A supplier spreadsheet with hundreds of products is not ready to import just because it opens in Excel. Colour names are inconsistent, image links sit in notes, price units vary and bundles are often mixed with single items. Asking AI to “clean the whole sheet” can create a more polished but less trustworthy file. Its better role is finding patterns, applying approved transformations and isolating rows that need human review.
This workflow turns a supplier file into a structured product-import sheet. ChatGPT or Claude can analyse samples and produce suggestions, while the original values, mapping tables and final checks remain under team control.
Define import-ready before cleaning anything
Create a standard template first: SPU, unique SKU, title, variant_colour, variant_size, cost, retail_price, currency, stock, image_url, description_source, review_status, source_row and change_note. The final two columns make every change traceable. Do not decide the schema in the middle of each supplier file.
Run a data-quality check on a sample
Use 30–50 representative rows, not a full sheet containing contacts and settlement data. Ask AI only to identify duplicate SKUs or titles, inconsistent prices/currencies/stock, mixed attributes, missing images or descriptions, and rows that may confuse bundles with variants. Require row numbers and original values. The output should be a rule list, not a rewritten spreadsheet.
Review this product-sheet sample for data quality only. Do not invent values or modify data. Report inconsistent price, currency, inventory, colour, size and packaging formats; likely duplicate identifiers; missing critical fields; and rows that require manual review. Include row numbers, original values and a proposed transformation rule.
Turn approved decisions into mapping tables
Keep colour, size, packaging and banned-word mappings in a separate sheet. For example, BK, Black and 黑色 can map to a standard Black value; 0.5m, 50CM and 50 centimetres can map to 50 cm. Preserve the raw value in another column. Ask AI for suggested mappings with a confidence level, and route low-confidence rows to manual review instead of forcing a match.
Do not let AI merge SKUs
Size variants can share an SPU but require individual SKUs when customers choose them or stock is deducted separately. A skincare set and the single serum should not be merged because their titles look similar. Use AI only to propose candidate groups, then review them. Do not allow it to generate or replace master SKU values.
Write titles from confirmed facts only
A dependable title combines category, confirmed material or use, and a key specification. Ask AI to use only non-empty confirmed fields and to return a missing-field notice instead of guessing. Keep supplier wording in description_source and AI copy in description_draft. Claims about safety, certification, compatibility, health effects or delivery time need an original or official source.
Test a small batch before the full import
Check SKU uniqueness, required values for ready rows, consistency within each SPU and a random sample against source_row. Import 10–20 SKUs into a draft environment first. Inspect variant selection, inventory, image order and mobile titles before importing the full file. Save the import-file version when it passes.
With a maintained schema and mapping dictionary, AI becomes a controlled assistant for spotting exceptions and preparing candidates—not an untraceable editor of product data.