Predicting flight risk. Prescribing retention.
A machine-learning model that finds front-line associates at risk of leaving, then hands their manager a specific, ready-to-use retention action inside the tools they already use.
A revolving door at the front door
Front-line customer care is the front door of a health plan. It is also one of the hardest jobs to keep staffed: first-year attrition ran as high as 90% in some markets, driving an estimated $200M a year in recruiting, training, and upskilling, and putting member service and experience at risk.
Leadership wanted to stabilize the talent base and reduce reliance on reductions in force. The ask was not another report. It was a way to see risk early enough to act, and to act at the level of the individual associate and their manager.
Predict early, prescribe clearly, deliver in-flow
Predict, don’t react
An XGBoost model trained on 18-to-6-month history scores flight risk for recent hires with at least one survey response, before they quietly disengage.
Auditable by design
Protected attributes are excluded and parity is monitored, so HR and Legal can trust and defend every score. Nothing is a black box.
Prescribe the next step
Embedded GenAI turns each score into a specific, manager-ready retention action. Not just a number, but a move.
Enable the manager
Training, playbooks, and coaching helped managers act on the insight, delivered inside Workday and Tableau with automated notifications.
From survey signal to manager action
1Data sources (Workday RaaS + systems)
Signals flow in from Workday reporting and operational systems: onboarding and 30/90-day surveys, terminations, scheduling, call-quality, and member feedback.
2AWS data lake (S3)
Signals are ingested into a secure S3 data lake and joined with Workday data into a governed feature store.
3Model & ML ops
An XGBoost model trained on 18-to-6-month history scores flight risk, with parameter tuning and continuous bias monitoring. Inference runs on recent hires with at least one survey response.
4Back into Workday (Prism)
Predicted leavers are written to a Workday Prism dataset and surfaced as a security-aware manager report, with automated Workday notifications.
5Manager insight & action
Each manager sees risk scores, trends, and a GenAI-recommended action for their own team, inside the tools they already use.
6Enablement & change management
Managers did not act alone. Targeted training, action playbooks, and hands-on coaching helped them use the insights well and hold the retention conversations that matter. This enablement ran through the pilot and was built into the scaled rollout, so adoption held as the program grew.
What leaders and managers saw
Executives triaged risk by market; managers received a ranked action list for their own team. The view below is an interactive reconstruction. Explore it.
How a risk score is built
Every score is explainable. Pick an associate to see the signals driving their risk, and the action the model recommends.
Attrition bent down in ten weeks
In the pilot markets, monthly attrition fell about 25%, from 19.5% to 14.7% (a 4.8-point drop), against a matched control, saving an estimated $1.8M during the pilot and a projected $9.6M in year one. Manager enablement and change-management support were a deliberate part of the design, and were carried into the scaled rollout to keep adoption strong.
Whether you’re modernizing people analytics, deploying AI responsibly at scale, or sharpening product and market strategy, I’d be glad to talk.
The dashboard, associates, and figures shown here are an illustrative reconstruction built from synthetic data. This work was completed while employed at a Fortune 25 health insurer; the original model, data, and production dashboards remain the property of the company.