Agent Performance Scorecard

🧮 KPI Calculation Methodology

How each metric is calculated, data sources, and formulas

📞 Contact Metrics

✅ Long Handled

Definition: Contacts with meaningful customer interaction time

Formula:


IF Channel = Voice AND Talk Time > 50 seconds THEN Long Handled
IF Channel = Chat/Email AND (Active Talk + Hold + Conference) > 50 seconds THEN Long Handled
IF Channel = Store Line AND Talk Time > 40 seconds THEN Long Handled

Data Source: call_detail (talk_sec, active_talk_sec, hold_sec, conference_sec)

Purpose: Counts contacts where agent provided substantive assistance

⚡ Short Handled

Definition: Brief contacts (quick questions, transfers, disconnects)

Formula:


IF Channel = Voice AND Talk Time ≤ 50 seconds THEN Short Handled
IF Channel = Chat/Email AND (Active Talk + Hold + Conference) ≤ 50 seconds THEN Short Handled
IF Channel = Store Line AND Talk Time ≤ 40 seconds THEN Short Handled

Data Source: call_detail

Quality Gate (hybrid): Hard exclude when short-handled > vertical p95; soft “review” flag when > vertical p90 (still scored). Replaces the old flat 80% rule — respects that Fraud & Risk legitimately runs hotter than Digital Care.

📊 Total Handled

Definition: All contacts handled by agent

Formula:

Total Handled = Long Handled + Short Handled

Quality Gate: Minimum 100 contacts required for scoring

⏱️ Efficiency Metrics

Average Handle Time (AHT)

Definition: Time attributable to the agent per contact. On chat/email this is NOT the customer-experienced session duration — read the callout below.

Important: two flavors of AHT, one column

Row type	Formula	What it measures
Voice	(TALK + ACW) / handled	Session AHT — real per-call duration (chat and customer experience align here)
Chat / Email attrib	(ACTIVE_TALK + HLD + CONF + ACW) / handled	Attributable AHT — agent-effort per contact. Discounts idle time when agent was handling a parallel chat.

Why: On chat, talk_sec = full session span (chat-window open-to-close). When an agent runs 2 chats in parallel, those spans overlap — summing raw TALK double-counts wall-clock time. Using ACTIVE_TALK gives real attributable effort.

Real example (Asfand Ali, Mar 2026): Session AHT = 807s / Attributable AHT = 548s. Ratio 1.47 matches his observed avg concurrency of 1.51 chats. His scorecard AHT is 548s — fair for peer comparison because it doesn't punish him for a multi-chat policy he doesn't control.

When to use which:

This scorecard → Attributable AHT. Fair peer comparison, correct for capacity planning, apples-to-apples across voice + chat.
Customer experience / "how long is one chat" -> Session AHT (naive TALK+ACW). Not shown here — run against call_detail directly.
Vendor benchmarks (vendor WFM defaults) → usually Session AHT. Expect our numbers to look LOWER than vendor dashboards for chat channels — that's the concurrency discount, not a bug.

Formula (channel-aware branch):


IF Channel IN ('Chat','Chat Spanish','Chat Escalation','Email','Email Escalation'):
  AHT = (ACTIVE_TALK + HLD + CONF + ACW) / handled     -- attributable effort

ELSE (Voice / Voice Escalation / Outbound / Offline):
  AHT = (TALK + ACW) / handled                        -- session duration

Data Source: call_detail (talk_sec, active_talk_sec, hold_sec, conference_sec, acw_sec) joined to queue_lookup for channel.

Skill-Based Benchmarks:


Personalized AHT Benchmark = Weighted average of skill benchmarks
  - Chat agents: ~550s target (attributable)
  - Voice agents: ~517s target (session)
  - Escalations: ~1200s+ target
  - Multi-skill: Weighted by contact volume per skill

Scoring: 100 points if AHT ≤ P25 benchmark, 0 points if AHT ≥ P75 benchmark, linear between.

🚀 Contacts Per Hour (CPH)

Definition: Productivity rate - contacts handled per login hour

Formula:

CPH = Total Handled / Login Hours

Data Sources:

Contacts: call_detail
Login Hours: agent_state_events (status_duration_sec / 3600)

Benchmark: Target ≥ 1.5 CPH, Minimum 0.8 CPH

⏸️ Hold Time %

Definition: Percentage of handle time customer was on hold

Formula:

Hold Time % = (Total Hold Time / Total Handle Time) × 100

Data Source: call_detail (hold_sec)

Benchmark: Target ≤ 3.3%, Maximum 18.4% (lower is better)

❌ Refused Contacts

Definition: Number of contacts agent declined to accept

Formula:


Refused = COUNT(contacts WHERE refuse_reason IS NOT NULL AND refuse_reason ≠ 'none')

Data Source: call_detail (refuse_reason)

Benchmark: Target ≤ 2 refusals, Maximum 6 refusals (lower is better)

[WFM] Workforce Management Metrics

v2 note (2026-07-01): Occupancy %, Productive %, and Utilization % are shown for visibility but NO LONGER FACTOR INTO THE GRADE. Leadership feedback: these are staffing / scheduling outcomes agents cannot directly control. They have been replaced in the composite by Shift Conformance and Break Conformance (see below), which measure behavior agents CAN influence (showing up on time, staying on shift, not overrunning breaks).

