Every business holding stock knows the gap between what the system says is on the shelf and what is on the shelf. Items get moved, miscounted, damaged, returned to the wrong place or taken without being recorded, and each event is small. Accumulated over months, the records and the reality separate, and the business finds out by promising a customer something it does not have.
The traditional answer is the stock-take: close, count everything by hand, reconcile. It is accurate on the day and wrong again within weeks, because it is a snapshot of something that changes continuously. It is also expensive, disruptive and disliked, which is why it happens rarely - and rarity is precisely what makes it ineffective, since the value of a count decays from the moment it is finished.
The machines being deployed against this do something narrower than most people expect. They do not pick, lift or restock. They travel the aisles, usually at night, photographing and scanning shelves, and compare what they see against what the system believes. Some are ground robots, some are fixed cameras, some are drones inside high-bay warehouses reading labels at heights that otherwise require a lift and a person. The output is not a tidy shelf; it is a list of discrepancies for someone to act on in the morning.
That framing is what makes it work. A count nobody can perform more than twice a year becomes something that happens every night, and the errors are caught while they are still small and while it is still possible to work out what happened. Detecting a gap the day it appears is a different problem from detecting it six months later, and the second one is usually unsolvable.
The limits are the familiar ones. This suits places with long uniform aisles, consistent labelling and enough stock to justify the cost. It struggles where items are irregular, stacked, stored loosely or identified by anything a camera cannot read, and it does not remove the need for people - it redirects them from counting, which machines do better, to investigating and correcting, which they do not do at all. The pattern is the same as elsewhere in this field: the machine takes the measuring, and the judgement stays where it was.