AI Call QA: Build a Scorecard Your Agents Actually Trust
A call center QA scorecard template with 7 categories, what AI call review can score reliably, and how to use transcripts to coach agents, not police.
A call center QA scorecard is a fixed rubric that scores every reviewed call on the same categories, so agents and reviewers read the same number the same way. A workable one has seven categories, from greeting to compliance, each scored 0 to 10 and scaled to 100, with compliance treated as a gate rather than just another line.
Most templates were written for a supervisor with a headset and a few hours a month, listening to a small sample and hoping it is fair. Agents know it usually isn't. This guide gives you a scorecard to copy today, explains where AI call review helps and where it doesn't, and shows how to turn scores into weekly coaching. The template works with or without AI.
Why does manual call QA miss most of the story?
Manual QA misses most of the story because a reviewer can only hear a tiny, often unrepresentative sample of each agent's calls. Say a 10-agent team handles a few thousand calls a month and one QA lead reviews four calls per agent. That is 40 calls. Every conclusion about every agent rests on a tiny, often random slice of their work.
Three problems follow from that:
- Skewed samples. Reviewers tend to pull escalated or unusually long calls, which are real but not typical.
- Inconsistent scoring. Two reviewers listening to the same call often disagree on "showed empathy" or "handled the objection well." Without a tight rubric, the score reflects the reviewer as much as the agent.
- Slow feedback. By the time a call is discussed, the mistake has been repeated dozens of times.
What AI review changes
AI review changes three numbers. Coverage: it can score far more calls than a person can listen to. Consistency: it applies the same rubric to every call, the same way. Speed: a transcript, summary and score are ready shortly after the call ends, not weeks later.
The QA lead's job shifts from listening to checking the AI's work, finding patterns and coaching.
What should a call center QA scorecard include?
A call center QA scorecard should include seven observable categories: greeting, discovery, solution, objection handling, next steps, tone and compliance. That is few enough for agents to remember and specific enough that two reviewers land on the same score.
Score each category from 0 to 10, then add them up (a maximum of 70) and divide by 0.7 to get a score out of 100, so everyone reads the same kind of number.
- Greeting and opening. Did the agent identify themselves and the business, confirm who they were speaking with and set a reason for the call in the first 20 seconds?
- Discovery. Did they ask open questions about the customer's situation before pitching or solving? Did they listen to the answers and follow up on them?
- Solution and product knowledge. Was the recommendation or answer accurate, and did it match what the customer said they needed?
- Objection handling. When the customer pushed back on price, timing or trust, did the agent acknowledge the concern, ask a clarifying question and respond with something relevant?
- Next steps. Did the call end with a clear, agreed action: a booked appointment, a callback time, a sent quote, a resolved ticket?
- Tone and empathy. Was the agent courteous, patient and easy to follow? Did they match the customer's pace without talking over them?
- Compliance. Were the required disclosures given, consent confirmed where needed, and any stop or do-not-call request honored on the spot?
Treat compliance differently from the rest. A great call that skipped a required disclosure is not a great call. Many teams make compliance a gate: if a required item is missed, the call is flagged for review regardless of the total score.
What can AI call review judge, and what still needs a person?
AI call review is strongest on things that can be checked against the words in a transcript. It is weakest on things that depend on context the transcript does not contain.
| AI handles well | Needs a person |
|---|---|
| Whether a disclosure or recording notice was said | Whether the disclosure was clear enough for this customer |
| Whether discovery questions were asked | Whether they were the right questions for this account |
| Which objections the customer raised | Whether the agent's answer was the best one available |
| Overall customer sentiment | Sarcasm, cultural nuance, a regular customer's humor |
| Whether a stop or DNC request was made | Edge cases, disputes and anything with legal weight |
| A short summary and a suggested outcome | The final outcome on a deal or a complaint |
The practical rule: let AI score every call it can, and have a person spot-check a weekly sample of high scores, low scores and every compliance flag. If AI and reviewer scores drift apart, tighten the rubric wording.
In AI call review with QA scores on Phonify, a reviewed call gets a redacted transcript, a summary, the customer's sentiment, the objections raised and a QA score out of 100 across seven categories, each scored 0 to 10. It also suggests an outcome and the next follow-up, and flags the disclosure, recording notice, consent, DNC requests, abusive language and misleading claims. You can queue a recording for analysis from the report and search the AI call report by number, type or campaign. AI call review is part of Phonify Enterprise, and review of your agents' recorded calls is available on request.
How do you protect customer data in call transcripts?
Protect transcript data by redacting spoken card numbers and SSNs before storage, limiting access by role and deleting on a schedule. Transcripts are more searchable than recordings, which makes them more useful and more sensitive. Customers read out card numbers, Social Security numbers, account numbers and dates of birth. Once that text sits in a report, anyone with access can copy it.
Before you scale up transcript-based QA, set a few rules:
- Redact before anyone reads. Card numbers, SSNs and other spoken numbers should be removed before the transcript is stored, not after a reviewer has seen it. Phonify's AI call review redacts these from the transcript before it is stored.
