Denominator integrity and scorecard interpretation

Diagnosis begins with a metric definition, not the color of a management dashboard. Two dashboards can show different “on-time delivery” rates and each be arithmetically correct: one uses dispatch and another customer receipt; one excludes orders canceled before the promise and another includes them; one counts orders and another order lines. Before comparing months, write the numerator, denominator, event, exclusion rule and data source. This is consequential rather than ceremonial. Ninety successful cases become 90% with 100 orders but 75% with 120 lines; choosing the flattering denominator changes the question being answered.

SCOR distinguishes reliability, or meeting the commitment, from responsiveness, or speed of execution, as well as cost and assets. ASCM's perfect-order definition requires more than on-time arrival: product, quantity, customer, place, condition and documentation must also be correct, with all relevant lines satisfying their conditions for an order to count. Suppose 95 of 100 orders arrive on time, but five of those 95 are short and three have incorrect documents. On-time arrival is 95%. We cannot assert that perfect-order performance is 87% unless the two defect groups do not overlap and every other condition is satisfied. If two orders have both defects, the most that can qualify from these data is 95−(5+3−2)=89, before checking other defects. Precise metric names prevent an intervention aimed at the wrong outcome.

Cycle time is measured between agreed events for completed orders, with its distribution disclosed. A mean of 2.2 days can hide a long tail: nine orders take one day and the tenth 13 days, giving (9×1+13)/10=2.2 days, while nine customers experienced much faster service. Show a median, threshold rate or relevant percentile when the decision calls for one; do not describe every customer's experience with a mean. Segmentation matters as well. A larger share of easy orders can lower the aggregate mean even if no process improves within either segment. Compare standard and custom orders separately, keeping the segment definition fixed across periods.

Cost tracing must match the decision. Transport cost per shipment is not full cost to serve an order that may split into deliveries or require returns and exception support. Build a model from identifiable, nonduplicated categories: transport, pick and pack, packaging, damage or resend, and exception support. An expedited shipment costs 18 units instead of 10 but prevents an estimated lost-order cost of 5 on average. Its expected incremental cost is 18−10−5=3 units for comparable cases, before validating the loss probability and estimate. A better on-time rate alone does not prove that the extra spend pays for itself.

A missing record is not a zero. If 15 of 200 orders lack proof of receipt, do not call them all late or all on time. Show what is confirmed and conditional bounds. If 170 of the 185 recorded orders arrived on time, confirmed on-time orders are 170/200=85% of all orders. The upper bound, if all 15 missing cases arrived on time, is 185/200=92.5%. The observed-case rate is 170/185≈91.9%. That 7.5-percentage-point uncertainty span can change a carrier decision. Determine whether missing records concentrate in one carrier or region; otherwise a biased observed sample may be mistaken for the whole population.

Do not multiply component success percentages to derive a perfect-order rate unless independence of failures is established, which is seldom a safe assumption. The orders that arrive late may also be more likely to have damage or document errors; multiplying marginal rates can mislead. Keep pass/fail status for each component at the order-line level, then aggregate lines to an order under the stated rule. If an order has three lines and one is short, the order is incomplete even though two thirds of its lines are correct. Compare cost and assets alongside reliability: reducing stock may free cash yet increase shortages, while adding it may cut shortages but raise obsolescence risk. A single measure cannot choose the sponsor's trade-off.