Performance

Performance answers one question in several shapes: how is the floor doing, and who needs attention today. It is built around an idea the rest of the product does not assume anywhere else — an agent's work is not split by where it happened. Conversations run through Kilevel and recordings imported from a contact centre are the same person's output, so they are counted together and the mix stays visible.

Nothing here defines what "good" is. The metrics are whatever your own structured outputs and metric packs actually produced — discovered from the data rather than from the schemas, so a column is never offered that would be empty.

Where the numbers come from

Calls, chats and imported recordings are all counted. Which of them is included is a filter, not a setting, so you can always ask the narrower question.

The filter row

One row above the tabs decides what every view is looking at.

Changing view starts that view on its own settings rather than on the last one's. A window picked to read this morning's briefing should not silently become the window the scorecard matrix judges people over. If the view has a saved layout starred as the one it opens on, and that layout was saved with its filters, those are applied instead of the defaults.

Saved layouts

Any view can have named arrangements: which panels appear, in what order, under a name and an icon. Star one and that view opens on it.

A layout can optionally carry the filters as well — "Monday, remote pool, against the induction scorecard" — which turns it from an arrangement into a saved view. That is opt-in per layout, because a switch that silently moved somebody's date range would be doing something they did not ask for.

The views

Clicking any agent opens their detail modal: their periods, their criteria, their classifications and their conversations.

Scorecards

A scorecard turns a set of metrics into one score and a pass mark. Each criterion names a metric and carries:

• Weight — how much it counts.
• Better when — higher or lower. If a criterion scores a metric against its declared direction, the editor says so: a good conversation would be losing points.
• Scale — the range it is normalised against. If a metric's observed range changes, the editor flags the mismatch rather than quietly scoring everybody low.
• Target — the per-criterion goal. Something can miss a target and still pass overall, which is the point of weighting.
• Auto-fail — marks a criterion non-tradeable. It is reported as its own rate — "3 of 90 calls failed to verify identity" — rather than folded into the weighted score, because one catastrophic breach among ninety good calls would otherwise average away to nothing. That is also how real QA programmes report them.

An account with no scorecard is given one. On the first visit, a standard rubric is built from the metrics that account actually produces and selected. It is an ordinary scorecard: review it, edit it, rename it, archive it. Presets for common shapes are offered in the editor, each naming the packs it scores so you can see what it needs.

Review & calibration

A scorecard is only worth what its scores are worth, so this view measures agreement: how often the model and a human reviewer land within ten points of each other on the same conversation. Ten, because two competent reviewers routinely differ by that much.

Draw a random sample per agent — so the quietest person is reviewed as often as the busiest — score a handful by hand, and the view reports agreement, the typical gap, any consistent bias, and auto-fail mismatches, which are the most serious kind of disagreement. Around twenty completed reviews is enough for the figure to mean anything.

It also names the criteria the two sides disagree about most. A criterion people disagree about is usually worded ambiguously rather than scored wrongly, and rewriting it is normally the fix.

Alerts & rules

A rule is a question asked on your behalf every night, after the rollup: "did anyone's score drop by more than five points this week".

• Watch a metric, a scorecard score, an auto-fail rate or conversation volume.
• For each agent, or the account as a whole.
• Compare against a fixed number, against the previous period in points, or in percent.

Before saving, the editor previews how many agents would fire today, and says so when a threshold can never fire, always fires, or fires for more than half the population — which is usually a threshold set where the population already sits, and produces a daily list nobody reads. A rule that stays true for three weeks is one alert that stays open for three weeks, not twenty-one notifications.

Privacy and retention

In a workplace with a works council, per-person performance measurement is co-determined — in Germany under BetrVG §87(1) no. 6. These settings exist so the page can be run the way an agreement permits, instead of being switched off entirely while one is negotiated.

Outside "names shown" you also set a minimum number of people per combined figure and a minimum number of conversations per person (below it, the row is hidden rather than shown under a pseudonym), whether the rules apply to AI agents too (off by default — the co-determination right protects employees), and whether identities may be revealed for cause. Every reveal is logged with who did it and why; most agreements permit a defined cause rather than forbidding it, and the log is what makes that acceptable.

Retention permanently deletes daily per-person rollups older than the limit, every night. It cannot be undone or rebuilt.

Getting numbers onto this page