What problem does it solve?
By the time an employee formally resigns, the decision is often close to final, and the manager finds out at the point retention options are narrowest. Exit interviews describe why people left, but only after they are gone, so the same pattern repeats with the next person on the team. HR teams add engagement surveys and stay interviews, but neither reaches every employee often enough to catch a change in risk as it happens.
Replacing an employee is also expensive in ways that rarely show up in one line of a budget: recruiting cost, the vacancy itself, ramp time for a replacement, and the knowledge that leaves with the person. A model that scores risk continuously, from data the organization already has, turns retention from a reaction into something a manager can act on before the resignation letter.
How does it work?
- Assemble the signal. The model draws on tenure, role, compensation relative to the market and to peers, manager changes, performance ratings, promotion history, engagement survey responses and, where available, internal mobility activity.
- Score risk. Each employee gets a risk score and the factors behind it (for example, below market pay for the role, a recent change of manager, or a stalled promotion), refreshed on a regular cycle rather than once a year.
- Route to a person, not a dashboard alone. The manager or HR business partner sees which specific employees are at risk and why, so a stay conversation or a compensation review can happen before the person is already interviewing elsewhere.
- Validate and adjust. The model's predictions are checked against who actually leaves, stays or is promoted, and retrained; workforce planning uses the aggregate pattern to see where attrition risk is concentrated by team, level or location.
- Audience
- Employee facing
- Autonomy
- Assist
- Adoption
- Early adopters
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Accuracy | Too few to pool | about 95% | 1 | 1 organization |
| Cost savings | Not pooled | about USD 300 million | 1 | 1 organization |
Value drivers: Risk and loss reduction, Lower cost to serve, Employee productivity.
Indicative value
A company with 10,000 employees and 12% annual voluntary turnover
USD 900,000 to USD 4.8 million
Avoidable turnover cost prevented per year
How this is calculated
Formula: employees * baselineTurnover * reducibleShare * costPerDeparture. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Employees employees, employees | 10,000 | 10,000 | The reference company. |
| Baseline annual voluntary turnover rate baselineTurnover, fraction of employees | 0.12 | 0.12 | Editorial assumption, replace with your own turnover rate. |
| Share of voluntary departures a timely intervention can prevent reducibleShare, fraction of voluntary departures | 0.05 | 0.1 | Editorial assumption, conservative against Visier's reported 10% reduction in truck driver turnover at Pitney Bowes, https://www.visier.com/customers/pitney-bowes/, since that result is for one role from a broader people analytics rollout, not a per employee risk score applied to the whole workforce. |
| Fully loaded cost of an avoidable voluntary departure (recruiting, vacancy, ramp time) costPerDeparture, USD per departure | 15,000 | 40,000 | Editorial assumption, varies widely by role; replace with your own cost model. |
What it leaves out: Assumes a share of voluntary departures are genuinely preventable with a timely, well targeted intervention; many departures are not (relocation, retirement, a role that no longer exists). It leaves out the cost of running the model, the risk of employees reacting badly to being scored, and any effect of retention actions on people who were never actually at risk.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
IBM
United States · Technology and software · 2019
IBM built its own "predictive attrition program" using AI on internal HR data to flag employees at risk of resigning. Then CEO Ginni Rometty described the results publicly at a CNBC @Work conference in April 2019: the model was in the 95% accuracy range at identifying workers planning to leave, and the program had saved IBM nearly $300 million in retention costs.
- Accuracy: about 95%
"IBM artificial intelligence technology is now 95 percent accurate in predicting workers who are planning to leave their jobs, said Rometty."
Claimed by: organization - Cost savings: about USD 300 million
"AI has so far saved IBM nearly $300 million in retention costs."
Claimed by: organization
Pitney Bowes
United States · Technology and software · 2019
Pitney Bowes, a global shipping and mailing technology company, adopted Visier's people analytics platform to analyze, identify and predict issues across the employee lifecycle. Visier reports a 10% reduction in truck driver turnover and over 400 self service people analytics users, and Pitney Bowes' VP of Total Rewards and HR Technology said the platform let the company change how it onboards and engages new hires to improve retention. The platform gives a general view of retention drivers across the workforce; neither Visier nor Pitney Bowes describes it scoring an individual employee's risk of leaving.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Clean, joined HR, payroll, performance and engagement survey data over enough history to model against actual departures
- A defined population and time horizon for what counts as a resignation the model should predict
- Manager and HR business partner training on what the score means and does not mean
- A retention playbook describing what a manager is expected to do with a high risk score
Systems to integrate
- HR information system for tenure, role, compensation and performance data
- Engagement or pulse survey platform
- Payroll, for compensation relative to market and peers
- Manager facing dashboard or existing people analytics tool
Complexity: Medium
The modelling itself is standard machine learning on structured HR data. The harder parts are data quality across systems that were never built to talk to each other, and change management: managers have to act on the score, and the organization needs a policy for how the score is shown and used so it does not feel like it decided who gets fired.
