Abstract:

Artificial intelligence can identify sales orders at risk of delay, but a risk score alone does not determine which orders deserve scarce operational attention. This paper studies eligibility-constrained, cost-aware triage for SAP Sales and Distribution. A supervised predictor estimates baseline delay probability; a deterministic policy combines that probability with delay loss, assumed intervention effectiveness and review cost. We provide an allocation rule, an error bound, an integration contract and a reproducible simulation. Thirty independent replications use 20,000 training orders, 5,000 calibration orders and five test batches of 10,000 orders each. At 10% review capacity, gradient-boosting gain ranking reduces simulated cost by 15.70% against no review, by 6.07% against probability ranking and by 1.69% against loss-weighted probability ranking. The last comparison has a paired 95% confidence interval of 1.38% to 2.01%. However, misspecified intervention effectiveness reverses this advantage, increasing cost by 5.51% against the loss-weighted baseline. Logistic gain ranking remains competitive, and calibration provides no clear stationary cost improvement. These results support evaluating decision utility and intervention assumptions alongside predictive accuracy. All observations and cost parameters are synthetic; no SAP installation or enterprise dataset was used.