When a Warehouse Drone Finds an Empty Bay

Inventory drones are already in use

In August 2024, IKEA’s parent group, Ingka, reported that more than 250 drones were operating across 73 locations in nine countries. It described drones using AI to identify and photograph storage locations, with an upgraded system tested in Belgium that could work alongside staff around the clock. Those figures describe the deployment at the time of the announcement.

Maersk described a similar application in January 2023. Its inventory drones scanned barcodes and photographed pallets. The findings were compared with the warehouse management system, and discrepancies were presented to staff through a dashboard. That account provides a useful view of the workflow: observation, comparison and information for someone to act on.

The question becomes more complicated when an AI assistant can propose—or make—the resulting stock correction.

The pallet has moved

Consider a hypothetical warehouse.

The stock system places a pallet in bay B17. A drone photographs B17 and finds it empty. An AI assistant reviews the image and proposes marking the pallet as missing.

A forklift has already moved it to C08. The movement has been recorded, but that update has yet to reach the inventory snapshot the assistant is using.

The image accurately shows an empty bay. The proposed stock correction would still be wrong.

Marking the pallet missing could remove available goods from sale, disrupt an order or send staff looking for stock that is already accounted for elsewhere. The problem sits in the relationship between the observation, the records available at that moment and the change being proposed.

A useful review would ask when the image was taken, which inventory snapshot was checked and whether any movements occurred between the two.

Give the assistant a precise job

An operator could authorise an assistant to flag a location discrepancy and request a recount. Changing the recorded quantity or removing stock from sale would require separate authority.

That gives the assistant a useful job while keeping the consequences of each action clear.

In the B17 example, it could raise a task asking someone to confirm the pallet’s current location. If a reliable movement record identifies C08, it could propose a location correction. If the image and movement records conflict, the proposal would go to an authorised reviewer.

The reviewer needs to see the exact proposed change. “Resolve the inventory discrepancy” leaves too much open. “Change pallet P1042’s recorded location from B17 to C08, leaving its quantity unchanged” is something a person can examine and approve.

The same precision matters if software carries out the update.

Keep enough evidence to explain the correction

A record saying “inventory adjusted” would leave several questions unanswered.

For this workflow, the evidence should connect:

• the image, its capture time and the location scanned;
• the pallet identity and the assistant’s interpretation;
• the inventory snapshot and relevant movement records;
• the exact change proposed and the authority applied;
• the update accepted by the warehouse system.

If a later check finds the pallet elsewhere, the original interpretation should remain available alongside the correction. That allows reviewers to understand how the mistake arose.

This is a specific place to examine PF Core’s role. A proposed trial could use it to preserve linked material from the observation through to the recorded decision, so a reviewer can follow the stock correction without piecing together several unrelated logs.

The warehouse system’s acknowledgement would still be needed to establish which update it accepted. A permission decision alone would leave that question open.

PF Kernel could also be evaluated against proposed amendments: whether the assistant may raise a recount, correct a location or refer a quantity change for approval. Any later enforcement would depend on the warehouse integration applying that decision to the actual update.

Test the awkward cases

A first trial could use historical or synthetic discrepancies and copies of proposed corrections, leaving the existing stock process in charge.

The test set should include a genuinely missing pallet, a recently moved pallet, an unreadable label and the same correction proposed twice.

For each case, the warehouse team could examine whether the proposed change was justified, which information was missing and whether someone could reconstruct the decision afterwards. It could also count how often a person needed to intervene and why.

That would give the business a basis for deciding where automation is useful and where the workflow needs more work.

The IKEA and Maersk examples show inventory drones being used for a defined operational task. For businesses considering further automation, the follow-up deserves equally careful attention.

Before a stock record changes, the organisation should be able to explain what was observed, why the change was authorised and what its system actually accepted.