Turnover prediction is one of the most widespread HR applications of AI in large organisations. But the gap between commercial demonstrations and real-world usage remains significant.
How these models work
Turnover prediction models aggregate signals from multiple sources: frequency of internal tool logins, number of weekly meetings, tenure, relative salary progression, annual review results, LinkedIn activity.
Each signal in isolation has little predictive value. It is their combination that allows the model to identify at-risk profiles with accuracy that surpasses managerial intuition.
What an industrial mid-market company measured
An industrial company with 3,200 employees (interviewed on condition of anonymity) deployed such a tool two years ago. Result: its manager turnover rate fell by 14% over two years, thanks to targeted retention actions triggered by the model’s alerts.
The actions themselves are straightforward: a one-on-one conversation with the manager, an early salary review, an internal mobility proposal. The tool does not recommend actions; it signals a priority.
What raises concerns
Several HR directors interviewed describe a tension that is hard to resolve. Using behavioural data to assess an employee’s intention to leave raises legitimate questions about workplace surveillance.
In France, using such data requires informing employees and filing a declaration with the CNIL (data protection authority). Companies that have done this correctly describe a broadly neutral reception from employee representatives, provided the tool is used to trigger support, not sanctions.
Key takeaway
Turnover prediction works technically. Ethical and legal deployment is possible, but it requires a level of transparency that few organisations have genuinely put in place.