Skip to content

Part 5 of 7 · Cash drawer reconciler series ~5 min read

How a cash finding gets closed

A surfaced pattern is a question, and questions need answers or they come back every week until somebody mutes the whole system. This post is about closing the loop: recording what the cause turned out to be, using that to suppress the same message, and reading the monthly summary that tells you whether any of the thresholds are set anywhere near right.

Key takeaways

  • A finding closes by recording a cause, in free text, against a till and slot.
  • A recorded cause suppresses the same pattern until the pattern changes shape.
  • Suppressions expire. A cause recorded in March stops suppressing by summer.
  • The monthly summary is four numbers: two about money, two about the thresholds.
  • If nothing is ever surfaced, the thresholds are too loose — and that is also a finding.

Recording a cause

The one button on a surfaced pattern opens a free-text line. Not a dropdown: the useful causes are always the ones nobody anticipated, and a dropdown of six options produces six months of “Other”.

FieldExampleWhy it is there
tilltill-2Which drawer
slotthu-eveningWhich shift pattern, not which person
patternshort 5-8 on 4 of 5What was reported
causeThursday bakery delivery paid cash from the drawerIn their words
actionpetty cash tin from 1 AugWhat changed, if anything
recorded_bymanager@example.comWho answered
suppress_until2026-10-11When this stops muting the pattern
shapeshort|5.00|8.00|thu-eveningWhat it suppresses, exactly

The shape field is what makes suppression safe. A recorded cause suppresses that pattern — same till, same slot, same direction, similar magnitude. If till 2 starts being short by thirty pounds on Thursday evenings, that is a different shape and it surfaces immediately, cause on record or not. Suppressing by till alone would be how a genuine problem hides behind a bakery delivery.

Why suppressions expire

The cause recorded in March is often no longer true in July. The delivery driver changed, the petty cash tin was introduced, the till was moved. A suppression that never expires quietly turns into a permanent blind spot, and the blind spots accumulate one at a time until the system reports nothing at all. Ninety days is a reasonable default and it lives in the thresholds sheet.

The monthly summary

One month of cash reconciliation in five numbersA horizontal row of five boxes. Closes: one hundred and twelve shifts. Explained: four hundred and eighteen pounds of four hundred and ninety-one pounds of raw variance. Residual: seventy-three pounds net. Surfaced: two patterns. Still open: zero. A note says the second number is the system's actual output and the fourth says whether the thresholds fit.ONE MONTH, IN FIVE NUMBERSCloses112 shiftsExplained£418 of £491Residual£73 netSurfaced2 patternsStill open0The second number is the system's actual output. The fourth says if the thresholds fit.
Fig 1. A month of closes in five numbers. The explained figure is the one that justifies the system; the surfaced count is the one that tells you whether the thresholds need moving.
  • Machine learning
  • Management
  • Analytics
  • Explained versus raw is the number that justifies the whole project. Eighty-five per cent of gross variance having a documented cause is a different operational reality from a variance book full of unexplained figures, and it is the number to put in front of whoever asked what this was for.
  • Net residual matters more than gross. A month that is seventy-three pounds net across a hundred and twelve closes is noise. The same month with four hundred pounds short and three hundred and thirty over is not noise at all — it is two problems cancelling each other out in a total.
  • Patterns surfaced calibrates the thresholds. Zero for three months running means they are too loose and the system is decorative. More than a handful a month means they are too tight and people are learning to dismiss it.
  • Still open is the only one that is about people rather than money. A surfaced pattern nobody has answered in a month means the messages are going somewhere nobody reads.

The case for reporting nothing

It is worth saying plainly that the best possible month for this system is one where it sends no messages at all — a hundred and twelve closes read, every variance either explained or too small to matter, and nobody’s evening interrupted. That is not the system failing to find anything. It is the system doing the fifteen minutes of lookup work per shift that nobody was ever going to do, and correctly concluding that there was nothing there.

The value in those months is not the messages that were not sent. It is that the variance book now has causes in it, that the numbers are computed the same way every time by something that is not tired, and that when a real pattern does appear, it appears against three months of quiet rather than three months of noise.

Next: what all of this costs to run.

All posts