AI Inventory Counting
to Catch Record Errors Between Counts

Most warehouses count a small set of locations each day or week instead of shutting down for a full count. ABC analysis decides how often each item is counted, with valuable or fast-moving items counted most. It keeps records reasonably accurate, but it depends on people being free to count and checks each location only occasionally, so errors can sit unnoticed for weeks. Counting also competes with picking for the same people, and picking usually wins on a busy day.

With camera inventory counting, locations are watched all day, so a difference between what was moved and what the record says can be flagged as soon as it happens. Counters then check only the locations with a problem, instead of working through a fixed list. The count frequency for each location effectively rises without adding labour or overtime. Scheduled counts can continue for items the cameras cannot see, so nothing falls out of the programme.

Often, yes. P&G's deployment used cameras already installed in its warehouse, with no additional hardware. Coverage is reviewed first: locations a camera cannot see clearly may need a repositioned or additional camera, and very high racking may need a different view. The review shows which zones can be counted from what you already have and which need changes.

Phantom stock appears when the system shows stock that is not physically there, usually because of a mis-pick, a putaway to the wrong location, damage that was not recorded or a transaction posted twice. Pickers sent to an empty location lose time, and orders are delayed. Watching the movements that create these errors lets them be corrected at the source rather than at the next count.

An ongoing stock record depends on accurate transactions as well as a clear record of the stock each observation can cover. Camera evidence can support a comparison of visible movements with the warehouse transaction record where the identifiers, interfaces and timing are agreed. It does not establish every hidden item or unobserved transaction. Review discrepancies and obtain the required approval before adjustments, then track count variance on a comparable basis. Define the correction workflow rather than assuming counts write back automatically.

That depends on your auditors and regulations. Many organisations still perform an annual stocktake, but continuous counting can reduce its scope or support a move to cycle-based audits. Image records for each count give auditors useful audit evidence, because they show what was counted, where and when, without relying on handwritten sheets.

Start where errors cost most: high-velocity SKUs that are picked many times a day, high-value items and locations with a history of variances. These give the fastest improvement in accuracy and the clearest return. Slower items can follow once the process is running and the team trusts the results it produces. Seasonal items can be added before their peak, when errors are most costly.

Send a layout of the warehouse zone with camera positions, recent count variances by location and an export of the items stored there. Note how WMS records are corrected today and who approves adjustments. With that, we can check camera coverage, propose a starting zone and show how differences would be flagged and resolved.