A drawer that is forty dollars short is rarely a theft problem. It is usually a chain-of-custody problem, and the fix is a routine that makes every dollar attributable to a person and a moment.
August 20, 2026
A drawer comes up forty dollars short on a Tuesday. Almost certainly nobody stole anything. The opening count was done by whoever unlocked the door, two people rang on the same terminal through the lunch push, someone made change for the beer delivery out of the till, and the closing count happened at eleven at night by a tired assistant manager who already knew roughly what the number was supposed to be. By Thursday, when anyone actually looks, there is no way to reconstruct what happened. That is the real cost of loose cash handling. Not the forty dollars — the fact that forty dollars is unexplainable, which means four hundred would be unexplainable too.
Cash volume keeps falling in most North American restaurants, and that is precisely why the discipline slips. A category that used to be half of sales and is now single digits stops getting the attention it needs, right up until the month the variance stops looking like rounding. The routine below is not complicated and does not take long. It just has to be the same every single shift.
The useful way to think about cash is that it has an owner at every moment, and ownership only transfers when two people agree on an amount. Starting bank goes from the safe to a named person. Sales go into that person’s drawer. A mid-shift drop moves a counted amount from the drawer back to the safe. The closing count moves the rest. The deposit moves it to the bank. Every link in that chain has a person, a time and a number attached to it.
When a link is shared — two cashiers on one drawer, an unassigned till, a manager who counts alone and records the result later — the chain breaks, and no amount of investigating afterwards will fix it. This is why the answer to a recurring shortage is almost never a camera. It is closing whichever link stopped naming a person.
There are two workable models and one that is not. Terminal banking assigns a drawer to a named cashier for a shift, and only that cashier rings on it. Server banking gives each server their own bank, they keep their own cash and settle up at the end of the night against what the system says they owe. The model that does not work is a shared open drawer that several people use because it is faster during a rush. It is faster, and it also guarantees that any variance is permanently unattributable.
Quick service and counter operations usually land on terminal banking, because the cashier is fixed to a station anyway — the same logic that shapes most of the setup advice in a QSR environment. Full service tends to do better with server banking, since servers are already handling their own checks and the settle-up gives you a natural per-person reconciliation. Either way the point is the same: at any moment, one named person is responsible for a specific pile of money.
That model only holds if the system underneath supports it. The things worth confirming before you build a routine on top of your POS:
If the person counting can see what the drawer is supposed to hold, the drawer will usually be counted to that number. Not dishonestly — human beings recount until they find the expected figure and stop. A blind count means the system does not display the expected total until the counted amount has been entered and submitted. Then it shows the variance.
It is a small setting and it changes what the number means. A blind count produces a real measurement. A sighted count produces an agreement. Only one of those is worth reviewing later, and if your closing figures have been suspiciously clean for months, this is the first thing to check.
Deciding what counts as a problem in the moment, with a specific person standing in front of you, is how policies end up applied unevenly. Write the thresholds down first and apply them to everyone.
| Variance | What it usually is | What to do |
|---|---|---|
| Under a couple of dollars | Change-making and rounding | Record it, take no action |
| Small but occasional | A miscount, or change given fast under pressure | Log it and watch for a pattern by person and shift |
| Large, or small and repeating | A process problem — shared drawers, or a new hire nobody trained | Recount with a second person present, then review the drawer report |
| Cash short while the POS ties out | Sales were rung correctly and money left afterwards | Pull the no-sale and manual drawer-open events for that shift |
| Consistently over | Undercharging, unrecorded sales, or change errors in your favour | Review comps, voids and manual price overrides for the period |
That last row deserves more attention than it gets. Most operators only investigate shortages, but a drawer that is regularly over is telling you that money is arriving without a matching sale, and that is a bigger reporting problem than a shortage of the same size. The discount and void side of this is worked through in comps, voids and discounts.
Set a ceiling for how much cash sits in a drawer, and drop the excess to the safe when it is reached rather than at a fixed clock time. The ceiling depends on your volume and your neighbourhood, but the principle does not: a drawer holding a whole day of cash is both a robbery target and a much harder count at the end of the night.
Each drop should be counted, sealed, recorded in the system at the time, and witnessed if you have a second person on. The recording matters more than the sealing. A drop that exists only as an envelope in the safe is not a transfer of custody, it is a gap in the chain that shows up as a shortage in one place and an unexplained surplus in another.
Most closeouts take too long because they are done in the wrong order — counting starts before the day is actually finished, so the target keeps moving. Run it in this sequence and it usually settles in one pass.
The card batch step is the one that quietly costs the most. A batch that fails to settle looks completely normal on the floor and only surfaces days later as missing deposits, by which point nobody remembers the night in question. Check that it settled while you are still standing there. Disputes that arrive weeks later are a separate discipline, covered in restaurant chargebacks.
One short drawer tells you nothing. Twelve weeks of drawer variance sorted by person, shift and terminal tells you almost everything. Real theft shows as a consistent one-directional bias tied to one name. A training gap shows as noise that starts when someone was hired and settles as they learn. A process problem shows as variance tied to a shift or a terminal rather than to a person — which is usually the shared-drawer problem announcing itself.
Track it as a running number the way you track anything else that matters. Variance as a percentage of cash sales, reviewed weekly next to the rest of your operating metrics, turns cash from something you react to into something you can see coming. Most operators find the number improves within a month of simply being looked at, before any policy changes at all.
Do not roll out the whole routine at once. Start with drawer assignment and per-user login, because nothing else produces useful data until every transaction carries a name. Turn on blind counts the following week. Add the drop ceiling after that, then the variance review. Four changes over a month, each one small enough that the team can absorb it during service, and by the end you have a cash process that survives staff turnover — which is the actual test, since the person who currently holds all of this in their head will eventually leave.
Most operators treat a couple of dollars either way as normal change-making noise and investigate anything larger. The absolute figure matters less than consistency and direction: a drawer that is a few dollars off in random directions is a counting reality, while one that is off by the same rough amount in the same direction shift after shift is a process or a person, regardless of how small the number is.
A blind count means the system does not show the person counting what the drawer is expected to hold until after they have entered and submitted their figure. Without it, people naturally recount until they reach the number on screen and stop, so the closing figure becomes an agreement rather than a measurement. Blind counting is usually a single configuration setting and it is the cheapest control available.
Server banking works well in full service, where each server already owns their checks and settles up against what the system says they owe at the end of the night. Counter and quick-service operations usually do better assigning one drawer to one named cashier per shift. The model to avoid in either case is several people ringing on one open drawer, because any variance it produces cannot be attributed to anyone.
Drop on an amount rather than a clock — set a ceiling for how much may sit in a drawer and drop the excess whenever it is reached. Each drop should be counted, recorded in the POS at the moment it happens, and witnessed where staffing allows. A drop that exists only as a sealed envelope is not a transfer of custody and will show up later as a shortage in one place and a surplus in another.