People analytics predicting attrition: why 'flight-risk' scores are legally shaky

Short answer: AI that calculates a per-employee likelihood of leaving ("flight-risk") is — like absence prediction — one of the most sensitive HR applications. At the individual level it collides with purpose limitation, proportionality and the GDPR's protection against profiling, and it carries the real danger of a self-fulfilling prediction and discrimination. Only at an aggregated level is it sometimes defensible.

What such a model actually does

An attrition model combines data such as appraisals, salary progression, leave, email behaviour, tenure and sometimes external signals into a score: how likely is this person to leave within X months? That is profiling within the meaning of the GDPR — the automated evaluation of personal aspects to predict behaviour. The strictest requirements apply at once.

Purpose limitation and proportionality

Data collected for payroll, appraisal or access control may not simply be reused to predict departure. That is a new, different purpose (Art. 5 GDPR). The employer must show the processing is necessary and proportionate — and that is exactly where it pinches: the aim (retaining staff) is legitimate, but the means (permanently screening everyone for loyalty) is almost never the least intrusive option. A good conversation often achieves more than a score.

Article 22 and the human measure

If consequences are attached to a high flight-risk — no promotion, no training, no renewal — this approaches an automated decision with significant effects (Art. 22 GDPR), in principle prohibited without a valid basis and safeguards. And once the score steers decisions about the individual worker, the system can become high-risk under the AI Act as an HR application, with the corresponding obligations.

The self-fulfilling prediction

The most subtle risk is behavioural. Anyone labelled a "flight risk" receives — consciously or not — less investment, less trust, fewer opportunities. That is precisely what pushes people out the door. The prediction makes itself come true and can no longer be falsified. Moreover, many predictive features (age, part-time status, caring duties, origin via language or network proxies) correlate with protected grounds, so discrimination creeps in without anyone intending it.

When (rarely) it is defensible

At an aggregated, non-traceable level turnover analysis can be legitimate: which departments, job groups or pay structures show high attrition, and what organisational causes lie beneath? That targets the system, not the person, and can improve policy. The line is hard: as soon as the outcome is traceable to an individual or weighs in decisions about people, it tips into unlawful.

What to do

• Do not calculate individual departure scores — purpose limitation and proportionality are almost always missing.
• Attach no consequences to a predicted flight-risk in promotion, training or renewal.
• Work aggregated and non-traceable, aimed at causes (workload, pay, leadership).
• Test for proxy discrimination and involve the works council before deployment.

Reducing turnover is a legitimate aim, but permanently screening every employee for loyalty undermines the very trust that makes people stay. The law and common sense point the same way here.