An apparel or consumer brand replacing its product, planning and order systems should use AI on the data conversion, matching duplicate item and customer records and mapping old fields and codes to new ones, and start AI planning only after cutover, once the ledger ties and operational checks have held through month-end closes.
For a COO holding off on AI until the data is settled, part of the conversion is matching: entity matching, deciding which old records describe one style or one customer, and schema matching, deciding which old field and code fills which new one. A University of Mannheim paper at the March 2025 EDBT database conference calls entity matching "a central step in most data integration pipelines", and found both that the best large language models it tested matched records about as well with no or few training examples as older models trained on thousands and that "the prompt needs to be tuned for each model/dataset combination". A separate university-led July 2025 paper on schema matching, accepted at APWeb 2025, tested its method on a benchmark built from real schema pairs in three enterprise domains and reports that it "significantly improves matching accuracy in complex schema matching settings".
In each mock conversion, a test run that converts the old data into the new systems, the AI, whether the integrator's conversion tooling or a model the project team runs, proposes pairs of records it judges to be one style or one customer, a new field or none for each old one, and a new value for each old code such as size scales, colors and seasons, each with a confidence score and the values behind it; the EDBT paper found GPT4 can "generate structured explanations for matching decisions". Before anything merges, owners review every proposal below a confidence threshold they set, plus a random sample of the rest: the product data lead for items and styles, the credit manager for wholesale accounts, the ecommerce lead for DTC customers, and the business owner of each field for its mapping.
Within a tolerance finance sets in advance and where the new systems carry them, the controller ties inventory value by location to the general ledger's inventory balance and open wholesale customer balances to the receivables ledger. The order and inventory leads run operational checks of open orders by customer and units by style, color and size against the old order book and stock report. AI planning tools start after cutover, on sales history converted to the new styles and tied to the old monthly sales, once the ledger ties and operational checks have held through month-end closes.
A line from a brand's matching log might read: records, the same fleece hoodie entered once by product development under its style number and again by the wholesale team under a department store's item code; proposal, one style, with the store's code kept as that customer's cross-reference; evidence, same vendor style number, fabric, size run and first cost; reviewed by, the product data lead; check, units on hand under both codes add up to the merged style's count.
In the ninety days after the new systems' data design is signed off, the project team can run the matching, keep every review in that log, and finish one mock conversion the controller reconciles.
Rejected proposals can show the team why the AI matched wrongly before the next mock conversion: the EDBT paper also found that GPT4 can "automatically identify potential causes of matching errors by analyzing explanations of wrong decisions" and that its error descriptions "can help data engineers to improve entity matching pipelines".
