Logistics becomes autonomous and predictive.

Demand forecasting estimates demand over a longer horizon from historical sales, seasonality and trend data. Demand sensing corrects that estimate in the short term using signals arriving now, such as open orders, customer inventory positions, weather and promotions. Forecasting drives what you buy and where you position stock across the network. Sensing drives the weekly or daily correction. Most supply chains run both, because a long-horizon forecast cannot react fast enough and a short-horizon signal cannot plan capacity.

Start with a decision that repeats daily and whose outcome you already record, because that record is your training data. Route planning, slotting and demand forecasting usually qualify. Limit the first project to a single depot, lane or product group rather than the whole network. Run the model in shadow mode alongside the existing planning process for a full demand cycle, including a peak period, and let it steer only once its results are consistently better than the current method.

Route optimization needs what was actually driven, not only what was planned: real departure, arrival and dwell times for every stop. You also need geocoded addresses, delivery windows, vehicle capacity in both weight and volume, realistic loading and unloading times per customer type, and driver working-time rules. Without actual outcomes, the model mainly reproduces the assumptions behind the old plan. In most projects, cleaning up this master data takes longer than building and tuning the model itself.

Always measure against a baseline, which is your current planning or forecasting method. A result without that comparison proves little. Agree on the metric before you start, for example forecast error per item, cost per stop, load utilization or on-time delivery, and hold back a period the model has never seen. Then measure in live operations rather than in the demo. A model that scores better statistically but that planners override every morning delivers no benefit.

The most common mistake is starting with the model instead of the master data and the planner. Wrong package dimensions or outdated delivery windows produce a system that calculates flawlessly on the wrong assumptions. A second mistake is optimizing a single metric: minimizing kilometers alone tends to cost service reliability. A third is treating the planner as an obstacle. If people cannot see why the system proposes a route, they override it and the gain disappears.