Business Case · People Analytics
A single accuracy number doesn't tell HR who needs what. So the 294 held-out employees were broken into four real-world outcomes — and each one was priced.
Every one of the 294 held-out employees falls into one of four groups. That is how we briefed HR: not a single accuracy number, but who needs what.
No intervention. Confirms the model is not over-flagging.
A brief, low-key check-in. Cheap even when the flag is wrong.
The blind spot — about 4 in 10 leavers. High-value roles still need a manual safety net.
Where the value lives. Retention conversation within days.
SHRM puts replacement cost at 50–200% of annual salary. Average pay in this dataset is about $57K. Of the 29 leavers the model caught, retaining even one in three — roughly 10 people — avoids about $290K to $1.1M in replacement cost. False alarms are conversations. Missed resignations are replacement cycles. That is why we optimised for recall, not accuracy.
The 84 flagged employees are not treated the same:
Quarterly check-in for the HRBP. No urgency; logged for trend tracking.
Retention conversation within two weeks. The human still makes the call.
Escalate to manager and HRBP together — the costliest miss to allow.