Duplicate Vendor Payments and the Matching Rule
A mid-sized wholesaler runs an accounts payable analytics routine that flags potential duplicate payments by matching on vendor ID, invoice number and gross amount. In FY2024 it flagged 412 records worth $1.9m. The AP manager manually reviewed all 412, cleared 388 as legitimate recurring invoices (same vendor, same amount, different periods) and recovered $61,000 from the remaining 24. Encouraged, the CFO wants to tighten the rule to exact matches on vendor ID, invoice number, amount and invoice date, and to reduce manual review to flagged items only. Last year's external audit found three genuine duplicates that the rule missed because the second entry had been posted with a transposed invoice number and a $0.01 rounding difference. The AP team has no documentation of how the 388 clearances were decided.
1. Evaluate the CFO's proposed change against the data quality dimensions of accuracy, completeness and uniqueness. What would you advise?
Read the worked answer
The proposed tightening improves precision but worsens completeness of detection, and it does nothing for accuracy of the underlying records.
- Uniqueness: matching on vendor + invoice number + amount + date assumes each real invoice appears once with consistent attributes. The audit finding shows this is false: a transposed invoice number and a $0.01 rounding difference are exactly the kind of variation a stricter rule will not catch. Tightening the rule reduces false positives but guarantees the false negatives persist.
- Accuracy: the $0.01 difference and transposition are data entry errors, not business events. The control should tolerate them (fuzzy matching on invoice number, amount tolerance of a few cents) rather than assume them away.
- Completeness: the 388 clearances were undocumented. Without a recorded reason code, there is no evidence the review was effective, and no way to tune the rule using past outcomes.
Advice: do not tighten to exact match. Keep a tolerance-based rule, add fuzzy invoice-number matching (e.g. edit distance of 1-2), and require a documented clearance reason code for every item. Use the historical clearances as a labelled sample to measure the rule's recall before changing thresholds. Reducing review to flagged items only is acceptable if the flag logic has demonstrated recall against known duplicates.
2. The CFO argues the $61,000 recovered justifies the 412-item review. How would you frame the cost-benefit analysis and what additional metric would you ask for?
Read the worked answer
The $61,000 is a gross recovery figure. It should be compared with the full cost of the review: analyst hours x loaded rate, plus the risk cost of the three missed duplicates (which were not quantified). If 412 reviews took, say, 60 hours at $85 loaded, that is $5,100 - a favourable ratio, but the ratio is misleading because the three misses represent an unquantified control failure.
The right framing is a detection-rate question, not a recovery question. Ask for:
- Recall against a known-duplicate test set (the three audit findings plus any others confirmed).
- False positive rate, so the review burden is predictable.
- Cost per dollar recovered, trended over time.
- Value of duplicates not detected, estimated by sampling cleared items and re-testing with a looser rule.
Present this as a precision/recall trade-off curve: the CFO can choose the operating point, but the choice must be explicit and documented, not presented as a free saving.