Your flight-risk dashboard flags a valued engineer. Her manager books a stay conversation. Three weeks later, she resigns.
We usually conclude the conversation came too late. But there is another possibility: Have we targeted the wrong person?
For years, we’ve judged retention analytics on one question: How accurately can we predict who will leave? Vendors advertise accuracy. Dashboards rank flight risk. Managers get a list of names. But prediction accuracy is the wrong scoreboard.
The problem is not necessarily that organizations are predicting turnover incorrectly. It is that they are using predictions to make decisions; those predictions were never designed to support.
A model can correctly identify everyone likely to leave and still send your retention budget toward people who will go regardless, people whose reasons for leaving aren’t addressed by the intervention or departures the organization can absorb.
The question talent leaders need to answer is different: Where will an intervention change an outcome that matters? As organizations set retention priorities and budgets for 2027, that’s the question worth asking.
Risk is not intervention value
A good retention decision rests on three questions. Most flight-risk models answer only the first: Will they leave? Predictive models estimate the probability of departure.
What would it cost if they did? Not a generic replacement multiplier, but the actual organizational value at stake: scarce skills, client relationships, institutional knowledge, ramp-up time, succession depth and strain on the remaining team.
Would anything we do change their decision? And, if so, what? That third question changes the problem. At its simplest: Expected value of intervention = reduction in departure probability × cost of departure avoided − cost of intervention
This is a decision framework, not a precise accounting formula. Estimates should reflect realistic, avoidable costs and uncertainty about whether the intervention works.
Call the first part retention lift: How much does a particular action reduce the probability that an employee or employee group will leave? There is a catch: Retention lift cannot be read directly from a flight-risk score. It requires evidence about what would have happened without the intervention.
Historical correlations can identify patterns, but they cannot establish whether a retention action caused someone to stay. Someone can have a 90 percent probability of leaving but almost no retention lift from the intervention you’re considering. Another employee may have only a 40 percent probability of leaving, but the right intervention could reduce it substantially.
The traditional model targets the first employee, but it may be abetter choice to invest in the second.
Other fields have already found the problem
There is evidence for this distinction outside HR. Marketing researcher Eva Ascarza combined field experiments with machine learning to examine whether organizations should target retention efforts toward customers most likely to leave. The answer was no. High-risk customers were not necessarily the most responsive to intervention. Targeting based on what researchers call “lift,” the change attributable to the intervention, was more effective than targeting based on churn risk.
Economists Susan Athey, Niall Keleher and Jann Spiess found a related pattern in a field experiment involving more than 53,000 college students. Their question was who should receive nudges encouraging financial-aid renewal. Simply targeting students based on their predicted outcome was not the best criterion. What mattered was the expected effect of the intervention.
These studies do not prove that intervention-based targeting will improve employee retention. They demonstrate why predicting an outcome and predicting the effect of an intervention are different analytical problems. Talent leaders should test whether that distinction holds true in their own organizations.
The employee at greatest risk is not necessarily the one most likely to respond to your intervention. Gallup’s research suggests there is room to find out. Forty-two percent of voluntary leavers said their manager or organization could have done something to prevent their departure. Yet 45 percent said neither a manager nor another leader had proactively discussed their job satisfaction, performance or future with them during the three months before they left.
Some turnover appears potentially responsive to organizational action. Some is not. The challenge is learning to spot the difference.
Not every departure costs the same
Risk-first retention strategies have another problem: They can implicitly treat departures as equivalent. They aren’t.
The steady analyst who is the only person who understands a legacy billing system may represent a far more consequential loss than a highly rated generalist with three ready successors.
Recruiting and onboarding costs are only part of the equation. A departure can also mean lost client relationships, delayed projects, reduced team capacity, specialized knowledge walking out the door and months required to rebuild expertise.
That means “regrettable turnover” needs a more rigorous definition. Instead of equating regrettable loss with high performance, assess criticality, skill scarcity, replacement difficulty, ramp-up time, succession depth and team dependency. That turns regrettable turnover from an HR label into a business calculation.
What does this look like?
Consider a hypothetical regional health services company. Its flight-risk model identifies patient access employees as a high-risk group. Clinical informatics analysts show moderate risk.
