Last updated: September 12, 2026
What an agent scorecard shows
An agent scorecard summarizes one agent over one period instead of one call. The row carries calls handled, calls scored, the share of their calls the score rests on, the average with its movement against last period, the critical fail count, and the section that lost the most points. Coverage decides whether the average is worth discussing.
This document exists for a conversation about a person. Somebody opens a folder of these on the second Tuesday of the month with a dozen agents to get through, or writes the case for a promotion, or starts a performance plan. One call's score carries none of that, however carefully it was scored.
Two readers decide whether the sheet holds up. The agent gets handed it and reads every line back, usually starting with the line they disagree with. HR comes later, if the row ever turns into a performance plan or a termination file, and their question is always what the number is built on.
So a per call scorecard grades a call and this one grades a period, which is why no criteria appear on it anywhere. The criteria behind the average stay on the page that owns them.
The agent scorecard
Sixteen fields, one row per agent per period. Every value below is hypothetical, and the month it describes is deliberately not a flattering one.
| Field | What it holds | September example |
|---|---|---|
| Agent | The person the row belongs to | Agent A |
| Period | Fixed start and end dates, the same window every month | 2026-09-01 to 2026-09-30 |
| Scorecard version | The rubric version in force for the whole window | v1.2, unchanged since 2026-06-01 |
| Calls handled | Every call the agent took in the window | 612 |
| Calls scored | Calls that got a completed evaluation | 14 |
| Coverage | Calls scored divided by calls handled | 2.3% |
| Average score | Mean of the scored calls, with auto fails already zeroed | 88.0 |
| Prior period average | The same rubric version, previous window | 85.6 |
| Change | This period minus last period | +2.4 |
| Critical fails | Scored calls carrying at least one auto fail | 1 |
| Disputes raised | Scores the agent formally challenged | 2 |
| Disputes resolved | Challenges closed, with the outcome recorded on each | 2 closed, 1 score changed |
| Coaching sessions held | Documented sessions against the sessions the scores called for | 2 of 3 |
| Top deduction criterion | The one criterion that lost the most points across the scored calls | Asked at least two probing questions before proposing a solution |
| Development focus | One behavior for next period, written as an observable event | Two probing questions before the first solution, on every billing call |
| Next review date | When this row gets read back to the agent | 2026-10-14 |
Fourteen scored calls against 612 handled is 2.3% coverage, which is roughly what one supervisor gets through in a month alongside everything else the job involves. The 88.0 is honest arithmetic on those 14 calls and still a number to hold loosely, for reasons the last section of this page works through.
Score by rubric section
An average tells you a problem exists somewhere. The section split tells you where to go looking, and it takes about four extra columns in the file.
| Section | Points available | September | August | Change |
|---|---|---|---|---|
| Opening and identification | 15 | 14.2 | 14.0 | +0.2 |
| Discovery and listening | 25 | 21.0 | 19.4 | +1.6 |
| Communication quality | 20 | 17.8 | 17.5 | +0.3 |
| Resolution and accuracy | 25 | 21.6 | 21.2 | +0.4 |
| Compliance and process | 10 | 8.7 | 8.9 | -0.2 |
| Closing | 5 | 4.7 | 4.6 | +0.1 |
| Total | 100 | 88.0 | 85.6 | +2.4 |
Discovery and listening gained the most ground and is still the weakest section against what it is worth, sitting at 84% of the points available while opening sits at 95%. Walk into a review holding "you're at 88" and there is nothing to coach. The discovery number gives you one behavior, one criterion, and two calls to play.
Download the rollup
The rollup downloads as a CSV or a spreadsheet, with no form in front of either. Columns match the table above, plus each section score and its prior period value, one row per agent per period. Our example row ships filled in so the shape is visible, and it should come out before you file a real month.
Per call scoring and this rollup belong in the same workbook on different tabs, which is the spreadsheet that holds both tabs.
Reading the trend against last period
The change column earns its place only when both periods ran on the same rubric version. Move eight points out of communication quality and into compliance halfway through a quarter, and every agent who was weak on tone improves on the chart without changing a word they say on a call. A reweighting shows up in the review as a coaching win, and the supervisor delivering it has no way to know.
That is why the version sits on the row, and why the first period under a new version gets reported with no trend at all.
The rule is easier to write than to keep. A rubric frozen for a year is usually one nobody is maintaining, and the criteria that produce arguments do need rewriting. The compromise worth defending is changing the instrument between quarters rather than inside one, then re baselining out loud, and how to build a call scoring rubric covers the change log discipline this depends on.
Before you compare two agents
Two agents scored by two different reviewers are not on the same scale until somebody checks the variance. The check is blind double scoring: both reviewers take the same ten calls separately, then you compare the spread per criterion. No published threshold for acceptable calibration variance held up when we went looking on 2026-09-08, so the signal to watch is a criterion where the two reviewers disagree every single time, which is a wording problem rather than a reviewer problem.
Stack ranking a floor before that check produces an argument nobody wins. The agent at the bottom asks who scored them, and the honest answer is that it mattered. Calibration before you compare reviewers has the session format and the tracker.
The one to one agenda
Five moves off this sheet, in this order, and the meeting runs about twenty minutes.