Occupancy % Context only — not scored

Definition: Productive time vs. productive + available time

Formula (Formula B):

Occupancy % = (Production Hours / (Production Hours + Available Hours)) × 100

Data Source: agent_state_events

Production: ACD state codes marked as 'Productive' (handling contacts, ACW)
Available: Agent logged in and ready for contacts

Code Mapping: activity_code_map (code_type = 'Productive')

Benchmark: Target ≥ 83.3%, Minimum 61%

Productive % Context only — not scored

Definition: Productive time vs. total login time

Formula:

Productive % = (Production Hours / Login Hours) × 100

Data Source: agent_state_events

Benchmark: Target ≥ 65.4%, Minimum 40.6%

Utilization % Context only — not scored

Definition: Available + productive time vs. total login time

Formula:

Utilization % = ((Production Hours + Available Hours) / Login Hours) × 100

Data Source: agent_state_events

Benchmark: Target ≥ 81.4%, Minimum 66.8%

📉 Total Shrinkage %

Definition: Non-productive scheduled time (meetings, breaks, training, leave)

Formula:


Total Shrinkage % = ((In-Office Shrinkage + Out-of-Office Shrinkage) / Scheduled Hours) × 100

In-Office Shrinkage = Huddles, coaching, meetings, training, team activities
Out-of-Office Shrinkage = Leave, PTO, sick time, absences

Data Sources:

Schedule: agent_schedule (source = WFM scheduler)
Activity Mapping: activity_code_map (scheduler activity codes)
Account Linking: agent_roster (linked_agent_ids)

✨ Multi-Account ID Support: Schedules searched across all linked account IDs using linked_agent_ids

Excluded Activities: Lunch, breaks, holidays, VTO (not counted as shrinkage)

Benchmark: Target ≤ 16.2%, Maximum 27.6% (lower is better)

Coverage: 100.0% of Internal agents (External agents excluded - no schedules)

Shift Conformance % v2 NEW · Weight 10% Internal only

Definition: How much of the scheduled shift the agent actually logged in for. Capped at 100% (logging in EXTRA doesn't earn bonus points; it's often overtime or unpaid time not the model wants to reward).

Formula:

Shift Conformance % = min(1.0, Login Hours / Scheduled Hours) × 100

Data Sources:

Login Hours: agent_state_events (all ACD aux time)
Scheduled Hours: agent_schedule (source = WFM scheduler)

Why it’s here: Replaces Utilization % as the agent-side accountability metric. Utilization varied wildly with staffing decisions; conformance is purely about the agent's own attendance. Standard call-center “schedule conformance” concept (per Reynolds, Call Center Staffing).

Vendor handling: External agents have scheduled_hrs = NULL (not on our WFM scheduler) → sconf is excluded and the composite renormalizes over remaining KPIs.

Benchmark: Target ≥ 95%, Minimum 70% (higher is better).

Coverage: 100.0% of scored Internal agents in current dataset.

Break Conformance % v2 NEW · Weight 10% Internal only

Definition: How well the agent stayed inside their scheduled break/lunch allowance. Zero overrun = 100%; a 30-minute overrun on 60-min allowance = 50%; overrun ≥ allowance = 0%.

Formula:

Break Conformance % = max(0, 1 − (Actual Break/Lunch min − Allowed min) / Allowed min) × 100

Included aux codes (numerator): Break, Extra Break, Training Break, Overtime Break, Lunch.

Scheduled allowance codes (denominator): Break, Training Break, Overtime Break, Lunch, Training Lunch, Overtime Lunch (from agent_schedule).

Personal time is EXCLUDED from this metric because it has no scheduled allowance — counting it in the numerator would guarantee a 0%. Personal / RTA Personal is surfaced separately as a diagnostic column.

Data hygiene: Agents with <120 min of scheduled Break+Lunch per month are excluded to avoid divide-by-tiny-number spikes. Three shared/kiosk accounts (25939388, 34457731, 20906022) with chronic unclosed-Break patterns are flagged in the agent table but not removed from data.

Distribution: 54% at 0% floor (blew past allowance), 25.8% useful middle, 16.4% at 100% ceiling. Bimodal by design — break overruns are typically all-or-nothing.

Vendor handling: External agents' ACD state doesn't stream into agent_state_events at the fidelity needed → bconf is excluded and composite renormalizes.

Benchmark: Target ≥ 95%, Minimum 50% (soft floor because of bimodal distribution).

Coverage: 89.3% of scored Internal agents in current dataset (5,708 of 6,392 with break data).

