A revenue-intelligence platform sees the macro view: the whole call, the whole pipeline, what happened. What it does not carry is a graded, evidenced read on how well each skill was actually demonstrated. That is the micro view, and it is what Peer supplies — individual moments of demonstrated selling skill, graded against the universal sales-skills standard, each returned with the quote and timestamp that earned it.
The question we set out to answer is what happens when an AI assistant has both.
The finding: the same assistant, measurably better work
We ran the same AI assistant on the same questions twice. Once with the recorder's data alone. Once with Peer's graded skills observations served to it alongside. Judging was blinded — the judge did not know which answer came from which condition.
| Measure | Recorder only | With Peer |
|---|---|---|
| Blinded judge preference, head to head | 0 of 12 | 12 of 12 |
| Leadership decisions that changed | — | 10 of 12 |
| Fact-checked claim accuracy | 3, 4, 3 of 10 | 9, 9, 10 of 10 |
| Honest answers under 20 trap questions | 13 | 17 |
| Self-consistency across 8 runs | 86.7% on 3 competencies | 100% on 15 |
| Leadership work products, blinded preference | 0 of 9 | 9 of 9 |
The clarity measure is the one worth pausing on. Ask the assistant the same question eight times and, on recorder data alone, its answers held together on three competencies. With Peer's data underneath, they held together on fifteen. That is the difference between a tool you can quote in a promotion debate and one you cannot.
Why the two views need each other
Seven independent assessors then marked real calls blind — no AI grade in front of them, no sight of anyone else's marks — to establish what the micro view is actually worth and where it stops.
On a single moment, the engine grades about as well as a trained assessor. Judging the same moment, engine and human land within one rung of each other 95.5% of the time — marginally more often than two trained assessors agree with one another (93.5%). The micro view is sound.
But it is deliberately narrow. The engine grades one moment per competency in 99.1% of cases, and reaches a median of 15 of the 43 competencies in a call (mean 21). Across the study it returned a grade on 34.8% of the competency cells a human assessor graded — 541 of 1,554. In a replication on a second organisation's corpus, the share of the engine's emitted observations that carried a usable grade fell with framework level: 80.9% at foundation level, 56.8% at enhanced level and 11.4% at advanced level. On the hardest competencies, current AI grading produces no usable grade nine times in ten.
So neither view is sufficient alone. The macro picture knows what happened across the whole conversation but not how well each skill was demonstrated; the micro view grades that precisely, on the moments it reaches. The uplift at the top of this page is what happens when an assistant is given both.
How we know the data is sound
The same seven assessors are our check on our own grading. Judging whole calls, three independent assessors agreed with one another on the exact grade 62.5% of the time and within one rung 98.0% of the time. Against the engine's single graded moment for that competency, they agreed exactly 39.0% of the time and within one rung 84.1%.
That call-level gap is not the engine grading badly — on the moment itself it matches a human closely. It is the engine answering a different question: it grades one instance of a competency, where an assessor weighs every instance across the call. Whether an agentic layer that reasons across all the moments of a competency would close that gap is our next experiment. It has not been tested, and we are not claiming it.
We do not sell an agreement coefficient and no leader should buy one. It is how we audit ourselves, and it is why we can tell you precisely where the micro view is strong and where it stops.
Two different things, often confused
- Blind Human Observation Grading is internal quality assurance for the Peer platform. It is how we know the graded data we serve your stack has integrity. It is not a service.
- Blind Human Call Grading is a service, available on its own or alongside Peer: trained assessors mark whole calls blind, against your framework, with the evidence attached.
How the marking was run
Every assessor saw the transcript and the rubric. None saw the engine's grade, and none saw another assessor's marks; blindness was enforced at the read layer rather than asked for in an instruction. Names of people and companies were stripped before marking. Each assessor was tested against an independence check registered in advance — note originality, provenance, and declared exposure — before a single figure was computed, and marks from anyone failing it are excluded rather than weighted down. Eighteen real calls, all 43 competencies of the Rep's framework, three whole-call assessors marking eight calls each; separately 93 pinned single observations marked by all four observation-lane assessors.
What was not measured
Revenue, win rate, ramp time and deal outcomes were never joined to this data, and no claim is made about them. There is also no ground truth here: the agreement work measures whether two raters reach the same judgement, never which one is right.
What this means if you lead a revenue team
You already own the stack. The question is what it is working from. On these measures, giving that stack graded skills data changed what it told a leader — more of its claims survived a fact check, its answers held together across repeat questions, and a blinded judge preferred its work every time.
And the coverage figures tell you where the ceiling currently sits. When nine advanced competencies in ten come back empty, your Reps read as thin on advanced material across the board — and the coaching cadence you build off that reading is shaped by what the tool could see, not by what the Rep can do.
Limitations
This is a first study, not a population estimate: one organisation, one recorder, 18 calls, one framework family, and small samples throughout — twelve decisions, twenty probes, nine work products. The judge models used in the comparison come from the same model family as the assistant under test, so self-preference bias is not excluded. One measure moved the wrong way: with Peer's data the assistant invented fewer claims overall, six against eleven, but more in one category, four against one. On the agreement side, the human-versus-engine interval is wide — 0.123 to 0.474 — because it rests on eight call clusters, and the level-coverage figures come from a second organisation's corpus, normalised before joining. Nothing here shows our engine is more accurate than a human. It is not, and the study does not test that.