The Shelf Photo Says Twelve. The Scans Say Eleven. What Should AI Do?

The shelf photograph appears to show twelve folded shirts. The barcode scans total eleven. The inventory system says thirteen.

An AI assistant should not choose the number that looks most plausible. It should create a clear exception, preserve all three observations and help the warehouse team perform a targeted recount.

Images can locate objects, not establish every variant

Google’s Cloud Vision object-localisation documentation describes detecting multiple objects and returning their positions within an image. That kind of capability can support a visual count or identify areas for review.

Folded apparel is a harder operational problem than a demonstration image. Garments overlap, labels are hidden, colours change under lighting and several units can appear as one stack. Object detection alone cannot reliably prove the size, SKU or ownership status of each garment.

This is a proposed cycle-count workflow, not a claim that TRYBUY.IN uses AI warehouse cameras or the Google service referenced.

Make variant identity the foundation

GS1 US explains that apparel variants such as each size and colour can receive unique GTINs represented in barcodes, supporting inventory identification. Whatever approved identification standard a business uses, a count needs to distinguish the actual variant—not simply “grey shirt.”

Keep the product identifier, size, colour, storage location and inventory status separate. Saleable, returned, damaged, held and awaiting inspection units should not be collapsed into one available total.

Men’s shirts, men’s kurtas and boys’ kurtas require distinct locations or unmistakable labels. A similar fold or colour should never move a boys’ garment into an adult count.

Design the photo as an observation

If images are part of a pilot, standardise the camera position, shelf boundaries, lighting and capture time. Include a location marker in the frame. Record who captured the image and whether the shelf was being replenished or picked at that moment.

Use privacy-aware framing that excludes people and unrelated sensitive information. Define a retention period and access controls for warehouse imagery.

Let the model say “cannot count”

Occluded stacks, open cartons and reflective packaging deserve an uncertain result. A forced number creates false precision. The system should mark the region and explain which condition blocked assessment.

Do not read a hidden size from appearance

Even when a visible product resembles a known catalog image, variant identity should come from the approved label or record. A large folded shirt can look like a smaller one; a blue tone can shift under warehouse lamps.

Reconcile three sources in order

Compare the visual estimate, the variant-level scan count and the recorded system quantity. If all agree, sample a portion for manual verification during the pilot. If they differ, generate an exception showing the location, timestamp, identifiers and values.

Then conduct a controlled recount. Check nearby shelves, recent moves, returns awaiting disposition and unprocessed picks. Correct inventory only through the authorised stock-adjustment process, with a reason and reviewer.

The AI can prioritise exceptions by value, booking risk or repeated discrepancy, but the business should decide those rules. It should not silently adjust quantities because a photograph contains one more rectangle.

Test on difficult shelves

Include neat stacks, mixed colours, partially empty shelves, hanging items, open cartons and garments whose barcodes are not visible. Compare results with a carefully verified count.

Report missed units, double counts, wrong locations and variant-identification failures separately. A high total-count score can hide a serious problem if the system consistently confuses M and L.

Measure whether the pilot reduces time to investigate discrepancies without increasing bad adjustments. If simple barcode scanning already solves the location, visual AI may add little value.

Cycle-count questions

Can a shelf image replace barcode or RFID identification?

Not where exact variant identity is required. Use the approved item-identification process.

Should the system correct stock automatically?

Begin with advisory exceptions and authorised human reconciliation.

What is the best first scope?

One stable storage area with clear labels, known boundaries and a manual reference count.

Explore TRYBUY.IN for the product choices customers see. Behind accurate availability is the quieter work of counting every real variant correctly.

Technical sources checked on 26 September 2026.