Why warehouse records drift away from the floor

01 / FIELD NOTE

Keep the decision tied to the operating context.

Ask a warehouse team whether the system is right and the answer is usually a qualification: right about most things, most of the time. That qualification is the whole problem. A stock figure is used as though it were a fact, but it is an accumulation of events, and every event that was not recorded — or was recorded twice — moves it further from the floor.

Drift has two directions and they have different causes. Stock goes missing from the record when a movement happens and nobody posts it: goods moved to make room, stock pulled for a sample, a pallet relocated to seasonal overflow. Stock appears in the record when a post happens without a movement: a receipt entered from paperwork that never arrived, a transfer created to clear a discrepancy, a return credited before the goods were checked in.

Manual entry is the most visible contributor and the least interesting one. The larger effect is latency. When the record is updated at the end of a shift rather than at the point of movement, everything that happened in between was transacted against a figure already known to be stale. Two pickers working the same aisle from the same wrong number produce two wrong picks, and the error multiplies rather than adds.

Line-of-sight capture has a structural weakness that is easy to underrate. A barcode is read one item at a time at a moment someone chooses, so the record only learns about what somebody decided to scan. The pallet that was counted at case level, the carton whose label faced the wrong way, the box that was never opened — all of them are invisible to a method that has to be aimed. This is not carelessness; it is what the technique can do.

Misplacement is the failure that most often gets misread as loss. Goods are on the premises, in the wrong place, and a quantity-based system cannot tell the difference between absent and elsewhere. It reports a shortage, the business reorders, and a location error has been converted into a purchasing cost. Then the original stock is found, and now there is a surplus that no one explains.

The chain of custody breaks at every handover, and there are more handovers than the process diagram suggests. Inbound to put-away, put-away to storage, storage to picking, picking to packing, packing to dispatch. Each transition is a chance for the record to stop following the goods. A system with five handovers and a high success rate at each still loses a meaningful share of its movements to the compounding.

The consequence is not only a wrong number. It is a set of compensating behaviours that make the operation slower and less predictable. Safety stock is raised to cover the uncertainty. Staff double-check the system against the shelf before trusting it. Supervisors keep a parallel count on paper for the categories they care about. Each of those is a rational response to an unreliable record, and each one adds cost that never appears as an inventory line.

Reducing drift is a matter of moving the capture to the moment of movement and removing the human decision about whether to record it. Capture that happens as a by-product of the work — goods passing a point, a container being handled — does not depend on anyone remembering, which is what makes it different in kind from a scanning discipline rather than merely better at it.

Item-level identity changes what can be recovered after the fact. If the record says a product has a quantity of ten, then finding nine tells you only that something is wrong. If it says which ten units, then finding nine identifies the missing unit, its last known location and its last recorded movement, which is the difference between an investigation and a search.

Frequency does more for accuracy than thoroughness. A complete count is disruptive enough that it happens rarely, so the record is corrected rarely and drifts badly in between. A frequent count of a small area fits into ordinary work and corrects drift while it is still small enough to trace. This is the practical case for counting continuously in pieces rather than perfectly at intervals.

Exceptions are where the diagnosis lives, and they are usually discarded. A count that reports a single variance figure says you were wrong. A count that separates expected-and-not-found from found-but-unexpected, and both from found-in-the-wrong-place, tells you whether the problem is a movement that was never posted, a receipt that was never made, or a location that was never updated. Those need different fixes, and averaging them together hides which one you have.

The last requirement is ownership. A record with no one responsible for its exceptions decays, because correcting a discrepancy is always less urgent than the work in front of you. Assigning the exception queue to a named role, even a shared one, and making the resolution visible is what keeps the figure trustworthy after the project that produced it has ended and attention has moved on.