A pallet jack pauses at the warehouse gate while the handheld scanner records the same crate twice under different batch codes. The master data layer treats both entries as distinct because the deduplication rule only fires after the nightly reconciliation window closes.
Procurement feeds arrive in daily batches that bypass the fuzzy match service entirely. Each row receives an auto-generated key before any existing supplier or asset profile is consulted, so the duplicate sits in the primary table untouched.
AI models trained on these records learn two different demand signals for one physical item. Forecast variance widens and reorder triggers fire twice, draining working capital on phantom stock.
The single-source dashboard still shows 94 percent uniqueness because the metric counts only exact string matches. Any variation in packaging codes or legacy system suffixes defeats the comparison.
Fixing the flow requires forcing every inbound record through the entity resolution engine at ingestion time, not after the fact. That change also means accepting a temporary dip in reported throughput while the backlog clears.
Most contracts still pay vendors on volume of records processed rather than on net unique entities created. The commercial incentive therefore rewards the very duplication the organisation claims to hate.
Run a weekly audit that samples the last seven days of new master records against the full graph instead of the previous day only. The gap between claimed and actual uniqueness becomes visible fast.