Last updated: September 12, 2026
What a customer service scorecard measures
A customer service scorecard grades one support interaction against a fixed rubric covering whether the agent understood the issue, the accuracy of the answer, how the reply read, the quality of the case notes, and whether the promised follow up happened. The template below runs to 100 points and covers voice and written channels on one card.
This is a separate instrument from a call center scorecard, and the difference shows up first in the auto fail list. A compliance driven call voids on the disclosure, because an agent who skipped the recording notice created exposure that no amount of warm handling buys back. Support interactions void somewhere else. The two failures that kill a support score are giving the customer information that turns out to be wrong, and closing a case the customer did not think was closed.
That inversion runs through the weights. Script adherence carries real points on a compliance card, and here the opening is worth 8, with the 44 points that decide the score sitting in accuracy of the answer and resolution with follow through. Those two are what the customer actually experienced. How to build a call scoring rubric covers defending a weight after somebody disputes a score.
Queue shape is the other difference. A support lead scoring last week's work is looking at tickets, and one ticket can hold a phone call, two emails, and a chat session about the same broken thing. If everything arrives on the phone under a script, the voice heavy version of this rubric is the better starting point.
The 100 point customer service scorecard
Eighteen criteria across eight sections. Section totals sit in parentheses, and the channel column says where each criterion applies.
| Section | Criterion | Channel | Points |
|---|---|---|---|
| Opening and identification (8) | Opened with the approved greeting and gave a name the customer can ask for again | Voice | 4 |
| Opening and identification | Verified the customer's identity to the level policy requires before touching the account | Both | 4 |
| Understanding the issue (18) | Restated the issue in the customer's own words before proposing a fix | Both | 7 |
| Understanding the issue | Read the case history instead of asking the customer to repeat what they already sent | Both | 6 |
| Understanding the issue | Asked a clarifying question where the request could mean two things | Both | 5 |
| Accuracy of the answer (24) | Every factual statement matches current policy or documentation | Both | 12 |
| Accuracy of the answer | Quoted the right entitlement, timeline, or charge for this customer's plan | Both | 7 |
| Accuracy of the answer | Named the limit of what could be done rather than implying more was coming | Both | 5 |
| Resolution and follow through (20) | Resolved the issue on this contact, or set a next step with an owner and a date | Both | 8 |
| Resolution and follow through | Delivered the promised callback or update inside the window the customer was given | Both | 7 |
| Resolution and follow through | Confirmed with the customer that the issue was handled before closing the case | Both | 5 |
| Tone and clarity (12) | Held a steady tone and kept answering the question after the customer escalated | Both | 7 |
| Tone and clarity | Acknowledged the impact on the customer before moving to the fix | Both | 5 |
| Channel handling (8) | Used the right macro or template and edited it so it answers this customer in plain language | Written | 4 |
| Channel handling | Escalated to the queue that owns the issue rather than the nearest one | Both | 4 |
| Documentation (6) | Case notes carry the issue, the action taken, and the outcome in a form the next agent can act on | Both | 4 |
| Documentation | Case tagged and categorized so it comes back in a search | Both | 2 |
| The close (4) | Offered further help and told the customer how to reopen the case | Both | 4 |
| Total | 18 criteria | 100 |
Tone sits at 12, and it is the section we are least sure about. Both of those rows depend on knowing which voice belonged to the agent, and the section on what we can score gets into why that matters.
Auto fail items
Five failures zero the card regardless of the arithmetic. A support auto fail is almost always about what the customer walked away believing.
| Auto fail | Why it is absolute |
|---|---|
| Gave information the customer could act on that contradicts current policy | The customer acts on it, and the next contact starts from a worse position than the first |
| Released account or personal information to a requester whose identity was never established | Data exposure |
| Closed the case with the issue unresolved and no follow up scheduled | The customer has to start over to get back to a person |
| Committed the company to a refund, credit, or timeline nobody had authority to give | Somebody senior takes it back, in writing, from a customer who was already unhappy |
| Ended the contact while the customer was still asking the original question | The interaction produced nothing |
Download the scorecard
The rubric downloads as a CSV or a spreadsheet, with no form in front of either. Columns are section, criterion, points, auto fail, and channel, so you can filter the written rows out when you score a phone queue.