A traditional retention strategy prioritizes the first group. But exit data reveal many patient access employees leave for school, relocation or other reasons a retention bonus rarely changes. The organization could spend heavily without materially affecting turnover.
The informatics analysts present the opposite problem. They are difficult to replace, hold specialized organizational knowledge and require long ramp-up periods. Interviews suggest stalled career progression is a significant reason for leaving. A career pathway and title review could plausibly change that decision.
The organization has been asking: Who is most likely to leave? But it should be asking: Where can we create the greatest retention lift on the organizational value most worth protecting?
No more accurate flight-risk model solves that problem. It requires a different decision rule.
When you don’t know what works, test
Sometimes organizations know turnover is costly but don’t know what will change it. That’s not a reason to guess. It’s a reason to test.
Suppose newly promoted supervisors are leaving at an elevated rate. HR believes coaching could help but has no evidence that it affects retention. Instead of immediately scaling a companywide program, pilot the intervention with an eligible group. A wait-list design can extend the opportunity to everyone while still allowing the organization to compare outcomes before scaling the investment.
Start with lower-risk, non-monetary interventions such as coaching, career conversations, mentoring, internal mobility, scheduling flexibility or manager support. The objective isn’t to turn employees into laboratory subjects. It’s to stop spending substantial resources on retention practices without knowing whether they change the outcome they’re intending to change.
Handle with care
There are limits to this approach. Retention decisions are visible. A bonus offered to one employee affects another employee’s perception of fairness. “Invest where it works” can also become an excuse for rewarding whoever threatens to leave or withholding opportunities from groups an algorithm labels unlikely to respond.
That’s why the targeting rule needs governance. Compensation decisions require legal and pay-equity review. Models should be tested for differential effects across employee groups. Leaders should be able to explain why one population received an intervention, and another did not. Organizational value should not be confused with compensation, seniority or visibility. Criticality assessments must account for overlooked contributions and be tested for inequitable outcomes.
Regulation is moving in the same direction. Colorado’s 2026 Automated Decision-Making Technology Act establishes requirements, beginning January 1, 2027, for certain technologies that materially influence consequential decisions, including covered employment decisions. Not every use of flight-risk analytics necessarily falls within its scope. But the broader lesson matters beyond any one state’s requirements: A model should inform a talent decision, not quietly become the decision.
5 questions to ask now
Talent leaders don’t need to wait for a new generation of analytics. They can start by changing the questions they ask.
What losses matter? Define regrettable loss using criticality, scarcity, replacement difficulty, ramp-up time, succession depth and team dependency.
What changed after last year’s retention spending? Review bonuses, counteroffers, stay programs, career interventions and other investments. Look for evidence that they changed outcomes.
Where are we measuring risk but not response? Ask whether the employees with the highest predicted flight risk the employees who are most likely to respond to the intervention being offered.
What could we test before scaling? Use small, fair pilots to estimate retention lift, particularly for non-monetary interventions.
What are our vendors optimizing? Don’t ask only how accurately a system predicts attrition. Ask what evidence shows that acting on its recommendations improves retention outcomes.
Before approving another retention investment, ask what evidence suggests the action will change an outcome that matters, rather than simply reward someone predicted to leave.
Start with one critical employee population, not the entire workforce. Identify the costs associated with turnover, examine why employees leave and select one intervention the organization can reasonably test. Compare outcomes, examine unintended consequences and use the findings to decide whether to expand, modify or discontinue investment. Begin with group-level evidence rather than assuming the model can reliably predict how a particular employee will respond.
The question that matters
The next generation of retention analytics shouldn’t simply get better at predicting behavior. It should help leaders make better decisions. That means changing the unit of analysis.
Instead of asking only about an employee’s probability of leaving, ask about the incremental value of a specific intervention for a specific employee or talent segment.
- Who might leave?
- What would that loss cost?
- What action could change the outcome?
- Is the expected benefit worth the investment?
Predicting who might leave is getting cheap. Knowing which actions change outcomes, for whom and at what cost, is not. If we keep perfecting flight-risk accuracy, we may simply get better at watching the right people leave.
