A single variance figure mixes unlike things. Purchase price variance says the supplier charged differently from the model. Freight variance says transport cost more, or was billed on different dimensions. Duty variance says the rate, the classification, or the value moved. Data variance says one of the systems is simply wrong, a stale exchange rate, an old divisor, a missing cost line. Each cause has a different owner and a different fix, and a report that cannot tell them apart sends every line to a meeting instead of to a person.
Threshold each cause separately, in money and in percentage, and require both: a two percent freight variance on a large line is real money, while a forty percent variance on a tiny accessory shipment is arithmetic noise with a dramatic costume. Dual thresholds are a small piece of report design that decides whether the report survives its first month.
Individual alerts train reviewers to dismiss; the unit of decision is the pattern. One dutied shipment over model is an event to note. The same SKU over model six shipments running is a wrong rate, a wrong classification, or a supplier price change nobody recorded, and it will keep billing until someone changes a master record. Structure the review to surface repeats first: SKUs, suppliers, and lanes appearing again, ranked by cumulative money, with the single odd events in an appendix for anyone with time.
That ordering changes the meeting. Instead of sampling random lines and speculating, the review works a short list of recurring causes, assigns the master data fix, and checks next period whether the line disappeared. A variance process that does not end in master data corrections is a subscription to the same report forever.
Every alert line should carry the name of the role that can fix its cause: buyer for price, logistics for freight, broker or compliance for duty, systems for data. Unowned lines are read by everyone and fixed by no one, and they accumulate until the report is a museum of known problems. If a cause has no owner in the business, that is the finding, and it is worth more than any single variance in the file.
Close the loop visibly: the following report shows lines fixed, lines recurring, and lines accepted with a reason and an expiry. Accepted variance, we know the model is wrong for this lane and will fix it at contract renewal, is a legitimate state with a date on it. Accepted forever, with no date, is how a threshold quietly becomes a rug.
Persistent variance in one direction means the model is wrong, and the honest response is to update the model, not to widen the threshold until the report goes quiet. A model updated from observed costs, lane by lane and season by season, turns variance reporting back into what it is for: catching the unexpected, a billing error, a rate applied wrongly, a supplier change that never reached the file. Those are the catches that pay for the whole process in single incidents.
Keep the history when you recalibrate. The old model, the date it changed, and the data behind the change are the audit trail for every margin reported in between, and finance will ask for it at the least convenient moment. Version the cost model like the financial record it is, because it is one.
Keep reading: How do you spot landed cost drift when a supplier quietly changes pack sizes?, Which exception reports catch landed cost errors before month end close?.
Dual thresholds per cause: a money floor and a percentage, both required. This filters tiny line noise while catching small percentage moves on large lines. Thresholds should differ by cause, price, freight, duty, and data, because their owners and normal ranges differ.
Because early thresholds flag everything, the report runs to hundreds of unexplained lines, and no line has an owner. Reports survive when they are short, ranked by cumulative money and repetition, and every line names the role that can fix its cause.
An event is one shipment over model, worth noting. A pattern is the same SKU, supplier, or lane over model repeatedly, which means a master record, rate, classification, or price, is wrong and will keep costing money until it is corrected.
When variance persists in one direction, the model no longer describes reality. Recalibrate it from observed costs, keep the prior version and the evidence for the change, and let alerts go back to catching the genuinely unexpected.