Personal Time (min/month) Diagnostic only — not scored

Definition: Total minutes in Personal or RTA Personal aux state per month. Surfaced as a leader-visible diagnostic to spot outliers, but never contributes to the grade.

Why not scored: Personal has no schedule pair (agent-initiated), so any denominator we picked would be arbitrary. Instead we show the raw minutes and let leaders decide whether a given number is concerning for their team.

Reference distribution (Feb–Aug 2026): median 14 min/mo, p90 209 min, p95 327 min. Outliers >500 min/mo are worth a conversation.

⭐ Customer Satisfaction

CSAT % (Customer Satisfaction)

Definition: Percentage of customers who rated their experience 4 or 5 stars

Formula:

CSAT % = COUNTIF(overall_rating IN (4,5)) / COUNTIF(overall_rating IS NOT NULL) × 100

Data Sources:

interactions (overall_rating, created_at, ref_id, survey_agent_login)
survey_responses (ref_id, agent_login) — resolves the agent who handled the surveyed interaction
agent_roster (login_id → agent_id) — maps login to scorecard agent ID

Join: interactions.ref_id = survey_responses.ref_id, then login (with COALESCE(B.agent_login, A.survey_agent_login) + UPPER/TRIM) joined to agent_roster.login_id.

Pipeline change (2026-06-25): Replaced the old digital_contacts + store_survey_csat + call_detail join. The previous pipeline had two defects: (1) call_detail segment-grain fan-out multiplied each survey by ~11×, and (2) the INNER JOIN silently dropped ~13.6K digital/chat/email/stores agents. New pipeline: response volume 41M→3.6M (realistic, ~30K/day), agent coverage 11.9K→25.6K, positive_rate unchanged at ~86.6%.

Minimum Requirement: Must have at least 10 CSAT responses to be scored on CSAT.

Handling below threshold: CSAT is excluded from the weighted score (the KPI sits out, same pattern as Call Quality for non-voice agents). It is not a penalty — the remaining KPIs reweight to 100%. (Was: <5 responses = 0 points, which unfairly tanked specialists.)

Benchmark: Target ≥ 95.4%, Minimum 82.1%

🎙️ Call Quality Score VOICE ONLY

Definition: Average AI/QA-derived transcript quality score for voice contacts. 0–100 scale, higher is better.

Formula:

Call Quality = AVG(transcript_quality_score) over the agent's voice transcripts in the month

Data Source:

call_transcripts (column transcript_quality_score, partitioned on transcript_date)

Join: call_detail.call_id = call_transcripts.call_id

⚠️ Voice Only: Non-voice agents (chat / email / offline) skip this KPI entirely — their other weights re-normalize so they're never penalized for missing voice data.

Minimum Requirement: ≥ 5 scored transcripts in the month to earn points (anti-noise gate).

Benchmark: Target ≥ 75, Minimum 50 (linear in between).

Coverage: ~100.0% of voice agents in the current dataset.

🔄 Data Freshness & Period

Reporting Period: Feb 1 - Aug 31, 2026 (212 days, Q1+Apr+May+Jun+Jul+Aug 2026)

Monthly Dimension: Each agent gets one row per month (enables trend analysis)

Agent Deduplication: Latest record per agent_id from agent_roster (updated_at DESC)

Vertical Filter: Order Support, Fraud & Risk, Store Helpdesk, Delivery Partner Support, Marketplace & Specialty, Digital Care

Query Processing: ~170 GB of data analyzed per refresh

Technical Questions? All data in this demo build is synthetic.

📊 Scoring Methodology & Grading System

How we calculate your performance score (0-100) and letter grade

Overall Score Formula (v2)

Score = Weighted Average of up to 10 SCORED KPIs

Each KPI is scored 0–100, then weighted by importance. Voice-only KPIs (Call Quality) are skipped for non-voice agents and Internal-only KPIs (Shrinkage, Shift Conformance, Break Conformance) are skipped for vendors — the remaining weights re-normalize automatically. Occupancy, Productive %, and Utilization % are shown as context metrics but do NOT factor into the grade (removed 2026-07-01 per leadership guidance).

KPI Weights (v2 model — Total = 100%)

Changed 2026-07-01: Occupancy, Productive %, and Utilization % were removed from scoring per leadership guidance (not agent-controllable). Their combined 25% weight was reallocated to new Shift Conformance and Break Conformance KPIs. Updated 2026-08-04: Transfer % added as a scored KPI (5%), funded by trimming Shift Conformance to 10%. The old context metrics are still visible in the table, marked as context only.

