Detecting Flight Risk 73 Days Before Resignation
By the time an employee submits their resignation, the decision was made weeks or months ago. We built an AI that identifies risk early enough to act — here's what the model looks at.
Resignation is not an event. It is the visible end of a process that typically began 60 to 90 days earlier with a conversation the employee had with someone outside the organisation. By the time a resignation letter lands in HR's inbox, the mental departure has already happened. Retention conversations at that stage succeed less than 20% of the time.
This is why traditional HR analytics fail at retention. They measure what happened after the decision was made.
What happens before resignation
We studied 3,400 voluntary departures across seven organisations over a 30-month period, tracking every digital signal that could be measured without privacy violation: internal communication patterns, project assignment history, performance review trajectories, manager relationship signals, learning platform engagement, and survey response sentiment.
The patterns before departure are consistent and detectable. In the 90 days before a resignation:
- Internal communication volume drops 34% on average - Project contributions shift from leading to supporting roles - Learning platform engagement drops to near zero - Response time to internal messages increases - Survey sentiment scores, where available, decline 2-4 points on a 10-point scale
No single signal is definitive. In combination, they form a pattern that our model recognises with 84% precision at the 73-day mark — meaning 8 out of 10 employees flagged as flight risks do in fact leave within the following 90 days if no intervention occurs.
What the model does not use
We are deliberate about exclusions. The model does not look at individual message content, salary data, performance improvement plans, or any information that employees would consider private. The signals are all behavioural patterns — how people work, not what they say or earn.
This is both an ethical and a practical choice. Models trained on compensation data optimise for counter-offers, which have a poor long-term retention record. Models trained on engagement patterns optimise for the actual conditions that make people want to stay.
The intervention window
Seventy-three days is long enough to act meaningfully. A manager can restructure a role, open a career development conversation, address a team dynamic issue, or revisit a compensation bracket — actions that require organisational process time that a 2-week notice period does not allow.
Organisations using the system have reduced voluntary attrition by an average of 34% in the first year. The reduction is highest in high-performing employees — the group most likely to be recruited away — because the model is calibrated to be more sensitive to signals from employees with strong performance records.
The output
The system surfaces a weekly watchlist to HR business partners and managers — not an alarm, but a conversation starter. Each flagged employee comes with the top three contributing signals and a suggested engagement approach. The manager does not see "Employee X has a 78% flight risk score." They see "Employee X has reduced project leadership involvement and has not engaged with any development resources in 6 weeks. Suggested: check in on career growth priorities."
That framing changes the conversation from reactive to proactive — and changes the outcome.
