How Does AI Forecast Supply Chain Demand?

P&G evaluates 20 AI models every day to forecast six weeks of demand for Charmin and Bounty, Jaimie McIntyre Horstman, Director of Demand Forecasting at P&G, told a December 2025 IIoT World panel with Maria Araujo, formerly VP of Innovation Engineering at McCain Foods, and Srinivasan Narayanan, Oracle Solution Delivery Lead at Milwaukee Tool.

Which AI Models Forecast Supply Chain Demand?

P&G and Milwaukee Tool both run multiple AI models in parallel, selecting the one with the least prediction error for each planning window.

At P&G, three internal systems handle different forecasting needs. LDS (long-term demand sensing) runs machine learning across three years of sales history, identifies seasonality and trends, removes outliers, and picks whichever model produces the least error. P&G compared the process to how weather forecasters compare competing models to predict a hurricane’s path, selecting the one with the best track record.

A second tool called Cannibalizer calculates how a promotional event at one retailer pulls volume from competitors. When Costco runs a large promotion, the incremental volume is not all new demand. Some of it shifts from Target, Walmart, and local grocery customers. The model calculates that split automatically using historical data.

For the short term, P&G’s Intelligent Daily Forecaster evaluates 20 models against the next six weeks of demand data every day, selects the lowest-error model, and adjusts the forecast using real-time signals.

Milwaukee Tool uses machine learning ensembles like XGBoost and random forest for structured demand signals, deep learning models like LSTM and temporal convolutional neural networks for volatile and promotional demand, and probabilistic forecasting that outputs confidence ranges instead of a single number.

These models blend historical demand with external signals like promotions, market indexes, and customer behavior. The output is probability bands rather than point estimates. Human judgment still guides the final decisions.

How Can AI Predict Panic Buying?

Consumer panic buying events triggered by COVID, port strikes, and tariff announcements follow no seasonal pattern and have only a few years of data behind them. Personal care products like toilet paper carry a long shelf life, which separates these demand spikes from weather-driven runs on perishable food, where patterns are easier to model.

COVID was the first event that trained consumers to panic buy toilet paper, a product no one wants to run out of. Similar spikes followed a port strike in October 2024, the threat of a second strike in January 2025, and a tariff announcement earlier that year.

Regression models run against the limited panic buying history to isolate which variables correlated most strongly with past demand spikes. Those variables determine the keywords used to scan unstructured data sources in real time. When keyword hit rates rise in specific regions, the system signals a potential spike before orders arrive. A manual review layer still operates alongside the automated monitoring until enough history accumulates for fully autonomous detection.

Price matching between retailers creates a separate forecasting challenge. When Amazon runs a deal, Walmart often matches, and vice versa. Scanning unstructured data to detect these promotions early prevents inventory drain before the demand signal reaches the traditional forecast.

What Data Sources Feed Demand Forecasting?

Demand forecasting depends on three categories of internal data operating within a single model: historical shipment-versus-forecast data, inventory levels, and production throughput. If production rates drop, inventory drains faster than expected and customer orders go unfilled. If sales surge past the forecast, inventory falls before production can ramp. These three signals interact continuously. If any one falls, the forecast fails.

In pharma, cold chain shipments require IoT devices that monitor temperature and location in real time. COVID vaccine shipments at Johnson & Johnson needed to maintain close to minus 40 degrees Fahrenheit through the entire transit, including time spent clearing customs at international airports. Historical IoT monitoring delivered data only after shipment arrival. Switching to real-time sensing made it possible to reroute packages algorithmically when a shipment stalled at an airport or when a temperature excursion occurred.

When AI encounters missing or conflicting data in these environments, explainability matters. The largest risk is an algorithm producing an output based on an assumption the user is not aware of. In regulated industries like pharma and medtech, this is a key conflict. Synthetic data can fill some gaps in model training, but the context behind any workaround still matters.

In food and beverage manufacturing, sensor data from production lines combines with ERP records for orders, inventory, and procurement, plus demand signals from planning systems. Data quality matters more than volume. Clean, timely, and trusted data wins.