAll industries