The same measurement and causal rigor, turned on the workforce: predicting who leaves, then going past prediction to who an intervention can actually keep — and auditing the tools that make employment decisions.
Past predicting who leaves to estimating who an intervention can actually keep. A T-learner recovers heterogeneous treatment effects on semi-synthetic HR data and beats flight-risk targeting on the same budget — with the randomization caveat stated up front.
A turnover-prediction model on public HR data that goes past accuracy: interpretable feature importance, the organizational-behavior theory behind each predictor, and a fairness review of who the model flags and why.
Open-ended employee comments turned into structure: embeddings and topic modeling mapped to engagement constructs, with a read on how themes track measured outcomes rather than just word counts.
You have a retention budget and 120 employees. Target by flight risk, or by uplift, and watch the same budget save very different numbers — because the people most likely to leave often aren't the ones you can move.
Attrition models answer the wrong question. The one that matters — who stays because of what you do — is causal, and a risk score can't reach it.
A bias-audit standard for automated employment decision tools that goes past the four-fifths ratio to differential prediction and measurement invariance — the depth the new hiring-AI laws imply but rarely get.