On this page
The calculation
Variance is the difference between what you should have used and what you actually used, expressed against what you sold.
- Theoretical usage. For each product, multiply units sold by the recipe pour. Twelve hundred well vodka drinks at a 1.5 ounce spec is 1,800 ounces.
- Actual usage. Opening stock plus deliveries minus closing stock, in the same unit.
- Variance. Actual minus theoretical, then divide by theoretical for a percentage.
If theoretical was 1,800 ounces and actual was 1,890, you are 90 ounces over, which is five percent. At a 750 millilitre bottle, roughly 25.4 ounces, that is about three and a half bottles across the period.
Three notes that catch people out. Work in one unit throughout and convert once at the end. Count in the same order every time so the count itself does not become the variable. And expect a genuine floor of one to two percent from spillage and shake off that no bar eliminates; chasing zero is chasing noise.
The four causes, and how to tell them apart
This is the whole job. A variance number is a symptom, and treating the wrong cause is worse than doing nothing, because you spend credibility coaching somebody about a pour when the problem was the loading bay.
Overpouring
Usually concentrated in specific people on specific shifts rather than spread evenly. The tell is that variance tracks who was working rather than how busy it was. If the same two names are on every heavy period, that is your answer.
Unentered comps and voids
Product left the shelf but never appeared as a sale. The tell is in the transaction data rather than the count: void and comp rates that sit well above the team, particularly late in a shift or clustered around close.
A short delivery
The tell is a step change rather than a drift. Variance was normal, then it moved and stayed moved from a specific date. Check what was signed for around that date against what was invoiced. This one is often recoverable as a credit, which is why it is worth ruling in or out first.
Unlogged breakage
Diagnosis by elimination, and the money is already gone. The fix is a logging habit, not an investigation.
What to do the same night
The reason to care about timing is that the cost of a variance is roughly proportional to how long it runs. A pattern found on Thursday and fixed on Friday cost you one week. The same pattern found in next month's reconciliation cost you four or five, and by then the conversation is about something that happened long enough ago that nobody remembers the shift.
Practically, that means the useful cadence is not the count. It is the transaction stream, which is available continuously and can tell you about voids, comps and per person patterns without anyone touching a bottle. The count then becomes a confirmation rather than the discovery mechanism.
Turning it into something that gets done
A variance investigation that ends in a number is not finished. To actually change anything it needs to end in a piece of work: a named person, a specific action, a date, and a record of what happened. Otherwise the same variance shows up next period looking exactly like a new problem.
That is the argument in the data to action gap, and the reason a pour cost report on its own will not get you there. If you are comparing tools for this specifically, CoreTAP and Wisk.ai covers the counting versus attribution split honestly.