Category	KPI	Weight	Description
Customer Experience (45%)	CSAT %	25%	Customer satisfaction (4/5 star ratings). Most important metric!
	AHT (attrib for chat/email)	10%	Average Handle Time. Skill-based benchmarks! Lower = better efficiency. Chat/Email rows use ATTRIBUTABLE AHT (agent-effort per contact, concurrency-discounted) — not the customer-experienced session duration. Voice uses standard session AHT. See definitions for full explanation.
	Hold Time %	5%	% of talk time on hold. Lower = better customer experience.
	Refused	5%	Number of refused contacts. Lower = better acceptance rate.
Productivity (20%)	Contacts/Hour	12%	Contacts handled per login hour. Higher = more productive.
Productivity (20%)	Call Quality VOICE ONLY	8%	Avg `transcript_quality_score` from voice transcripts (0–100). Skipped for non-voice agents — their other weights re-normalize automatically.
Adherence & Conformance (30%) Internal agents only Vendors renormalize	Shift Conformance % v2 NEW	10%	min(1, login_hrs / scheduled_hrs). Did the agent show up for their shift? (trimmed 15→10 to fund Transfer%)
	Break Conformance % v2 NEW	10%	1 − (break_overrun / break_allowance). Did the agent stay inside their scheduled break window? Personal time excluded.
	Shrinkage %	10%	Non-productive scheduled time (meetings, coaching, leave). Lower = better.
Reliability (5%)	Transfer % new	5%	Share of handled contacts transferred out (transfer-events feed, distinct transferred ÷ distinct handled). Lower = better. Benchmark recalibrated on the internal population.
Context Only (0% — not scored)	Occupancy % removed	—	Still calculated and shown, but no longer scored. Reflects staffing decisions more than agent behavior.
	Productive % removed	—	Still calculated and shown, but no longer scored. Dominated by channel routing patterns.
	Utilization % removed	—	Replaced by Shift Conformance which measures the same intent (agent availability vs. schedule) but scoped to what the agent controls.

⚖️ Weight Source by Department — Ops-tuned vs Default (v2)

KPI weights are matched per Vertical → Sub-LOB → Department → Channel against the operations weight config. Overall 41.1% of scored agent-months (2,626 of 6,392) use Ops-tuned weights; the remaining 3,766 inherit Default (v2) weights.

Legend: All Ops-tuned Mixed All Default

Inventory Support

100%620 ops · 0 defAll Ops-tuned

Chargebacks

100%555 ops · 0 defAll Ops-tuned

Gift Cards

100%540 ops · 0 defAll Ops-tuned

Driver Appeals

100%525 ops · 0 defAll Ops-tuned

Photo Center

0%0 ops · 483 defAll Default

Order Status

0%0 ops · 482 defAll Default

Store Systems

0%0 ops · 395 defAll Default

Billing

0%0 ops · 374 defAll Default

Driver Payouts

0%0 ops · 324 defAll Default

Pharmacy Orders

100%293 ops · 0 defAll Ops-tuned

Facilities Requests

0%0 ops · 290 defAll Default

App & Web Support

0%0 ops · 288 defAll Default

Driver Onboarding

0%0 ops · 267 defAll Default

Seller Onboarding

0%0 ops · 257 defAll Default

Seller Payments

0%0 ops · 199 defAll Default

Returns & Refunds

0%0 ops · 148 defAll Default

Delivery Issues

0%0 ops · 143 defAll Default

Trust & Safety

0%0 ops · 116 defAll Default

Account Review

100%93 ops · 0 defAll Ops-tuned

🎯 NEW: Skill-Based AHT Benchmarking

AHT is now scored fairly based on YOUR actual work!

Each agent gets a personalized AHT benchmark based on their skill mix
Chat agents (avg 550s) vs. Voice agents (avg 517s) vs. Escalations (avg 1200s+) = different expectations
Multi-skill agents get weighted average (e.g., 60% Chat + 40% Voice)
You're compared to peers doing the SAME work - no more apples to oranges!
93% coverage - personalized benchmarks for most agents

🔬 How p25 / p75 Are Actually Calculated WORKED EXAMPLE

Personalized AHT benchmarks aren't magic — they're standard statistical percentiles computed from your peer cohort's actual performance. Here's the math end-to-end.

STEP 1 · WHAT P25 AND P75 ACTUALLY MEAN

Yes — the first step is sorting all the values from low (fast) to high (slow).

p25 = the value at the 25%-from-bottom mark — "only the top 25% of agents are faster than this"
p75 = the value at the 75%-from-bottom mark — "only the bottom 25% are slower than this"

STEP 2 · WORKED EXAMPLE — 11 CHAT AGENTS ON ONE SKILL

Sort their AHT values low → high (seconds):

Position: 1 2 3 4 5 6 7 8 9 10 11
AHT (s): 380 · 420 · 440 · 460 · 480 · 510 · 530 · 560 · 590 · 680 · 720