Scoring the same issue across two channels
Say a customer calls Monday about a duplicate charge, emails Tuesday with the statement attached, and opens a chat Thursday because nobody has written back. One problem, three interactions, two agents, and by Thursday that customer has explained the charge three separate times.
Two teams score that ticket two different ways, and both look defensible until you check what each one produces.
The common approach scores each touch independently against the card. Three interactions, three scores, each agent graded on their own work, which feels fair and is easy to administer. The Thursday chat agent can score 94 by being accurate, warm, and quick about a problem that should never have reached Thursday.
We would keep two numbers on that ticket instead. Each interaction gets scored against the 100 points above, and the ticket as a whole gets a short resolution read covering whether the issue was actually resolved, how many contacts it took, and whether the customer had to repeat themselves. That read takes about ninety seconds and is not scored out of anything.
Score every touch as an independent event and three things stop being visible. Repeat contact disappears into a set of good individual scores. The handoff goes ungraded, so nobody owns the moment the phone agent closed the call without a case note, which is what made the email agent start from zero. And the cross channel failure has nowhere to land except the 6 point line about reading case history, which is more weight than one line was built to carry.
Sampling makes this harder than it sounds. Manual QA gets eyes on 1 to 3% of calls (Verint, Gistly, ViciStack), and a mixed queue runs worse, because the written side rarely gets sampled at all. Say you run nine agents handling 3,100 contacts a month. Two scored interactions per agent per week is 72 interactions, a little over 2%, and none of those 72 will tell you the duplicate charge took three contacts.
So pair it with something cheaper. Once a month, pull ten tickets that took three or more contacts and read them end to end without scoring anything. That hour finds different problems than the sampling does. Then roll the scores up per agent for the monthly one to one.
What we can score and what your team still reads
CallSniper ingests recorded calls. Audio arrives after the call ends, gets transcribed and split by speaker, and gets scored against whatever card you built, with points on each criterion and a critical fail flag that zeroes the score the way the auto fail rows above do. Email, chat, and ticket text never enter the product, so every written row on this card gets read by a person on your team.
Saying that on a page written for a support lead with a mixed queue costs us traffic. It costs less than the second meeting where somebody asks which channels are covered and finds the answer narrower than the website implied.
| Rows an automated pass handles on recorded voice | Rows your team still reads |
|---|---|
| Greeting wording and whether the agent gave a name | Anything that happened in an email, a chat, or a ticket note |
| Identity verification questions asked before account access | Whether the answer given matches a policy that lives outside the transcript |
| Whether the issue was restated before a fix was proposed | Whether the case history was read before the customer repeated themselves |
| Filler words, slang, and terms on the do not say list | Whether the tone landed with a customer who was already angry |
| Whether a next step with a date was stated out loud | Whether the promised callback actually went out on time |
| Hold events, transfers, and where the call was escalated to | Whether the escalation reached the queue that owns the issue |
Rule based lines score reliably and tone scoring is measurably weaker, which Aircall and StackAI both report about automated QA in general. Every scored line comes back with a transcript timestamp and a per line confidence score, so a supervisor spends twenty minutes on the low confidence flags instead of relistening to the sample.
Both tone rows depend on knowing which voice is the agent, so run the rubric against your own recordings and check the speaker split before trusting those two numbers. Which channels we transcribe and score has the current list.
First contact resolution as a scored line
First contact resolution usually lives in a metrics dashboard next to average handle time, calculated from repeat contacts inside whatever window your team picked. That number gives you a rate and nothing about the interaction behind it.
On a scorecard it becomes a different instrument. The scored line asks whether the agent did the things that make a second contact unnecessary, since nobody knows at the time whether the customer is coming back. Three rows in the table above carry it and total 17 points, covering confirmation that the issue was handled before the case closed, a next step with an owner and a date where it could not close, and a case note the next agent can act on.