- 1
Define what the model predicts and for whom
Decide the population, the horizon (for example resignation risk in the next six months) and whether the model runs company wide from the start or is piloted in one function first.
- 2
Build from data the organization already has
Do not launch a new mandatory survey to feed the model. Start with HR, payroll and performance data that already exists, and add engagement signals the organization already collects.
- 3
Give managers the reason, not just the score
A bare risk number invites either panic or dismissal. Connect the score to specific, addressable drivers, such as below market pay for the role, a recent change of manager or a stalled promotion, so the manager has something concrete to raise in a stay conversation.
- 4
Validate before it drives action at scale
Check the model's predictions against actual departures, promotions and stays for at least one full cycle before managers are expected to act on it, and keep validating after launch.
- 5
Decide how employees are told, if at all
Decide, before launch, whether attrition scoring is disclosed as a workforce planning tool and whether an individual employee is ever told their own score. Make this a deliberate policy, with legal and works council input where required, rather than a default.
Guardrails
- The model never triggers an employment action (termination, non promotion) on its own; a person always decides what, if anything, to do with a risk score
- No protected characteristics or obvious proxies for them used as model features
- Regular fairness testing of scores and any resulting actions across employee groups
- Retention offers and stay conversations documented, so an unusual pattern can be reviewed
KPIs to instrument
- Precision and recall of the model against actual departures, tracked over time
- Voluntary turnover rate for flagged versus unflagged employees after an intervention
- Manager action rate on flagged employees (did a conversation actually happen)
- Fairness metrics across employee groups for both scores and outcomes
Human in the loop
HR and the manager decide whether and how to act on a risk score; the model surfaces risk and drivers, it does not decide pay, promotion or termination. HR owns model validation, fairness testing and the policy on what employees are told.
Common failure modes
- The model is accurate but nobody acts on it
- A dashboard nobody opens changes nothing. Route flagged employees to a specific person with a specific expected action and a deadline, and track whether it happened.
- Employees learn they are being scored and disengage further
- A leaked or mishandled disclosure that scoring exists can itself damage trust. Decide the communication policy deliberately, with legal and employee representative input.
- Compensation ends up as the only lever
- It is easiest to retain someone with a raise, but a raise driven by a flight risk score can create its own pay equity problem. Coordinate flight risk retention actions with the organization's pay equity process rather than running them separately.
- The model drifts as the business changes
- A model trained before a reorganization, a new competitor or a change in the labour market can quietly stop predicting well. Revalidate on a fixed schedule, not only when someone notices it is wrong.
What are the risks and rules?
EU AI Act
High risk
Annex III point 4(b) lists AI systems intended to monitor and evaluate the performance and behaviour of workers as high risk, and scoring an employee's likelihood of leaving from their performance, compensation and engagement data is that kind of evaluation, even though the output feeds a retention conversation rather than a punitive action. Deployers must inform workers' representatives and affected employees before use (Article 26) and keep human oversight over any action taken on the score.
Guidance
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 4 covers employment, workers' management and access to self employment, including monitoring and evaluating performance and behaviour.
- Article 26, obligations of deployers of high risk AI systems (European Union, Europe). Human oversight, logs, and informing workers' representatives and affected workers before a high risk system is used at the workplace.
Controls to put in place
- Named HR owner for the model, its validation and its fairness testing
- Documented policy on whether and how employees are told scoring exists
- Retention actions and stay conversations logged for review
- Coordination with the pay equity process before compensation is used as a retention lever
Frequently asked questions
- What results have companies reported from attrition prediction?
- IBM's then CEO Ginni Rometty said in 2019 that the company's AI could predict which employees were about to leave with about 95% accuracy and had saved IBM nearly $300 million in retention costs. Vendors of broader people analytics platforms report retention gains too: Visier says Pitney Bowes cut truck driver turnover by 10% after using its platform to understand what was driving departures, though that result is about people analytics generally rather than an individual attrition score.
- Does the model decide who gets a raise or is let go?
- No. It scores risk and surfaces the likely drivers; a manager or HR business partner decides whether and how to act, whether that is a stay conversation, a role change or nothing at all.
- Is attrition scoring high risk under the EU AI Act?
- Generally yes. Scoring an employee's likelihood of leaving from their performance, pay and engagement data is a form of monitoring and evaluating worker behaviour under Annex III point 4(b), which brings the usual obligations: human oversight, logging and informing workers' representatives before use.
- Should employees be told their own risk score?
- There is no single right answer. Decide it deliberately, as a policy on whether the score is an internal workforce planning signal or something an employee can see about themselves, with legal and, where applicable, works council input before launch, rather than leaving it to default.
How to cite this page
Blits.ai AI Use Case Library, "AI for employee attrition prediction and retention analytics", last verified 29 September 2026, https://www.blits.ai/ai-use-cases/employee-attrition-prediction-and-retention-analytics. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 29 September 2026: First published