Position formula for n values:

position = (percentile ÷ 100) × (n − 1) + 1

p25 position = 0.25 × 10 + 1 = 3.5

Interpolate between #3 (440) and #4 (460):

p25 = 440 + 0.5 × (460 − 440) = 450s

p75 position = 0.75 × 10 + 1 = 8.5

Interpolate between #8 (560) and #9 (590):

p75 = 560 + 0.5 × (590 − 560) = 575s

STEP 3 · CONVERT MAYA'S ACTUAL AHT TO A 0–100 SCORE

Maya's actual AHT = 510s (position #6 in the sorted list). Apply the linear interpolation:

score = 100 − ((510 − 450) ÷ (575 − 450)) × 100
      = 100 − (60 ÷ 125) × 100
      = 100 − 48
      = 52

✓ Capped at 100 — anyone at-or-below p25 (450s) gets a perfect AHT score

✓ Floored at 0 — anyone at-or-above p75 (575s) gets zero

✓ Linear between — every second of improvement directly raises the score

STEP 4 · MULTI-SKILL AGENTS — VOLUME-WEIGHTED BLEND

Most agents handle multiple skills. Their personal p25/p75 is a contact-volume-weighted blend across every skill they touched:

Maya's skill mix	Contacts	% Volume	Skill p25	Skill p75
Chat — Returns	850	68%	450	600
Chat — General	300	24%	420	560
Voice — Outbound	100	8%	240	360
Maya's blended benchmark	1,250	100%	~426s	~563s

blended_p25 = (450 × 0.68) + (420 × 0.24) + (240 × 0.08)
= 306 + 100.8 + 19.2 = ~426s

Computed at scale via APPROX_QUANTILES(handle_time_sec, 100)[OFFSET(25)] — internally sorts and slices into 100 buckets; the 25th and 75th cut-points are returned.

⚖️ How the 9 Static Benchmarks Were Calibrated CSAT EXAMPLE

AHT gets a per-agent personalized benchmark (above). The other 9 KPIs use global, fixed benchmarks derived from the same p25/p75 percentile math — just computed once across the qualified population. Here's exactly how the CSAT target of 95.4% was derived.

STEP 1 · LOAD THE CALIBRATION DATASET

Read scorecard_rows.csv — one row per agent per month. Today: 8,809 rows.

STEP 2 · APPLY THE TWO QUALITY GATES

Drop rows with total_handled < 100 (statistical noise)
Drop rows where short_handled / total_handled > per-vertical p95 (gaming or mis-routing) and flag p90–p95 for supervisor review (still scored)

Result: 6,392 qualified agent-month rows

STEP 3 · EXTRACT CSAT VALUES, DROP NULLS AND ZEROS

values = [r['csat_pct'] for r in qualified_agents if r['csat_pct'] is not None and r['csat_pct'] > 0]

Result: 5,946 valid CSAT values · ⚠️ Zeros are treated as "no real data" and excluded.

STEP 4 · COMPUTE QUARTILES (PYTHON STANDARD LIBRARY)

import statistics
p25 = statistics.quantiles(values, n=4)[0] # 25th percentile
p75 = statistics.quantiles(values, n=4)[2] # 75th percentile

Sorts the 5,946 values low → high, returns the 3 cut-points between 4 equal-size groups. Uses the "exclusive" method (linear interpolation) — same math as Excel's PERCENTILE.EXC and numpy's default percentile.

STEP 5 · MAP PERCENTILES → TARGET / MIN BY DIRECTION

Higher-is-better (CSAT, CPH, Occ, Prdp, Util, CQS)

target = p75 (top quartile bar)
min = p25 (score 0 below this)

Lower-is-better (AHT, Hold%, Refused, Shrinkage)

target = p25 (top quartile bar)
max = p75 (score 0 above this)

📊 LIVE CSAT CALIBRATION OUTPUT (synthetic demo data)

Agents with data: 5,946
Average:         78.2%
Floor: 64.4%   ← becomes "min" (score 0 here)
Median (50th):   78.2%
Target: 85.4%   ← becomes "target" (score 100 here)
Range:           24.8% – 100.0%

CSAT scoring formula becomes: score = ((csat − 64.4) ÷ (85.4 − 64.4)) × 100 [capped 0–100]

📜 ALL STATIC BENCHMARKS — PROVENANCE TABLE

KPI	Target	Floor/Ceiling	Agents in pool	Direction
CSAT %	85.4	64.4	5,946	Higher = better
AHT (sec)	332.2	503.7	6,392	Lower = better
Hold %	0.6	5.7	6,392	Lower = better
Refused	0.0	2.0	6,392	Lower = better
CPH	10.0	7.3	6,392	Higher = better
Call Quality	74.2	57.1	4,307	Higher = better
Transfer %	1.3	5.6	6,392	Lower = better
Shrinkage %	10.6	17.5	6,392	Lower = better
Shift Conformance %	100.0	90.8	6,392	Higher = better
Break Conformance %	100.0	77.8	5,708	Higher = better
Occupancy % CONTEXT	90.0	78.2	6,392	Higher = better (not scored)
Productive % CONTEXT	77.5	62.0	6,392	Higher = better (not scored)
Utilization % CONTEXT	87.7	71.5	6,392	Higher = better (not scored)

✨ Re-derive any time: the benchmark calibration job

🔍 Stability: Benchmarks are fixed at calibration. If the population improves, the bar stays the same (more agents score 100). If the population regresses, the bar stays the same (more agents score 0). Calibrating quarterly is the recommended discipline.

