The same month scored green, amber and red on three supply chain scorecards
Nov 17, 2026, 9 min read

A supply chain scorecard is four things, the measures, the weights, the cascade and the review, and most businesses have deliberately chosen only the first. Weighted achievement indices let a strong result on a cheap measure cover a miss on the one that mattered, which is how one honest month scores green, amber and red on three different scorecards. Targets set as round numbers by people who never looked at how much the measure moves on its own turn ordinary variation into reported performance. The design rule that does most of the work is giving every objective measure a constraint measure, owned by the same person.
The pack has thirty-four measures on it. Twenty-nine are green, four are amber, one is red, and the red one has been red for nine months with the same commentary copied forward each time. The business missed its service commitment to its second largest customer that month and finished the quarter 3 million over its inventory plan. Nobody reading the pack would know either of those things happened, and nobody in the review asked, because the review works through the measures in order and there were thirty-four of them.
A scorecard is a set of measures, a set of weights, a cascade and a review. Most businesses have the measures. The rest of it gets inherited from whatever template the last consultant left behind, and the parts nobody chose deliberately are the parts that decide what the scorecard says.
Five numbers and three verdicts
Take one month, honestly reported, five measures.
Unit fill rate came in at 97.2 percent against a target of 97.0. On time in full came in at 91.4 against 93.0. Forecast accuracy at lag one was 82 percent against 80. Cost per case shipped was 1.41 against 1.45. Inventory cover finished at 9.1 weeks against 8.5.
Score that month three ways.
A count of targets met gives three passes and two misses. Sixty percent, amber, and the pack says the month was mixed.
A weighted achievement index gives something else. Score each measure as actual over target where higher is better and target over actual where lower is better, then weight service at 40 (split between fill and on time in full), cost at 30, accuracy at 20 and inventory at 10. Fill scores 1.002 at weight 0.20, contributing 0.2004. On time in full scores 0.983 at 0.20, contributing 0.1966. Cost scores 1.028 at 0.30, contributing 0.3085. Accuracy scores 1.025 at 0.20, contributing 0.2050. Inventory scores 0.934 at 0.10, contributing 0.0934. Total 1.0039, or 100.4 percent of plan. Green.
A gated design gives a third answer. Service is a gate: nothing scores green in a month where on time in full sits below 93. Red, and the other four measures do not get read until the gate clears.
One month, one set of honest numbers, three verdicts. The verdict was decided by the design, and the design was chosen by whoever built the spreadsheet.
The weights nobody derived
Weighted achievement indices carry two problems that survive almost every review because they are arithmetic rather than judgement.
The first is that ratio scoring is blind to what a percentage point means. Inventory at 9.1 weeks against 8.5 scores 0.934. Fill rate at 90.6 percent against 97.0 also scores 0.934. One of those is a slightly heavy month that a planner fixes over a quarter, the other loses you a customer, and the composite treats them as identical because both are 6.6 percent off target. Ratio scoring works on measures whose natural scales are comparable, and service percentages and cover weeks are not comparable in any sense that matters.
The second is that the weights bear no relation to money. Translate the month above. The cost beat is 0.04 a case on 1.2 million cases, which is 48,000 for the month. The inventory overshoot is 0.6 weeks against a weekly cost of goods of 5.0 million, so 3.0 million of extra stock, which at a 12 percent carrying rate costs 30,000 a month. Cost carried three times the weight of inventory in the index, while the money runs about 1.6 to one. Meanwhile the on time in full miss of 1.6 points on 33 million of order value put 528,000 of orders late or short, and if the customer agreement deducts 3 percent on non-compliant lines that is 15,840 plus whatever the expedite freight cost, which makes it the smallest financial item on the page and the one with by far the largest consequence for whether that customer renews next year.
Deriving weights from money fixes the first half of that and exposes the second half honestly. Convert each measure to a monthly financial equivalent, weight in proportion, and the argument about weights becomes an argument about carrying rates and margin, which is at least an argument with evidence in it. The measures whose consequence sits outside the month, customer relationships being the obvious one, then need to be handled as gates rather than as weighted terms, because a weighted term will always let a good cost month buy off a bad service month.
What each measure invites you to do
Every measure has a cheapest path to improvement, and it is worth writing that path down next to the measure before you commit to it.
- Unit fill rate improves if you measure it at the distribution centre and count a substitution as a fill. The customer's experience does not change. Measure at the line the customer ordered.
- On time in full improves when promise dates move. A date renegotiated on the day before shipment produces a compliant delivery and an unhappy planner at the other end; the definitional variants that decide this belong to BB4.
