prepared by the NADALI editorial team using AI
A customer selects a large green bag, sees a photograph of a small one, and the warehouse receives the SKU of a black one. Every individual field can look plausible. The error sits between them: colour, price, photograph and inventory belong to different variants.
For a small shop, a useful AI assignment is to explain these conflicts and prepare checks for a reviewer. This editorial workflow covers ordinary physical products sold in several variants. It is not a software test or a guide for regulated goods, used items or complex bundles. Start with one product family whose individual items you can verify.
1. Define what the customer buys
Separate three levels. A parent record groups a product family, such as one bag design. A variant represents the particular combination ordered, such as black and large. Inventory describes the available quantity of that variant at a particular location and time.
This is a working audit model, not a requirement for every platform’s architecture. Establish how your system represents those levels. If a parent only groups variants, do not count it as another sellable unit. Two inventory rows for the same variant at different warehouses do not represent two different products either.
Record the unit of sale: one bag, a pack of two or another configuration. Personalisation or a service option may work differently from a physical size. Do not turn every website selector into a separate stock item without an agreed rule.
2. Separate internal SKUs from GTINs
Here, SKU means the internal reference used to manage a variant. Preserve the existing SKU and system ID. If a reference is hard to read, add a clear label alongside it rather than renaming it during the audit. Orders and integrations may depend on it.
A GTIN identifies a trade item within the GS1 system; a barcode enables machine reading of encoded information. An internal SKU or barcode image is not a substitute for a verified GTIN. GS1 explains the number/barcode distinction.
GS1’s guidance assigns separate GTINs to distinct size-and-colour combinations. This concerns the corresponding products, not automatic numbering of every export row. GS1 guidance on variants.
Avoid a blanket rule that every listing edit requires a new GTIN. The GTIN Management Standard distinguishes new products from changes to existing products. Refer uncertain cases to the person responsible for identification under the applicable rules. A language model cannot decide from a title alone. For a supplier’s product, request confirmation of the appropriate identifier rather than borrowing one from a neighbouring variant.
A correct check digit checks structure, not correct allocation. GS1 Denmark explicitly states this limitation. Keep a missing identifier unresolved instead of generating plausible digits.
3. Build the allowed-combination matrix
Ask the assortment owner which variants actually exist. Two colours and two sizes produce four mathematical combinations, but the supplier may manufacture only three. AI must not create the fourth to make the grid symmetrical.
Give each combination a state: offered, temporarily out of stock, discontinued or never part of the range. The last two are different from zero inventory. Use a specification, approved range list or confirmation from the responsible person as evidence, rather than the mere absence of a row.
Agree a vocabulary separately. Do “чорний” and “black” mean the same colour? Does “large” map to L for this design? Preserve original values. Replacing every M with “medium” can be wrong when another category uses M differently.
4. Check relationships before asking for explanations
Retain a working snapshot with its date, filter and source list. Include parent ID, variant ID, SKU, options, sales unit, GTIN or verification status, image ID, price with currency and conditions, warehouse, available quantity and inventory timestamp. Use only approved data and processing destinations. Customer orders are usually unnecessary for this audit.
Use ordinary filters or system checks to identify:
- variants pointing to a missing parent;
- one SKU assigned to different combinations within the agreed inventory system;
- two active records for the same allowed combination;
- images, prices or inventory pointing to nonexistent IDs;
- prices without currencies and inventory without location or time.
A repeated GTIN across warehouse rows for one product is not automatically an error. The same number on different confirmed variants warrants investigation. Compare entities, not just repeated text.
Different prices for different sizes may be correct. Compare them with the approved price list using the same currency, tax basis and effective period. Do not “correct” them towards an average. Distinguish physical stock from availability after reservations; do not add quantities with different meanings.
5. Worked illustration: three bag variants
All records below are fictional. Family P-20 has three permitted variants: small black, large black and small green. The supplier does not make large green.
- V-201, SKU BAG-01: small black, image IMG-BS, UAH 600, available quantity 5.
- V-202, SKU BAG-02: large black, image IMG-GS, UAH 700, available quantity unknown.
- V-203, SKU BAG-03: small green, image IMG-GS, UAH 600, available quantity 0.
The approved image register identifies IMG-BS as small black and IMG-GS as small green. Prices share the same tax basis. Known inventory figures concern one warehouse and timestamp.
The expected findings are three existing variants, an image conflict for V-202, unknown inventory for V-202 and a known zero for V-203. Missing large green is not a defect. Known available quantities sum to 5, but that is not a confirmed complete family total. Replacing the unknown with zero hides the gap.
Correcting the image requires a verified photograph of the large black bag. Automatically substituting IMG-BS is inappropriate: matching colour does not establish matching size. GTIN numbers are deliberately not invented for this example.
6. Give AI a bounded assignment
Use this prompt with your approved snapshot:
Audit relationships between parents, variants, the image register, price list and inventory. Treat the allowed-combination matrix as a separate source. For each discrepancy return the variant ID, rule, exact source values, possible consequence and question for the responsible reviewer. Distinguish confirmed conflicts from missing evidence. Do not change data, merge records, create SKUs or GTINs, or fill blanks. Do not claim to have inspected an image when you only have its filename or URL.
Have a reviewer check every finding against its source. AI can explain inconsistent labels, but confidence does not verify colour, contents or identifier allocation. Manual checking may be simpler for a small family.
7. Accept the audit against explicit conditions
The audit is complete when every active variant has an explained place in the matrix, unchanged IDs, checked relationships and an owner for unresolved questions. Record the old and proposed values, evidence and reviewer for each approved correction. The report itself is not permission to modify the shop.
After separately authorised changes, check the storefront selector and a test cart without payment: image, options, price and SKU must describe the same item. Verify the warehouse mapping too. Do not create real orders without permission. Escalate an unresolved error that could cause incorrect fulfilment before the next sale.
Track confirmed conflicts, false alarms and open questions. Fewer rows or tidier names are not success criteria. A useful result lets staff explain exactly what the customer will receive after every permitted selection.
Sources checked on 7 November 2026. The workflow and example are editorial proposals; no product tests or measurements of time savings were performed.
Cover: an original conceptual editorial illustration generated with OpenAI for Nadali. It is not a documentary photograph, product interface or evidence of testing.