📝 CPH and Call Quality are manually set — they aren't yet wired into the calibration script.

🧪 Benchmarking Methodology — Three Lenses, One Score

Not every KPI is measured the same way. We use three different benchmark philosophies depending on the metric — and there's an alternative scoring lens (Department-Relative) that you can compare side-by-side in the agent table.

🌍

Lens 1 · Global (org-wide)

Used by: CSAT, Hold %, Refused, CPH, Occ, Productive, Util, Shrinkage, Call Quality

One target / floor for the whole org. A Sales agent and a Trust & Safety agent are judged against the same CSAT bar of 85.4%.

Pro: simple, comparable. Con: ignores work-type differences.

🎯

Lens 2 · Personalized (skill-mix)

Used by: AHT only

Each agent gets a unique p25/p75 benchmark built from a contact-volume-weighted blend of their actual skill mix (e.g. 60% Chat + 40% Voice).

Pro: fair to specialists. Con: opaque — every agent has a different bar.

NEW

🏢

Lens 3 · Department-Relative

Applied to: All 8 static-benchmark KPIs (AHT keeps personalization)

Each agent is scored against the p25/p75 of their own department. Falls back to Global if the dept has <10 agents with that KPI.

Pro: surfaces top performers WITHIN a department. Con: a weak dept makes mediocre look like A.

📐 How Department-Relative scoring is computed

For each (department, KPI) combination, gather every agent-month's raw value.
If ≥10 agents have data for that KPI in that dept, compute p25 and p75 from those values.
Higher-is-better (CSAT, Occ, etc.): score = 100 at dept-p75, 0 at dept-p25, linear between.
Lower-is-better (Hold %, Refused, Shrinkage): score = 100 at dept-p25, 0 at dept-p75, linear between.
Re-aggregate with the same KPI weights you see in this modal → produces score_dept.
Below the 10-agent threshold for a KPI → that KPI silently falls back to its global benchmark for that agent (so the dept score is never penalised for thin data).

Look at the new Score (Dept), Tier (Dept), and Δ columns in the main agent table to compare lenses side-by-side. The Δ arrow tells you how much the lens shifts the agent's tier (↑ = dept lens is more generous, ↓ = harsher).

📊 Department benchmark coverage (this build)

Loading…

⚠️ Why most agents grade LOWER under the Dept lens

The Global benchmarks were calibrated at a specific point in time and tend to be more lenient than the current population's actual p25/p75. The Department lens recomputes p25/p75 from the live data every build, so 25% of each dept is at-or-below the target by construction. Expect roughly 50% of agents to grade DOWN one tier when switching to the Dept lens — that's not a quality regression, it's just the math being stricter relative to peers.

If we re-calibrated the Global benchmarks today (the benchmark calibration job), the two lenses would converge.

🎯 Which lens should I trust?

Use case	Recommended lens	Why
Org-wide ranking / executive view	Global	Single bar, comparable across teams
Manager coaching within their team	Department	Surfaces top/bottom within peer group
Cross-department rewards / promotions	Global	Fair across the org
AHT analysis specifically	Personalized (auto)	Always skill-adjusted, regardless of lens choice
Spotting weak departments	Global	Dept lens HIDES dept-level weakness

Performance Tier Thresholds (4-Tier Model)

Each tier maps to a distinct management action. We moved away from the A/B/C/D/F letter grading model (June 2026) toward growth-oriented tier names so feedback feels developmental rather than punitive. Cutoffs were re-anchored on 2026-07-29 to documented composite percentiles (Internal, n=6,392 scored): Exceeding = P80, Meeting = P40–P80, Below Target = P15–P40, Improvement Focus = <P15. Because KPI benchmarks are peer-relative, the composite is effectively a percentile rank — so every cutoff is a named, reproducible percentile. Resulting distribution of graded agents: ≈20% / 40% / 25% / 15%.

Tier	Score	Peer Position	Performance	Management Action
Exceeding Target	≥ 77	Top 20% (≥ P80)	Top-tier across the board — sustained excellence relative to channel peers	Recognize publicly; promote into mentor role
Meeting Target	54 – 76	P40 – P80 (median sits here)	Consistently at expectations above the channel median	Maintain; set stretch goals to reach Exceeding Target
Below Target	35 – 53	P15 – P40 (lower-middle)	Meets most expectations — opportunities in specific KPIs	Targeted coaching on weak KPIs; weekly 1:1
Improvement Focus	< 35	Bottom 15% (< P15)	Multiple KPIs below channel norm — requires structured support	Structured development plan with clear milestones

Why these specific cutoffs? Under percentile-relative scoring (each KPI graded against channel peers with p25→target→p75 anchors), an agent hitting the median on every KPI lands around 55–60, so the composite is itself a percentile rank. Rather than eyeball a tidy shape, every boundary is pinned to a named percentile of the actual Internal composite distribution (n=6,392): Exceeding = P80 (77), Meeting = P40 (54), Below Target = P15 (35), Improvement Focus = <P15 (<35). This makes each cutoff reproducible and defensible. Note: because scoring is norm-referenced, “% below target” is a curve and can never reach zero — an absolute service standard would require switching KPI benchmarks to criterion values, not moving these cutoffs.

