A predictive retention model assigns an exit-risk score to each employee based on tenure, role, manager, compensation, engagement signal, and historical attrition patterns. The model’s output is a prioritization tool for HR business partners, not a decision tool for managers.

The structural definition

The model is a supervised machine learning system trained on historical employment data — past employees flagged as voluntary exits within 12 months. The training data carries demographic fields, tenure, role, manager hierarchy, compensation history, and engagement survey results. The model outputs a probability score (0 to 1) for each current employee. The 5 AI Applications Revolutionizing HR & Recruiting — Complete 2026 Guide expands the application context.

What the model does well

The model surfaces patterns the human network misses — combinations of signals (tenure plus manager turnover plus role family plus compression) that correlate with exit but escape intuition. Above 500 employees, the model produces signal that materially exceeds HR leadership intuition. Below 500, intuition outperforms the model. The 8 HR metrics guide covers the metric design that supports the model.

What the model must not do

The model must not inform compensation decisions, performance ratings, or termination decisions. The model produces a prioritization signal for retention conversations — nothing more. Using the model for adverse actions creates legal exposure and breaks the trust required for the model to surface honest patterns.

The governance requirements

The model requires four governance layers — data lineage from source systems to model output, quarterly bias audit on the model’s recommendations, explainability of model output for any flagged employee, and named escalation path when the model and HR business partner disagree. The governance burden is heavier than for other AI applications because the model’s outputs influence human conversations. The report design for strategic impact guide covers the reporting layer.

How the model integrates with HR business partners

The model produces a weekly or monthly list of employees ranked by exit risk. HR business partners review the list, apply their context (recent conversations, known life events, internal mobility plans), and prioritize retention conversations. The partners’ context overrides the model rank when the partners have information the model lacks.

The bias considerations

The model can encode and amplify historical biases — if historical attrition correlates with protected class, the model will surface protected-class employees disproportionately. The quarterly bias audit on model outputs is the control. If the audit reveals disparity, the model retrains with corrected weights or the feature set adjusts to remove the disparity driver. The data literacy for strategic HR guide covers the data understanding.

FAQ

What accuracy does the model deliver?

Production models achieve 65 to 75 percent precision at the top decile of risk — meaning 65 to 75 percent of employees flagged as highest risk actually exit within 12 months. The remaining 25 to 35 percent stay, in many cases because the retention conversation worked.

Does the model see compensation data?

Yes — compensation history and compression are strong predictors. The model sees the data; the model’s outputs do not inform compensation decisions. The separation is the governance discipline.

How often does the model retrain?

Quarterly retraining is standard. Major workforce changes (acquisition, layoff, reorg) trigger an interim retrain. The Make.com HR productivity guide covers the orchestration that supports the retraining cadence.