What changed
Three things: data availability (ERPs now capture granular transactional history), compute (training a per-SKU model on 3 years of data takes minutes), and model maturity (gradient boosted trees and LSTMs are now plug-and-play for time series).
What the model actually uses
- Sales history at SKU × location × week granularity.
- Promotional calendar: price changes, discounts, campaigns.
- Seasonality: true seasonality (monthly patterns) + holiday effects.
- Lead time: supplier responsiveness, truck schedules, port delays.
- External signals where relevant: weather, macro indices, category trends.
Where accuracy breaks down
No model can predict a pandemic, a viral TikTok, or a trade war. What a good forecast does is give you a confidence interval — 'we expect 1,000 units ± 150' — so you can plan inventory against the risk. The best teams use the model's confidence bands to set reorder points dynamically, not just the point estimate.
Before / after
A typical mid-market retailer before AI forecasting: 76% inventory accuracy, 15% stockout rate, 12% deadstock. After 90 days with an AI forecasting agent: 98% accuracy, 3% stockout, 5% deadstock. Working capital improves 20–30%.
See it in action
Want to walk through this with an engineer?
45 minutes, live on your ERP, no slides.