Shift Adherence Anomaly Detector

What it does

It processes a daily adherence/RTA dump at the agent × interval level and answers four questions with numbers: who/when/how much deviated from plan, is the deviation noise or a real anomaly, which root-cause class it belongs to, and which intervention — done first — recovers the most service level (SL).

The skill applies the field's real canon in sequence. The Adherence vs Conformance distinction (COPC/WFM) separates a "break in the wrong slot" (conformance OK → behavior/scheduling) from a "lost-time" event (both low → shrinkage/absence). Shewhart control charts + Western Electric (WECO) rules derive μ and σ from each agent/team baseline, build ±1σ/±2σ/±3σ zones, and use R1 (single point >3σ, acute), R2 (shift), R3 (drift), R4 (8 consecutive points = persistent bias) so that only rule-triggering deviations count as anomalies — "below target but stable" never raises a false alarm. An Erlang-C bridge converts lost agent-hours per interval into estimated SL-point loss and extra abandons. Shrinkage taxonomy (planned/unplanned, internal/external) + 5-Why/Ishikawa assigns each anomaly to a root-cause branch (agent behavior, scheduling, forecast, system, management, external). RICE (Reach×Impact×Confidence/Effort) ranks the interventions. The result is a 0–100 Adherence Health Score with a confidence level and owned "now / intraday / structural" actions.

When to use it

When you hold a daily conformance/RTA dump and see "low adherence" — especially when multiple agents deviate simultaneously (deciding individual vs systemic), when you need to separate an isolated acute event from a drifting trend, or when you want to point limited RTA-desk time at the highest-SL-impact fix. A daily / end-of-shift pulse for the team lead, RTA desk, or TP-BPO operations owner.

Method / frameworks

• COPC Customer Experience Standard — adherence/conformance/shrinkage definition canon.
• Shewhart control chart + Western Electric (1956) R1–R4 — statistical anomaly screening (±3σ, ~1/53 false-alarm rate).
• Erlang-C — converting adherence loss to SL (capacity ↔ service level).
• Ishikawa fishbone + 5-Why (TPS) — WFM-adapted 6M root-cause classification.
• RICE scoring (Intercom) — impact/effort prioritization.
• KVKK/GDPR — agent-level performance data is personal data; purpose limitation applies.