How AI Transforms DTC Supply‑Chain Replenishment: From Safety‑Stock Theory to Real‑World Deployment

Algorithm Design

Supply‑chain management has become critical as e‑commerce shifts from growth to a saturated market. Traditional safety‑stock theory assumes constant lead time and normally distributed demand, but DTC scenarios often violate these assumptions. By modeling promotion and non‑promotion periods separately and using the joint error distribution of forecast residuals, we compute safety stock that meets a target service level.

We define three core modules:

Prediction model : a fused model handling both promotion and non‑promotion periods.

Decision model : a safety‑stock calculation based on the joint error distribution, which can be replaced by operations‑research or reinforcement‑learning approaches.

Simulation model : a business‑logic driven simulator that ingests predictions, decisions, and historical data to evaluate KPI impact.

Simulation results show a >20% improvement in inventory days and stock‑out rates compared with manual replenishment.

Preparation Before Deployment

We combine MLOps with simulation to ensure model stability. Continuous monitoring of inventory days and stock‑out rates detects performance degradation, prompting retraining or data adjustments. Historical stability analysis across different periods confirms that the same hyper‑parameters yield consistent improvements.

For high‑impact promotions (e.g., 88 Day), we perform targeted hyper‑parameter searches, achieving a 4% reduction in stock‑out rate without significantly affecting inventory days.

Deployment Challenges

During pilot runs, low adoption of AI‑recommended replenishment was traced to over‑estimation of sales and lack of trust. By adding fields for AI‑recommended order points and maximum inventory, we reduced uncertainty and increased adoption by over 60%.

We also built an attribution analysis pipeline to classify reasons for non‑adoption, revealing issues such as untracked promotions and site‑driven over‑stocking.

Insufficient and Optimization Points

Our experiments reduced inventory days by 14.8% and stock‑out rates by 6.9%, but further work is needed to translate KPI gains into financial statements, consider logistics costs, and ensure long‑term model robustness.

Integration of Prediction, Decision and Simulation

Algorithm engineers can iterate from simple LightGBM models to transformers, adjust decision logic, and run automated simulation back‑tests before deployment. Product teams use the integrated platform to set service‑level targets, simulate cost impacts, and guide business users in adjusting replenishment plans.