- Forecast accuracy at lag one improves by refreshing the forecast late in the month using early actuals, which is free accuracy with no decision value attached, and BB15 has the horizon argument in full.
- Cost per case shipped improves by shipping the dense cases and deferring the awkward ones, and by pushing a shipment across a month end.
- Weeks of cover improves by holding receipts at the port over the reporting date. Four days of held receipts against monthly receipts of 5.0 million is about 660,000 of stock, which on a 42 million balance is 0.14 weeks. Do it two months running and you have produced a fifth of a typical cover reduction target while changing nothing at all.
None of these require dishonesty. Each of them is a defensible operational choice that a reasonable person makes when the measure is what they are judged on, which is why the design question is what the measure pays for rather than whether people are honest.
Targets without a distribution behind them
The other half of a scorecard is the target, and targets get set as round numbers by people who have never looked at how much the measure moves on its own.
Suppose unit fill rate has averaged 97.1 percent over the last two years with a month to month standard deviation of 0.8 points, and nothing about the process has changed. Set the target at 97.0 and the probability of any given month coming in below it is the probability of a normal variable falling more than 0.125 standard deviations below its mean, which is about 45 percent. So the measure goes red roughly five months in twelve on a process running exactly as it always has, and every one of those months generates commentary, an action and a small amount of organisational stress about nothing.
The fix is to treat the target and the action trigger as two different things. The target is where you want the mean to be. The trigger is the point at which a month is unusual enough to investigate, and that belongs two standard deviations out, at 95.5 in this case, which fires on about 2 percent of ordinary months. Between the two sits the run rule: several consecutive months on one side of the mean is evidence of a shift even when no individual month breaches anything, and it catches drift far earlier than a threshold does.
Doing this properly requires knowing the standard deviation of every measure on the page, which takes an afternoon with two years of history and is the single most useful preparatory step before any target-setting exercise.
Cascading without multiplying
The usual cascade failure is arithmetic. Five measures at executive level become nine at function level and thirty-one at team level, because each function adds the things it cares about and nobody removes anything. Measures should shrink as you go down, because a team owns fewer levers than a function does.
The test that decides whether a measure belongs to an owner is whether they can move it inside one planning cycle by doing their job differently. Fill rate at company level fails that test for everyone, which is why it belongs at the top and needs decomposing before it can be assigned. The decomposition that usually works: forecast bias at the lag the commitment is made, which belongs to demand; supply plan attainment against the published plan, which belongs to supply; lead time and unit of measure accuracy in the item master, which belongs to whoever owns master data (P6); and pick accuracy and dispatch cut-off adherence, which belong to the warehouse.
Each of those is movable by a named person within a month. A team handed company fill rate produces commentary, and a team handed supply plan attainment produces a change to how they schedule.
The paired constraint rule
The single design rule that does most of the work: every objective measure gets a constraint measure, and the same person owns both.
An inventory reduction target owned by a planner with no service floor produces a service problem in about six weeks. A service target owned by a planner with no inventory ceiling produces a cash problem in about twelve. A cost per case target with no delivery constraint produces consolidated shipments that arrive late. Forecast accuracy with no bias constraint (BB1) produces a forecast tuned to minimise absolute error by sitting under the peaks, which fills warehouses in a different way.
The ownership detail matters more than the pairing. When service belongs to one director and inventory belongs to another, the pair becomes a quarterly argument rather than a constraint, and the argument is won by whoever has the better relationship with the chief executive. When both belong to the same person, the trade-off gets made weekly by someone who can see both numbers, which is the only place it can be made well.
Four to six measures at the top, each with its constraint named alongside it, fits on one page and takes twenty minutes to review properly. Thirty-four measures fits on four pages and takes ninety minutes to read aloud.
Where this stops
A scorecard changes nothing on its own. Behaviour comes from the review: who attends, what happens when a measure misses, and whether a miss produces a decision with an owner and a date. A well-designed scorecard reviewed by nobody is a report, and a mediocre scorecard reviewed hard by someone with authority will outperform it.
Composites also destroy information by construction. Any index that turns five measures into one number can be produced by many different underlying months, and the whole point of the exercise was to know which month you had. Keep the components visible next to the composite, and treat the composite as an index for tracking direction rather than as a grade.
The last limit is comparison. Nearly every measure on a supply chain scorecard has three or four legitimate definitions, so a number benchmarked against another company is comparing definitions as much as performance. Benchmark against your own history on a definition you have written down and frozen, and treat external comparisons as a source of questions rather than of targets.
Take last month's scorecard, and for each measure on it write the single action a person could take to improve that number without helping the business; anything with a cheap answer needs a constraint attached before the next cycle.
More on this in metrics on the DemandLabs platform.




