When someone on a team hands in their resignation, it is rarely a complete surprise. In the preceding months there were signs — quieter meetings, a project delivered without the usual spark, holiday days piling up unused. The problem is that these signals sit in different systems and in perceptions no one joins up in time. By the time the decision is announced, it is too late to influence it.
People analytics proposes something simple to say and demanding to do: use the data a company already holds about work to estimate, honestly, who is most at risk of leaving — and to act while there is still room. It is not a crystal ball, nor a surveillance system. It is a structured way of turning scattered hunches into clear priorities for the people who lead teams.
This article shows how to approach attrition prediction without falling into two traps: the gut feel that ignores the data, and the technocentrism that treats people as rows in a table. Between the two there is a useful path, and that path is what we are going to talk about.
What predicting attrition means — and what it does not
Predicting attrition means estimating the probability that an employee leaves the company within a defined horizon — the next six or twelve months, say. The output is not a verdict, it is a risk score that ranks attention: who is worth a conversation, a role adjustment, a review of terms.
What the prediction is not: a justification for treating differently anyone "the model" flagged, nor a replacement for the judgment of those who manage. A high score is an invitation to ask why, not a label. And there are departures you do not even want to prevent in order to retain — people leaving to grow, people who no longer fit the project. The goal is to reduce unwanted attrition, the kind that costs knowledge, clients and morale.
Why attrition costs more than the payroll shows
It is tempting to see replacing a person as the cost of an advert and a few weeks of recruiting. The real bill is bigger. There is the team's time spent interviewing, the learning curve of whoever arrives, the normal mistakes of someone who does not yet know the processes, and the tacit knowledge that walks out of the door and was written down nowhere.
Conservative estimates put the cost of replacing a skilled professional at between half and twice their annual salary, depending on the role and seniority. Even if your company sits at the low end of that scale, it is enough to justify the effort of anticipating departures rather than managing them after the fact.
Which data to use — and which to leave out
The raw material is usually scattered but accessible. Useful, defensible signals include:
- Time in role and in the company: risk tends to have predictable peaks — around the first year, and when someone stagnates for years in the same role without progression.
- Progression: time since the last promotion or raise, and how it compares with peers in the same role.
- Relative pay: position against the market band and against colleagues, more than the absolute figure.
- Workload and work patterns: systematic overtime, unused holiday, prolonged peaks with no relief.
- Team context: recent changes of manager, nearby departures, reorganizations.
And what to leave out: private content such as emails or messages, health data, and any variable that acts as a proxy for gender, age or origin — not only because they are ethically problematic, but because they make the model legally fragile and unfair. Fewer sensitive data and more data about the work tends to produce models that are more robust and easier to defend.
How to build a first version without technical overkill
You do not need deep learning to start. A good first model is simple, explainable and honest about its limits. A pragmatic path:
- Define the target clearly: "voluntary departure within the next 12 months". Without this definition, everything else stays ambiguous.
- Gather history: two to three years of data on those who left and those who stayed, so the model learns real patterns.
- Start with a logistic regression or a simple tree: they give probabilities and, above all, let you see which factors weigh most.
- Validate on data the model has not seen: train on one period, test on the next. If it only works on the past it already knows, it is useless.
- Translate the score into bands: low, medium, high risk. A probability of 0.63 does not help a manager; three bands with suggested actions do.
Resist the temptation to chase the last decimal of accuracy. A model that is 80% good and that managers understand and use is worth more than one that is 90% good and no one trusts.
From score to action: the part that really matters
A model that only produces lists is an expensive report. The value appears when the score triggers a conversation. For each person at high risk, the question is not "how do we keep them", it is "what changed and what is within our reach". Sometimes the answer is a pay review; often it is something cheaper — a new project, recognition, clarity about the next career step, less load.
It is worth deciding in advance who receives the signals and how. As a rule, the right person is the direct manager, supported by HR, and never a list circulated without context. The signal exists to prepare a good conversation, not to excuse having one.
Mini-case: the company that stopped losing good people blindly
A services company with around 300 people was losing close to 18% of its headcount a year, with the added problem that departures were concentrated among its most experienced profiles. Recruiting was becoming a permanent cost and teams lived in replacement mode.
Instead of buying an expensive tool, they combined what they already had: tenure, progression history, pay position against the market, and training participation. A simple model highlighted two patterns no one had systematized — risk shot up among professionals in their second and third year with no role change at all, and it worsened when pay fell more than 15% below the market median.
With that, HR and managers began reviewing a short list of priority cases every quarter. They did not act on everyone — they acted on the few dozen right people, with timely promotions, selective pay adjustments and development plans. A year later, unwanted attrition had fallen to around 12%, and the saving in recruitment and replacement paid for the effort many times over. The model did not "predict" anything magical; it simply forced the organization to look in time.
Ethical and legal traps to take seriously
Predicting people's behaviour calls for care. A few principles that save trouble:
- Transparency of purpose: the goal is to retain better, not to punish whoever looks likely to leave. If a score ever harms someone, the project has lost its legitimacy.
- Explainability: it must be possible to say, in plain language, why someone appears at risk.
- GDPR compliance: a clear legal basis, data minimization, and informing people about the processing.
- Vigilance against bias: check whether the model systematically penalizes any group; if it does, fix it.
How to know if it is working
Measure what matters, not what is easy. The model's accuracy is secondary to the result on the ground: is unwanted attrition falling? Did the suggested interventions change decisions about real people? Is the time between "risk signal" and "actual conversation" shrinking? It is also worth tracking a rough counterfactual — comparing teams that used the signals with teams that did not — so you do not credit the model with improvements that would have happened anyway.
In practice
Predicting attrition is not about sophisticated technology; it is about looking in time at signals the company already has and turning that look into better conversations. Start small: define clearly what you want to predict, use data about the work rather than about people's private lives, prefer a model that explains itself to one that impresses, and measure success by real retention, not by accuracy on paper. Done with respect and transparency, people analytics stops being a buzzword and becomes what it should have been all along: a way to take better care of the people who make the company happen.
