Understand the calculation before acting on it.

Descriptive results

Turnover rates use qualifying departures divided by complete outcomes (population minus missing outcomes), over explicit equal-duration observation windows. Aggregated employee-period counts are not unique employee counts. Baseline and post-period observations are not silently pooled.

Missing outcomes are shown as counts and rates, with sensitivity bounds. Duplicate keys, overlapping source periods, invalid dates, impossible counts, and invalid financial inputs are rejected. Small-cell and complementary suppression withholds linked estimates to reduce disclosure risk.

Inferential comparisons

Independent-group retention comparisons report Wilson score intervals and a conservative approximate 95% difference interval from Bonferroni-adjusted marginal score limits. Independence, comparable follow-up, and consistent outcome definitions matter. With sparse observations, estimates can be unstable.

Before/after comparisons are descriptive when participants may repeat. Observational differences do not establish that an intervention caused an outcome. Selection, seasonality, manager effects, and concurrent changes may explain differences.

Documented randomization supports a stronger design but does not remove adherence, attrition, interference or analysis-plan concerns. The existing cluster difference-in-differences implementation is experimental, requires a balanced panel and administrator enablement, and is not a general causal estimator.

AI investment accounting

Gross hours saved = (baseline task minutes − AI task minutes) × task volume / 60. Net hours saved subtract incremental review, correction and rework once. If those increments are already included in AI task time, entering them again double-counts the burden.

Total costs include subscriptions, API use, implementation, integration, training, maintenance and other entered costs. Negative time savings remain negative. Quality adjustment applies a declared quality multiplier to positive net capacity.

Documented cash savings, avoided future spending, redeployed capacity and hypothetical productivity value are alternative benefit categories, never added together. Labor value is not a cash saving without separate evidence. Net benefit = selected benefit − total cost; ROI = net benefit / total cost × 100, and is not estimable when cost is zero. Break-even depends on the stated cost and capacity assumptions.

Simulated organizational scenarios

Role-specific capacity uses hours, part-time fraction, leave, training, onboarding and utilization. Workload uses volume and effort per unit. The simulator uses a recorded random seed, explicit bounded input distributions, and Monte Carlo percentiles. The same validated inputs and seed reproduce the numerical results.

P05–P95 ranges are simulated scenario percentiles, not confidence bounds on real outcomes. Input independence, hiring delays, attrition timing and role substitutability affect interpretation. The product does not produce automated individual termination recommendations.

Validation and limitations

Reference tests and synthetic demonstrations check calculation behavior. They do not establish external predictive validity, clinical validity, prospective workforce outcomes, regulatory compliance, or financial returns for customers. Legacy exploratory risk scores require external calibration before reliance in consequential decisions.

All public demonstration data and outcomes are synthetic. A real pilot needs agreed source definitions, legal/data-access review, an evaluation design, suitable follow-up, missing-data assessment and human interpretation. No scientific validation study or certification is claimed here.