Predictive modelling
Forecasts, anomalies, and scores that beat the baseline you already have.
Custom models on your history: demand, staffing, churn, and outliers that look like fraud. Structured data, time series, and the messy tables in between. A model that does not beat your current method on a holdout does not get a launch slide.
When does a model ship?
Only when it beats the method you already trust on a holdout. A launch slide is not the test.