Everyone scores a 10 when the scorecard asks whether a step happened instead of how well it was done, when its top level is never described, or when the scores carry consequences that make managers generous.
Three causes of a ceiling
Most call scorecards start as a list of steps. "Did the rep get buy-in on the agenda?" is one of Gong's own examples. Almost every rep sets an agenda, so almost every rep gets the point. Stack eight questions like that and the team clusters at the top. The scorecard measures attendance at each stage of the call and cannot see quality.
The second cause is a scale where only the bottom is described. If level 2 says "asked about the problem" and level 5 says "excellent," a manager who heard a decent call has nothing to hold it against, so it gets a 5. The third cause is what the score is used for. Yoodli's calibration guide names one version of it: "A call that ended well tends to get a generous score on every item." Tie scores to reviews or pay and managers have a reason to be generous on every call.
The full picture of how scoring should work is in the AI call scoring guide.
How to tell a ceiling from a strong team
Check the spread before you rewrite anything. Take ten recent calls that your best manager would put in a clear order from weakest to strongest, and compare that order with the order of the scores. If the scorecard ranks them close to how the manager does, the team may be strong. If the scores tie, the scorecard has a ceiling.
Ranking is the right test. In our research program (August to October 2026, more than 5,000 tests and more than 250,000 scoring decisions, with blind marking by PeerLab's trained assessors), a scorer that never read the transcript and gave every call the same middling score landed an average error of 0.41. The best real model landed 0.43. Across 34 configurations, average error and ranking correlated at 0.02 (field guide). A scorecard that gives everyone a 10 can look accurate on averages while telling a manager nothing about who to coach.
The fixes that work, and two that do not
Telling an AI scorer to spread out does not work. In our research, instructing a model to "use the full range" made it use less of the range, and a rewrite into "observable checks" ranked calls at 0.62 rank agreement, against 0.74 for the production engine the research started from (field guide).
What worked was describing each level with the reasons trained assessors wrote for it: the assessors' short written notes on why a real call sat at that level. The observable-checks rewrite turned each level into a list of checks; this version kept the levels and added the assessors' own reasons, and it lifted the AI scorer's average fit with human scores (a different measure from rank agreement) from 0.62 to 0.72 in our tests, the largest single gain we measured.
Behavior-based anchors have a long record outside sales too: HR University notes that behaviorally anchored scales "can reduce rating errors, including leniency bias" compared with plain rating scales. Then separate the uses. Scores that coach a rep before the next buyer call should stay out of compensation, so managers score what they heard.
Rewriting the budget row (illustrative)
This rewrite is illustrative. The old criterion read "Did the rep ask about budget? Yes/No." Every rep asked, so every rep scored. The new version has levels. Level 1: budget never came up. Level 2: the rep asked whether budget exists. Level 3: the rep learned the amount and who controls it. Level 4: the rep tied the amount to the cost of the problem the buyer described. Level 5: the buyer stated the cost of doing nothing in their own numbers. Scored this way, the row stops clustering at the top, and the reps sitting at 2 give the manager an obvious coaching ask before their next pricing conversation.
- Did the rep ask about budget?
- Every rep asked, so every rep scored
- Scores cluster at the top
- Level 2: asked whether budget exists
- Level 3: learned the amount and who controls it
- Level 4: tied it to the cost of the problem
- Level 5: buyer states the cost of doing nothing
Before you rewrite the scorecard
Run the ten-call ranking check first, then rewrite only the rows that tie. The sales call scorecard is a starting template with every level described, Can AI score a sales call? shows the ranking test in full, and the AI call scoring guide covers the rest. For a second opinion on the calls everyone gave a 10, talk to us.
FAQ
Is a team with high call scores a good team?
High call scores alone do not show that a team is good. Compare the scorecard's ranking of ten calls with your best manager's ranking. If they agree, the scores carry information. If the scores tie where the manager sees clear differences, the scorecard has a ceiling and needs described levels.
Do AI scorers give inflated scores too?
AI scorers tend to deflate scores. In our research every newer model generation marked reps more harshly than the human assessors, and the gap grew with each release. That is a calibration problem in the other direction, and the same fix applies: check rankings against your managers and add their reasons to each level.
Should call scores be tied to compensation?
Keep coaching scores out of pay. Once a score affects a review, managers soften it and reps argue it, and the scorecard stops showing who needs help. Use outcome measures such as quota attainment for compensation and keep call scores for deciding what to coach before the next conversation.