cx-attrition-early-warninglisted

# Attrition early warning Most "attrition prediction" projects fail in one of two ways: they **surveil what they should not** (private life, health proxies, social graph), or they **produce a list nobody acts on** — a risk score with no owner, no intervention, and no feedback loop. Both erode trust and do not reduce leavers. **The ethical boundary:** use **operational data the organisation already holds for work management**, aggregated or team-level where possible. The goal is **retention design** — fixing schedules, load, tooling, and manager capacity — not pre-emptively managing individuals out or ranking "flight risks" in one-to-ones. If the user asks for individual prediction to target people, redirect to team-level diagnostics and voluntary stay interviews. Individual lists belong behind HR, legal, and union review this skill does not substitute for. ## Signals that are in bounds These reflect **working conditions**, not private life: | Signal | What it may indicate | Caveats | | --- | --- | --- | | **Occupancy / utilisation trend** | Sustained overload; no recovery time | Confounded by shrinkage misclassification | | **Schedule instability** | Frequent last-minute changes, split shifts, cancelled rest days | Measure from WFM logs, not judgement | | **Overtime / involuntary extra hours** | Burnout path; childcare strain via work, not home surveillance | Distinguish voluntary vs mandated | | **QA score volatility** | Disengagement, confusion after change, or toolin