Historical note: The legacy A/B/C/D/F letter scale was retired in June 2026; the historical D band is permanently merged into Improvement Focus. Internal data keys (single character) are preserved for downstream compatibility; only display labels and cutoffs changed.

Worked Examples — See the Math in Action

Two end-to-end scoring walkthroughs from the methodology reference. Use these to explain a specific grade to an agent, or to verify the calculation yourself from raw KPIs.

⚠️ Heads up: Aisha's example below uses illustrative benchmark values (CSAT target 95.3 / min 82.2, AHT max 585.2, etc.) to keep the arithmetic easy to follow. Current benchmarks have drifted by <1%; her composite score would change by less than 1 point under the live benchmarks.

Example 1 — Aisha · Overall Composite VOICE AGENT Exceeding Target

Profile: Voice Agent at Northstar CX Hub · Tenure 2.4 years · Skill mix: 100% Voice · 1,247 contacts handled · 38 CSAT responses · Both quality gates passed.

KPI	Actual	Benchmark	Score	Wt	Math	Contribution
CSAT	92.0%	Target 95.3 · Min 82.2	75	25	(92−82.2)÷(95.3−82.2)×100	1,875
AHT	380s	p25=280 · p75=480	50	10	100−((380−280)÷(480−280))×100	500
Hold %	4.5%	Target 3.3 · Max 18.4	92	5	100−((4.5−3.3)÷(18.4−3.3))×100	460
Refused	1	Target 2 · Max 6	100	5	1 ≤ target → cap at 100	500
CPH	1.6	Target 1.5 · Min 0.8	100	12	1.6 ≥ target → cap at 100	1,200
Call Quality VOICE	82	Target 75 · Min 50	100	8	82 ≥ target → cap at 100	800
Shrinkage	14%	Target 16.2 · Max 27.6	100	10	14 ≤ target → cap at 100	1,000
Shift Conf % v2	92%	Target 95 · Min 70	88	15	(92−70)÷(95−70)×100	1,320
Break Conf % v2	75%	Target 95 · Min 50	56	10	(75−50)÷(95−50)×100	560
Occupancy CTX	80%	Target 83.3 · Min 61 (context only)	—	0	not scored in v2	0
Productive % CTX	62%	Target 65.4 · Min 40.6 (context only)	—	0	not scored in v2	0
Utilization CTX	78%	Target 81.4 · Min 66.8 (context only)	—	0	not scored in v2	0
TOTAL				100		8,215

Overall = 8,215 ÷ 100 = 82.15 → Tier Exceeding Target
# 7.15 points above the Exceeding Target cutoff (75). Areas to focus on: Break Conformance (56) and AHT (50) are the lowest sub-scores. Under v2, break-overrun behavior actually moves the grade — whereas the old Occ/Prod/Util scores didn't reflect anything the agent could act on.

🎯 Example 2 — Maya & Carlos · Personalized AHT in Action

Same global AHT benchmark, two very different outcomes. This is why we moved to skill-weighted personalized benchmarks.

Maya — Chat-dominant agent

Rescued from F → C

Skill mix: 95% Chat · 5% Voice
Actual AHT: 520s

Old global benchmark: 415s → AHT score = 38 (F-tier)

Personalized (chat-weighted): p25=440s, p75=600s → AHT score = 63 (C-tier)

Maya was never underperforming her peers — she was just measured against the wrong yardstick.

Carlos — Outbound-voice agent

Unmasked from A → D

Skill mix: 100% Outbound Voice
Actual AHT: 380s

Old global benchmark: 415s → AHT score = 100 (artificial A-tier)

Personalized (outbound-voice): p25=280s, p75=380s → AHT score = 50 (D-tier)

Carlos was getting credit for being average-fast on a fast-channel global benchmark — personalization correctly reframes him as bottom-quartile among his actual peers.

💡 Why this matters: Two agents, identical "old" AHT scoring logic, opposite directions of bias. Personalized benchmarks remove the channel-mix artifact so the grade reflects peer-relative performance, not which skill the agent happens to be assigned to.

⚠️ Quality Gates (Exclusions)

Agents are excluded from scoring if:

Insufficient Data — below the channel-specific volume threshold (see table below). Not a quality concern — just thin data.
Quality Gate Violation — short-handled rate above the vertical p95 threshold (likely gaming or mis-routing — warrants investigation). Agents between vertical p90 and p95 are still scored but tagged with a soft “review” flag for the supervisor.