The dashboard rate then works as a check on the card rather than a competing number. When those three lines run high and the repeat contact rate refuses to move, one of them is written wrong or the product is generating the repeat contacts on its own. We are not going to hand you a target rate either, since published FCR figures move with how the window gets defined, so take the median of your own last quarter and hold the team to improvement against that.
When a high score and an unhappy customer land on the same ticket
Every support QA program hits this wall eventually. A card measures process adherence. The customer experiences an outcome. Those two track each other most of the time, and the tickets worth reading are the ones where they come apart.
An agent can score 96 here by verifying identity, restating the issue, quoting the right policy, and documenting all of it cleanly, and the customer still cancels, because the right policy was a no. That 96 is accurate. It also tells you nothing about why the account left.
What keeps the two from drifting is a monthly audit that runs the other direction. Pull every interaction that scored above 90 and still produced a complaint, a detractor survey, or a repeat contact inside a week, then read those without scoring anything. Each one comes back as a problem with the card or a problem with the policy. Card problems are a missing criterion or a weight that is wrong, and both are cheap to fix. The policy version is the valuable one, an agent following the rules correctly to an outcome nobody wanted, and a QA team is the least equipped group in the building to act on it, which is why that read has to reach whoever can change the rule.
Run the same pass in reverse on low scores with happy customers. Those usually surface a criterion measuring internal preference rather than anything the customer noticed.
The obvious response is to stop arguing about process and score the outcome directly. We talked ourselves out of that one. Outcome depends on the shipping delay, the policy, the bug in the product, and the customer's mood before they dialed, none of which the agent controls, and grading people on things outside their control ends a QA program faster than a bad rubric does. So the card keeps scoring process, and the audit above is what keeps the process honest.
Schedule a demo and we will score six months of your own recorded calls against this card, voice rows only, at a fixed pilot price.
Frequently asked questions
What should a customer service scorecard include?
Eight sections cover it. Opening and identification, understanding the issue, accuracy of the answer, resolution and follow through, tone and clarity, channel handling, documentation, and the close. Every criterion has to name something observable, a phrase said or a callback that went out inside the window, so two reviewers scoring the same interaction separately land in roughly the same place.
How do you score email and chat support?
Same card, different evidence. The written rows are macro and template use, the plain language check, case note quality, and tagging, and somebody on your team reads those out of the ticket itself. CallSniper ingests recorded calls, so it does not score email, chat, or ticket text.
What is the difference between a customer service scorecard and a call center scorecard?
Channel mix and weighting. A call center scorecard grades a call handled under a script, where disclosure and script adherence carry the auto fails. This card grades a support interaction that may have arrived in writing, and the weight sits on accuracy and resolution. What call center QA covers walks through the discipline behind both.
Can AI score customer service interactions?
On recorded voice, automated scoring is dependable on the rule based lines and weaker on tone and empathy (Aircall, StackAI). Written channels sit outside what CallSniper ingests at all. The setup that holds up is machine scoring across every recorded call, with a human pass over the subjective lines and the whole written side.
What is a good customer service QA score?
SQM Group puts the industry average near 85% and treats 90 to 99% as good, with about 5% of agents reaching 100. Those figures come out of voice contact centers, so they travel poorly to a queue that is mostly email. Compare a team against its own prior quarter under the same rubric version before comparing it to anybody else.
Sources
- SQM Group, on benchmark QA scores. Third party, unverified by us, and drawn from voice contact centers.
- Aircall and StackAI, on rule based criteria scoring more reliably than tone. Third party, unverified by us.
- Verint, Gistly, and ViciStack, on the share of calls manual QA reviews. Third party, unverified by us.
- CallSniper ingests recorded calls and does not score written channels. Trinsic feature sheet, verified internally.
- Transcript timestamps, per line confidence scores, scorecards with points per criterion, and critical fail flags. Trinsic feature sheet, verified internally.
- Eight sections and their point values. Editorial judgment for this template.
- The duplicate charge ticket and the nine agent sampling arithmetic. Hypothetical, built for this page.