Skip to main content

​
Agent scorecard: one page per agent

Quality → Agents → <agent> is the per-agent scorecard. It composes the six metrics a supervisor needs to judge one agent — QA composite score, schedule adherence, AHT, FCR, CSAT, and compliance flags — with the LLM-judged rubric breakdown underneath, plus a coaching card that turns the numbers into next steps. Every number is computed on read from the data you already have, with drill-down to the interactions behind it. This guide covers what the panels enumerate, who can see the page, how the rubric view composes with Quality → Evaluations, where each click-through lands, and how to run three common supervisory scenarios from it.

​
1. Scope — who the page is for and who sees it

The page is per-agent and supervisor-facing. Access is enforced on the server, not just hidden in the UI:
  • Owner, admin, and super-admin roles can open any agent’s scorecard.
  • Supervisors can open agents in queues they supervise; an out-of-scope agent returns a permission notice, not the data.
  • Agents can read their own card — and that self-view has its own dedicated surface. The agent-facing path is My scorecard (/me/scorecard), which renders the same content with self-framing, and the My evaluations tab under Quality → Evaluations covers the agent’s acknowledgment and appeal flow. This page deliberately does not try to be the agent’s home; it is the supervisor’s deep-link into one person’s record. Self-evaluations an agent submits are also excluded from every QA metric the page reports, so a self-assessment never inflates the QA average, the trend, or the flag count.
Reach the page from Quality → Agents, or drill directly from the Quality leaderboard when a rank needs explanation.

​
2. Metric panels — what each tile reads and where it is computed

The Performance scorecard card at the top carries six KPI tiles and a day selector (7 / 30 / 90 days, defaulting to 30). Every tile is a server-computed aggregate over the selected window; the UI does no math beyond formatting. A metric the tenant does not capture yet degrades to a dash rather than an error — the rest of the card still loads. Below the tiles, the QA score trend renders a daily bar for each day in the window, with adherence and handled-call context on hover, and the Recent evaluated interactions list shows the latest evaluations with their QA score, CSAT, flag state, and date — each row a link to the source call.

​
3. LLM-judged rubric breakdown — how it relates to Quality → Evaluations

Beneath the performance card, the rubric scorecard shows the LLM-judged quality-management view: an overall pass rate with the total-scored count and average confidence, a per-rubric bar list sorted weakest-first, the top failure modes, and a daily trend. Channel filtering (All channels / Inbox) narrows the rubric set; rubrics that have produced no scored conversations show as an explicit empty state, never as fabricated numbers. Each rubric row is a drill-down into the scorecard-form rubric behind it — the same rubrics a reviewer grades under Quality → Evaluations. The composition to understand: under Quality → Evaluations, an AI pre-score can grade a conversation against the form before a human reviewer confirms or overwrites it (the Auto-scored tab); the corrected total is what the ledger keeps. This page’s rubric view reads the resulting rubric-level outcomes — rubrics whose scores come from an AI pre-score that a reviewer never overrode carry the AI’s confidence with them, which is why the tile reports average confidence alongside the pass rate. Treat a low-confidence rubric row as a prompt to spot-check the underlying evaluations, not as a settled human judgment. The full grading and appeal lifecycle is in QA Evaluations.

​
4. Drill-down to source interactions

Every KPI is a summary; supervision happens on the interactions beneath it. The page names its transitions explicitly:
  • Recent evaluated interactions → each row links to the source call on Voice → Calls (/voice/calls?call=<id>), opened at that call so you can listen and re-read the transcript against the score.
  • Coaching card compliance flags → the flagged-for-review interactions link to the same Voice → Calls detail view.
  • Rubric rows and failure modes → continue to Quality → Evaluations filtered to the agent to read the graded conversations, then Quality → Recordings for the recordings those evaluations grade.
Work one low outlier per coaching session: open the evaluation, listen to the recording, and decide whether the score survives your own read. If it does not, that is a calibration signal for your reviewers or the Auto-QA threshold — not automatically an agent problem.

​
5. Comparison shards — window plus coaching baseline

The page has no absolute numbers without a comparison shard, and it uses two:
  • Window shards. The 7 / 30 / 90-day selector re-aggregates every tile server-side on each switch. Use 7d for shift-floor checks, 30d (the default) for the standard coaching view, and 90d — the maximum the server accepts — for review-conversation inputs where the sample is large enough to be fair. A 7-day spike is noise until the 30-day shard confirms it.
  • Baseline vs. peek. For coaching measurement specifically, the supervisor-notes flow compares a pre-coaching baseline window against a post-coaching window of equal length anchored at the coaching note — the before/after report that tells you whether the intervention moved response time, resolution rate, or CSAT. The convention is symmetric windows; the after-window is capped at now so an in-flight measurement never reads a future, empty span.
There is no cross-agent comparison on this page by design — ranking, team averages, and cohort views live on the Quality leaderboard. This page answers “how is this one agent doing,” not “where do they rank.”

​
6. API equivalents

The two reads behind the page are public API endpoints. Both require the inbox:read scope and apply the role rules from section 1 — an agent can read their own card; an out-of-scope supervisor gets a 403. Fetch the unified performance scorecard:
Fetch the LLM-judged rubric view — per-rubric pass rates, daily trend, and top failure modes:
Both are computed at read time — no cached snapshot; days and window accept up to 90 days and default to 30. Metric values degrade to null when the underlying feed has no rows for the window, so a tenant without WFM schedules or without surveys gets a partial card, never an error. The coaching-comparison read is GET /api/v1/inbox/agents/<agentId>/coaching-effectiveness?window_days=14, which returns the symmetric before/after KPI deltas around an intervention.

​
7. How supervisors should use it in practice

Three scenarios the page is built for:
  1. Flag an agent for investigation. Compliance tile above zero, or a QA score sitting below your floor on the 30-day shard? Open the coaching card’s flagged interactions, listen to each on Voice → Calls, and check whether the flags cluster — one bad day is a coaching note; a pattern across the window is a plan.
  2. Detect calibration drift. When the rubric view’s average confidence is high but your own re-grades keep disagreeing with the scores, the drift is in the reviewers or the Auto-QA threshold, not the agent. Use Quality → Evaluations → Calibration to re-align reviewers, and re-check the Auto-QA threshold if AI pre-scores are the source of the disagreement. A supervisor who skips this step ends up coaching agents for scoring-system errors.
  3. Kick off a coaching plan. The coaching card surfaces the weakest rubrics as explicit focus areas and shows the agent’s active plan if one exists. Author the plan under Voice → Coaching → Coaching plans, note the baseline date, and return after the plan closes: the coaching-effectiveness read gives you the symmetric before/after report on the same metrics this page tracks, so the plan’s outcome is measured, not asserted.

​
See also