PRACTICAL RESOURCE
What retailers can contribute to AI training and evaluation
A retailer can demonstrate how staff restock a shelf, show how an inventory change is recorded or write a task that tests how a model resolves an exception.
A retailer can demonstrate how staff restock a shelf, show how an inventory change is recorded or write a task that tests how a model resolves an exception. Each route uses different evidence. Start with one job a reviewer can verify and clear the sources before sharing anything.
Match the sample to the evidence
| Public source | What it establishes | What to ask |
|---|---|---|
| Scale lists retail and consumer businesses in Data Partnerships. | A commercial intake for real operating records under an approved scope. | Which systems, records and rights does the current project require? |
| Troveo describes first-person capture and separately markets business data licensing. | Public video and operating-data offers. | Does the brief call for demonstration clips, linked records or both? |
| Dexset advertises shelf scanning, stocking, inventory and product recognition. Graviti publishes a specification that includes retail back-of-house footage. | Vendor-stated retail capture capabilities and dataset formats. | Which labels, views and variations are available? What can an independent reviewer verify? |
| OpenAI’s GDPval covers retail occupations, including first-line retail sales supervisors. | A public example of expert-authored work tasks and grading. | Can a retail expert write and score a realistic original exception case? |
Scale Data Partnerships · Troveo first-person video guide · Troveo business data · Dexset retail data · Graviti skilled work dataset · OpenAI GDPval
Try one closed-room stocking case
Use a tabletop shelf, two mock cases and original product labels. The approved stock card says case A contains six units of SKU A. Case B looks similar, but the unit count is missing. Film the label check and placement of two units from case A. Make a second take with the label hidden.
Score the product identity, visible count and placement times independently. Compare those observations with a separate fictional inventory event that records two units. The video supports what happened in the frame. The record shows what was entered. Keep the unknown case B quantity unresolved. If a buyer wants real records or more footage, agree on the required units, labels, edge cases, rights and review process first. GS1’s packaging levels give a public reference for representing the relationship between a case and the items it contains. GS1 US packaging levels
Clear the people, products and records
Use an authorized site and consenting adult staff. Keep shoppers, receipts, labels with personal details, supplier-only data, financial records and store security out of the pilot. Confirm product artwork, store footage and employee recording rights separately. Before supplying real business records, verify that the contract permits the proposed training, evaluation, redistribution and retention.
Try home center bundle picking, warehouse club package levels or pharmacy package tracing. Download the Retail Workflow Sample Worksheet · Describe an original retail workflow
Keep observed movement separate from inventory entry
Extend the practice shelf case with a video showing two units placed and a ledger entry recording three. The reference answer should report the disagreement with both sources. It should not silently change the observed count to match the ledger or treat a visible placement as proof of a completed system update.
For a missing-label take, preserve the unknown product identity even when the nearby stock card suggests an expected SKU. Hold out a new package design to test that the model can read the evidence rather than memorize the original prop. Use the worksheet to connect take, item, package level, event and reviewer decision before discussing a real collection.