- Read the coverage number out loud before the average, because it sets what the rest of the conversation is allowed to conclude.
- Put the section table in front of them, then the top deduction criterion.
- Play two calls cued to their timestamps, one where the behavior shows up clean and one where it goes missing.
- Take the dispute if there is one. A score the agent cannot challenge stops functioning as a measurement inside two months.
- Agree one development focus, written as an event somebody can watch for, with the next review date filled in before anybody stands up.
The agent leaves holding the same sheet you do. A rollup delivered out loud becomes whatever each of you remembers by the time the next one comes around.
Where the row comes from
Most rollups get assembled by hand in the last week of the month, which is why most of them describe a month that ended before anybody read them. The Agent Scoreboard builds this row out of the scored calls as they land, and the section split comes off those same criteria rather than a second spreadsheet somebody keeps in parallel. The agent scoreboard is the dashboard version of the table above.
Any rollup inherits whatever the rubric underneath it is worth. Automate a weak scorecard and the result is a confident wrong number arriving faster, with a trend line attached to it. Machine scoring also holds up better on rule based criteria than on tone, which Aircall and StackAI both note, so the communication rows are where a supervisor's reading time still goes.
When the period does not hold enough calls
Everything above falls apart if the month contains four scored calls, and plenty of months do.
Say those four come back at 92, 92, 92, and 60. The average is 84. Pull four different calls from the same agent in the same month and you might land on 92, 90, 88, and 86, which averages 89. Same person, same thirty days, five points apart depending on which calls the supervisor happened to open. Most of the gaps a program acts on are narrower than that.
Call Centre Helper's sample size math puts a defensible read at roughly 65 evaluations per agent per month at 90% confidence for an agent taking 1,400 calls. SQM Group reports that about 60% of call centers evaluate five or more calls per agent per month. The distance between five and sixty five is a staffing problem wearing a methodology costume, and no amount of scorecard design closes it.
When the volume is not there, the sheet still has work to do:
- Widen the window. A quarter scored under one rubric version gives you three times the sample and one trend line you can defend, in place of three monthly ones that mostly measure which calls got opened.
- Report the average as a band. Fourteen calls supports "mid to high eighties" and does not support 88.0 against 85.6 as a two point improvement.
- Keep the fields that survive a thin sample. A critical fail count counts events rather than estimating a population, and the same goes for disputes and coaching sessions held.
- Hold off on ranking agents against each other until coverage is even across the floor. An agent with 22 scored calls and an agent with 4 do not belong on the same chart.
- Score everything inside a narrow slice rather than a thin spread across all of it. Every after hours call for one month, or every call on one account, gives you a population you can actually describe.
That last one is the move most programs skip, and it is the cheapest. A supervisor who scores all 60 of an agent's overnight calls knows something real about how that agent works the on call script at 2 a.m., which 14 calls scattered across a month will never tell them.
If a trend line matters more than a first data point, the pilot pulls six months of your existing recordings and scores them against your own rubric, at a fixed pilot price.
Whatever you decide the average is worth, put the coverage number in the column before it, where the agent and anybody from HR can see what the score was built on.
Frequently asked questions
What is a call center agent scorecard?
A rollup covering one agent across one period, usually a month. It holds volume, coverage, the average and its movement, the critical fail count, the section scores, and the coaching record, with no individual criteria on it. The per call scorecard is the instrument that produces the numbers the row summarizes.
How many calls should you score per agent per month?
About 60% of call centers evaluate five or more calls per agent per month (SQM Group), while Call Centre Helper's sample size math puts a defensible read closer to 65 evaluations at 90% confidence for an agent taking 1,400 calls. Your own number sits between those and moves with volume. Report the coverage percentage next to the average whatever you land on.
What metrics belong on an agent scorecard?
Calls handled, calls scored, coverage, the average with its change against the prior period, critical fails, disputes raised and resolved, coaching sessions held, the top deduction criterion, the section split, and a development focus with a date. Coverage leads because every other number on the row depends on it being big enough to mean something.
How do you track agent performance over time?
Match the rubric version across every period you compare, and log each change with the date it took effect. Two periods scored under different versions produce a trend line that measures the edit rather than the agent. A fresh baseline after a rubric change costs you one month of trend and saves you a coaching conversation built on nothing.
What is a good QA score for an agent?
SQM Group puts the industry average near 85%, treats 90 to 99% as good, and reports that only about 5% of agents reach 100. Those numbers compare cleanly only inside one rubric version, since a scorecard with six auto fails and a scorecard with one produce different results for identical work. Compare an agent against their own prior periods first.
Sources
- SQM Group, on benchmark QA scores and evaluation volume per agent. Third party, unverified by us.
- Call Centre Helper, on sample size for a defensible read at 90% confidence. Third party, unverified by us.
- Aircall and StackAI, on rule based criteria scoring more reliably than tone. Third party, unverified by us.
- No published threshold for acceptable calibration variance. Our own research finding, 2026-09-08.
- Agent Scoreboard and dashboards by agent. Trinsic feature sheet, verified internally.
- Six rubric sections and their point values. From CallSniper's published call center QA scorecard, with editorial judgment on the weights.
- The example row, the section figures, and the four call arithmetic. Hypothetical, built for this page.