These are shown as separate bars in the Performance Tier Distribution chart so you can tell the difference at a glance.

Visibility-only statuses (not scored, hidden from tier chart):

Support Role — QA analyst, supervisor, or escalation-desk staff. They don't handle production contact volume by design, so no score is possible.
Ramp-Up — first month in the dataset (new hire still ramping). Explicitly NOT a performance concern.
RN (Legacy Email/Chat) — agent works on the legacy email platform. The the legacy email feed feed only gives us handled volume + identity for these agents; handle-time, CSAT, and adherence signals live in a separate WFM cache that isn't integrated yet. Look for the RN badge in the agent table — handled volume is real, everything else will show as —.

These agents still appear in the main table (for visibility and coaching context) and in the Excluded panel (grouped with their exclusion reason).

📊 Channel-Aware Volume Thresholds

A blanket 100-contact gate was unfair to specialist roles (Email, Escalation, Outbound) that handle fewer but more complex contacts. Thresholds are calibrated to each channel's actual baseline throughput.

Channel	Min Contacts	Rationale
Voice / Voice Spanish / Chat	100	High-volume channels; agents typically handle 1,000–2,000/mo
Offline	50	Mid-volume async work
Outbound / Email / Voice Escalation / Chat Escalation	25	Specialist roles; median monthly throughput ~40–60 contacts
Unknown channels (fallback)	100	Conservative default

✨ Impact: this change rescued ~2,000 specialist agents who were previously silenced as “N/A” but who legitimately performed work — reducing N/A from 17% → ~7–8% with no impact on scored agents.

CSAT Data Requirement

CSAT weighting: 25 of 100 possible points when scored.

Minimum 10 CSAT responses required for CSAT to be included in the weighted score (raised from 5 on 2026-06-29 — at n<10, the ±20pp margin of error is too noisy to grade on).
Below 10 responses: CSAT is excluded from the weighted average; the remaining KPIs reweight to 100% (same treatment as Call Quality for non-voice agents).
This is not a penalty — it's a fairness fix. The old behavior (0 points below 5 responses) unfairly punished specialists on low-survey channels (offline, escalation).

📈 Benchmark Data Source

All benchmarks are data-driven from actual performance:

Target (Good): 25th percentile of qualified agents (top 25% performance)
Minimum/Maximum (Acceptable): 75th percentile (bottom 25% performance)
Scoring Range: Linear scale between 25th and 75th percentiles
Period: Feb 1 - Aug 31, 2026 (212 days, Q1+Apr+May+Jun+Jul+Aug)

Questions? Contact your supervisor or WFM team for clarification on your score.

Top 10 Performers

Performance Tier Distribution

Tier Breakdown by Vertical

Tier Breakdown by Site

Tier Performance Quadrant

Tenure Insights — How Experience Shapes Performance

KPI Fingerprint by Tenure

Mastery Curve — Score vs Tenure

Heatmap — Tenure × Vertical

Bottom 10 Performers

Excluded Agents

🧮 KPI Calculation Methodology

📞 Contact Metrics

✅ Long Handled

⚡ Short Handled

📊 Total Handled

⏱️ Efficiency Metrics

Average Handle Time (AHT)

🚀 Contacts Per Hour (CPH)

⏸️ Hold Time %

❌ Refused Contacts

[WFM] Workforce Management Metrics

Occupancy % Context only — not scored

Productive % Context only — not scored

Utilization % Context only — not scored

📉 Total Shrinkage %

Shift Conformance % v2 NEW · Weight 10% Internal only

Break Conformance % v2 NEW · Weight 10% Internal only

Personal Time (min/month) Diagnostic only — not scored

⭐ Customer Satisfaction

CSAT % (Customer Satisfaction)

🎙️ Call Quality Score VOICE ONLY

🔄 Data Freshness & Period

📊 Scoring Methodology & Grading System

Overall Score Formula (v2)

KPI Weights (v2 model — Total = 100%)

⚖️ Weight Source by Department — Ops-tuned vs Default (v2)

🎯 NEW: Skill-Based AHT Benchmarking

🔬 How p25 / p75 Are Actually Calculated WORKED EXAMPLE

⚖️ How the 9 Static Benchmarks Were Calibrated CSAT EXAMPLE

🧪 Benchmarking Methodology — Three Lenses, One Score

Performance Tier Thresholds (4-Tier Model)

Worked Examples — See the Math in Action

Example 1 — Aisha · Overall Composite VOICE AGENT Exceeding Target

🎯 Example 2 — Maya & Carlos · Personalized AHT in Action

Maya — Chat-dominant agent

Carlos — Outbound-voice agent

⚠️ Quality Gates (Exclusions)

📊 Channel-Aware Volume Thresholds

CSAT Data Requirement

📈 Benchmark Data Source