The scanner at the gate flashes green on the second pass even though the trailer plate already sits in the system under a prior division code. The new record lands in a separate bucket because the batch prefix rule changed after the last vendor patch.
Procurement teams keep buying master-data platforms on the strength of lab deduplication percentages that never see live transaction variance. The tools match on static fields while every real consignment carries timestamp drift, site-specific codes and partial OCR output.
Downstream AI models then train on the split histories. Forecast engines treat the same supplier as two entities and generate contradictory lead-time signals that procurement overrides by hand each week.
Measurement dashboards reward the drop in total record count after each quarterly cleanse. They never track whether the surviving keys actually improve match rates inside the planning or invoicing systems that consume the data.
The real cost appears in exception queues that keep growing. Finance staff spend hours reconciling the same vendor against two supplier numbers because the original contract used one division code and the current shipment uses another.
Fix the loop by tying data-team incentives to downstream query accuracy rather than record reduction. Run weekly samples of AI output against actual supplier performance and penalise any model that still sees duplicate entities.
Stop accepting vendor claims based on curated test sets. Demand proof on a rolling two-week slice of production feeds that includes the address formats, batch codes and timestamp quirks your operation actually generates.