The ML is the easy part. The ethics is the hard part.

How you’d build it (technique):

Define the label carefully. “Voluntary departure within 180 days” is typical. Include only regrettable departures if the goal is retention. Exclude retirements and involuntary terminations from the positive class.
Feature set (what predicts voluntary exit):
- Tenure, time-since-last-promotion, compensation-vs-market-benchmark.
- Manager tenure; manager change in last 6 months.
- Recent engagement-survey score and trend.
- Performance rating history and trend.
- Internal mobility signals (applied for another internal role).
- Commute distance, remote-work status.
- Peer attrition (team members recently left).
Model. Gradient-boosted trees (XGBoost, LightGBM) are the workhorse. Logistic regression is an acceptable baseline and much easier to explain to HR.
Calibrate the output. A “70% flight risk” must actually mean ~70% of employees at that score leave within 180 days. Platt-scale or isotonic regression.
Validate. Time-based holdout (train on employees as of Jan, predict for Apr–Sep departures). IBM famously reports ~95% accuracy on their internal model — be skeptical of that number in your context until you’ve validated.

The ethical traps — load-bearing to name:

Prediction without action = surveillance. If you surface flight-risk scores to managers without a retention playbook, you’re just ranking people for their manager to judge. Every deployment must pair with a specific intervention: compensation review, stay-conversation protocol, mobility support.
Self-fulfilling prophecy. Managers who see a high flight-risk score may pre-emptively deprioritize that employee (less investment, less interesting work), causing the departure. Monitor for this.
Proxy discrimination. Features correlated with protected attributes (commute, age-via-tenure, gender-via-function) can make the model discriminatory even without those attributes as inputs. Run fairness audits: do the model's positive predictions differ across demographic groups in ways not explained by the outcome?
Employee consent and transparency. Unlike customer churn, employees haven’t opted in. Some companies disclose the existence of the model and what features it uses; others deploy silently. The legal / reputational landscape is evolving (EU AI Act specifically; US case law less settled).
Access control. Who sees the scores? HRBPs only? Direct managers? Anyone outside the department? Score visibility is as important as score quality.

What strong candidates say:

How would you build a flight-risk model, and what could go wrong?

Stuck? Peek at a hint