Warehouse AI works when it is pointed at a pick path that someone has already measured. In most manual picking operations, walking is the largest single part of a picker's day, so route, slotting and batching decisions move the numbers before any model does. A warehouse that cannot yet describe its own pick path in data is not ready for AI. It is ready for a tape measure and a clean location master.
What does walking the floor tell you?
It tells you what the system believes versus what happens. Follow a picker through ten orders. Count the steps, note every time they backtrack, wait or search, and compare the route to what the warehouse management system suggested. The gap between the two is the first finding, and it usually has nothing to do with software.
Four things to fix before any model
- Locations. Every storage position has a unique, scannable code, and the system knows where each location physically is. A route cannot be optimised on locations that exist only as labels.
- Slotting. Fast-moving items sit close to the dispatch point and in the pick faces that are easiest to reach. Slotting is the cheapest lever in the building, and it is judgement, not machine learning.
- Scan discipline. Every move, pick and put-away is scanned at the moment it happens. A model trained on timestamps that were entered at the end of the shift learns the shift, not the work.
- Order structure. Know whether you are picking single orders, batches, zones or waves, and why. The choice changes what an optimiser can usefully do.
Where does AI then help?
Once the path is measured, three uses are worth testing: sequencing the route inside a batch, re-slotting when demand shifts, and forecasting the labour a given day's orders will need. Each has a number you can check against the baseline you took on the floor.
How to measure it
Measure hours and steps per order line, not model accuracy. Take the baseline before the change, run the change on one zone or shift, and compare. If the hours saved cannot be named, and the team cannot say what they now do with that time, the project has produced a demonstration, not an improvement.
A test before you start
Can you reproduce yesterday's pick path for one order from your own data? If the answer is no, start there. Everything an AI system will do depends on it.