A staff member photographs the shelf on their phone. The system counts it, compares against the system quantity, and flags anything it isn't sure about — instead of guessing.
Illustrative demo — the shelf is a drawing and the counts are scripted to show system behaviour: per-item confidence, comparison against the stock system, and low-confidence items routed to a person rather than guessed. On a real build we measure accuracy against a labelled set of the client's own photos, and we quote that measured number — never an assumed one.