- Limit access by role. Agents see their own calls, team leads see their team, and only QA, compliance and a few admins see everything.
- Set a retention period. Decide how long recordings and transcripts are kept and delete them on schedule.
How do you turn QA scores into agent coaching?
Turn scores into coaching with a short weekly one-to-one built on two real calls: one strong, one to improve. A score on its own changes nothing; a regular conversation does. Here is a routine that fits into a normal week.
The 15-minute weekly coaching session
- Pick two calls. One the agent handled well and one with a clear improvement area. Use scores and summaries to find them quickly instead of listening to everything.
- Start with the strong call. Play the best 60 seconds. Ask the agent what they did there. People repeat what they can name.
- Move to the improvement call. Focus on one category, not seven. "Let's look at how this call ended" beats a full list of faults.
- Agree one change. Write it down in the agent's words. For example: "Before hanging up, I'll confirm the callback day and time."
- Check it next week. Pull a call where the situation came up again and see whether the change stuck.
Coaching, not policing
Agents trust a scorecard when nothing about it is a surprise. Share the full rubric, including what "great" sounds like for each category. Let agents read their own transcripts and scores. Invite them to dispute a score when they think the AI or the reviewer got it wrong, and change the score when they are right.
Avoid public leaderboards of QA scores; they push agents to game the rubric. Share team-level trends in meetings and keep individual scores for one-to-one coaching.
If an AI voice agent qualifies a lead and passes it to a rep, score the handoff too. Our guide to the AI-to-human handoff covers what a good transfer sounds like.
Your call center QA scorecard template
The template below lists the seven categories, what a 9 to 10 sounds like and the usual reasons for a low score. Copy it into a spreadsheet: one row per category, a column for the 0 to 10 score and a column for reviewer notes. Adjust the wording to your call types, but keep each line observable in a recording.
| Category | What "great" (9 to 10) looks like | Common reasons for a low score |
|---|---|---|
| Greeting and opening | Name, business and reason for the call within 20 seconds; confirms the right person | Rushed or missing introduction; no reason given |
| Discovery | Two or more open questions; follows up on the answers before recommending | Pitching before asking; yes/no questions only |
| Solution and product knowledge | Accurate answer that matches the stated need; admits and follows up on unknowns | Wrong information; generic answer that ignores the customer's situation |
| Objection handling | Acknowledges, asks a clarifying question, answers the real concern | Arguing, ignoring the objection or giving up at the first no |
| Next steps | Specific agreed action with a date or time; confirms contact details | "I'll follow up" with no time; call ends without a decision |
| Tone and empathy | Calm, clear, patient; lets the customer finish; matches their pace | Talking over the customer; impatience; jargon |
| Compliance | Every required item given and every request honored | Any missed disclosure or ignored stop request (flag the call) |
Compliance checklist for every reviewed call
- Required disclosure given (including an AI disclosure, where an AI agent spoke)
- Recording notice given where your state rules require it
- Consent confirmed for the type of call or message that followed
- Any do-not-call or stop request honored and logged during the call
- No abusive language from the agent
- No misleading claims about price, terms or results
- No sensitive numbers read back or written into notes
Total the seven scores, scale to 100, and treat any unchecked compliance item as a flag for a person to review. If you are also rethinking how calls reach your agents in the first place, our comparison of power, predictive and preview dialers is a good next read.
Frequently asked questions
How many categories should a call QA scorecard have?
Five to eight categories is a practical range, and seven works well for most teams: greeting, discovery, solution, objection handling, next steps, tone and compliance. Fewer categories hide useful detail, while more make the rubric hard for agents to remember and for two reviewers to score the same way. Keep every category observable in a recording or transcript.
How do you calculate a QA score out of 100?
Add the category scores and scale the total to 100. With seven categories scored 0 to 10, the maximum is 70, so divide the total by 0.7. A call scoring 56 out of 70 becomes 80 out of 100. Report compliance separately as pass or flag, so a strong total never hides a missed disclosure.
How many calls should QA review per agent each month?
There is no fixed rule, but more is better because a few calls per agent rarely represent their real work. When AI scores most calls, a supervisor can spot-check a weekly sample instead: a few high scores, a few low scores and every compliance flag. That keeps human review focused where judgment matters most.
Can AI replace human call quality monitoring?
No. AI call review is reliable for checks that depend on the words in a transcript, such as whether a disclosure was said or which objections came up. It is weaker on tone, sarcasm, account context and anything with legal weight. Use AI for coverage and consistency, and keep a person responsible for spot-checks, disputes and final outcomes.
Should agents see their own QA scores?
Yes. Agents trust a scorecard when nothing about it is a surprise, so share the full rubric, let them read their own transcripts and scores, and let them dispute a score they think is wrong. Keep individual scores for one-to-one coaching and share only team-level trends in meetings, because public leaderboards push agents to game the rubric.
