ATTRITION RISK INTELLIGENCE · SAP

The exit interview already knows who's leaving.

Most HR platforms report backward: who left, when, by department. WorkforceAI reads five ordinary signals as one story, months before the resignation letter, using explainable risk scoring built on SAP RPT-1-OSS.

Try Demo → Read SAP Community Blog
EMP04567 · Marketing · read one field at a time
months_since_promotion
avg
compa_ratio
near band
recognition_count_12m
low side
internal_job_applications
one
manager_change_count_24m
a few
On their own: nothing looks wrong. Read as one row, cross-column, cross-row, against everyone like her:
RISK BAND · CRITICAL
THE PROBLEM

A retention blind spot, hiding in plain sight

While working with SAP SuccessFactors and HR operations teams, the same pattern kept turning up, especially around review cycles. A top performer would give notice, and the team would be genuinely surprised. Not because the data was missing. Because it was scattered across five quiet columns that nobody had ever read as one story.

"It wasn't about the money. It was growth, mental peace."

That's what one departing critical resource said over a chai break at The Rameshwaram Cafe in Bangalore. Human psychology is genuinely hard to predict, but the signs were already sitting quietly in the fields nobody was reading together: a promotion cycle that kept slipping, a compa-ratio a little under band, recognition that stayed flat while performance kept climbing, an internal transfer she'd already tried and didn't get, and a manager relationship that kept resetting.

Not a people failure, exactly. Just a lot of rows, and not much time to read five fields as one story.

THE SIGNALS

Five fields. One story, read together.

These are the fields that were already moving before the resignation letter, if anyone had the time to connect them.

01
months_since_promotion
Climbing quietly, cycle after cycle. Easy to miss because it never spikes, it just never resets.
02
compa_ratio
A little under band. Easy to explain away on its own, it's the pattern next to it that matters.
03
recognition_count_12m
Low, even while performance stays high. Outperforming and unrecognized, at the same time.
04
internal_job_applications
She'd already tried to grow here first. An internal move attempt is a retention signal most systems don't score.
05
manager_change_count_24m
Too many changes. Nobody consistently in her corner long enough to notice the other four.
HOW IT WORKS

Reads job_title as meaning, not a label

Built after going through SAP's research on ConTextTab, a semantics-aware tabular in-context learner. A Sales Account Executive with a stalled promotion isn't the same case as an Engineer with the same numbers. The model is meant to know that, because it learned from millions of real tables, not synthetic ones.

Default engine
SAP RPT-1-OSS
In-context learning: shown past cases at inference time, no training step. Reads rows two ways at once, her whole profile, and how she compares to others like her, via cross-row / cross-column attention. WORKFORCEAI_PREDICTION_ENGINE=sap_rpt_oss
Fallback engine
LightGBM
Trades semantic meaning for structure: one-hot encodes roles as arbitrary IDs. In exchange, it's the only engine wired to full SHAP TreeExplainer output today. WORKFORCEAI_PREDICTION_ENGINE=lightgbm
  • ✓ Every score gets a risk band: critical, high, medium, or low, with explainable thresholds, not a black-box number.
  • ✓ Recommendations run off the explanation: compensation, promotion, engagement, and workload playbooks are wired today.
  • – Recognition, internal mobility, and manager-churn playbooks aren't wired yet; those cases fall back to "monitor quarterly."
  • – The honest ceiling: at real scale, tuned trees still win, the underlying research says so. The semantic engine's edge is on data this small and this messy, which is exactly what most HR teams actually have.
TRY IT

Load 800 synthetic employees. No upload required.

The workspace ships with a demo dataset so you can see the risk bands, explanations, and retention actions before connecting a real SuccessFactors extract.

1. Connect data
2. Score risk
3. Explain drivers
4. Act on retention
Try Demo →