Why ChatGPT and other general AI models can't replace shelf recognition

General multimodal models such as ChatGPT, Gemini or Claude can tell you a photo shows a shelf of soft drinks. They cannot reliably tell a 0.33 L can from a 0.5 L can, or the cherry variant from the original, across hundreds of SKUs — and that difference is the whole job. They also run in the cloud, so they stop working in a store with no signal.

The job is SKU level, not category level

Retail execution metrics — on-shelf availability, share of shelf, facings — are counted per SKU. Variants of one brand often differ only by a colour band, a flavour name in small print or a few millimetres of pack height. A model trained on the open internet has never seen most of these packs, and it has no way to learn the difference between your SKU 1042 and SKU 1043.

Counting has to be exact

A share-of-shelf number is only useful if the facings count is right. General models describe images; they are not built to return a consistent count of every pack on every shelf row. Two runs on the same photo can give two different answers.

Stores don't always have a signal

Basements, cold rooms, rural outlets and traditional trade in emerging markets often have weak or no mobile data. A cloud model needs the photo uploaded before it can answer. A model that runs on the device answers during the visit, while the merchandiser can still fix the shelf.

Cost at field scale

A field team taking thousands of shelf photos a day pays for every image sent to a cloud model. On-device recognition has no per-image cost, which is why it can be sold at a flat fee.

Where general AI does help

General models are useful around recognition, not instead of it: drafting visit reports, summarising audit results for managers, or answering questions about the data a recognition system has already produced.