Fourteen Stores, 4.6 Million Tickets, One Loss-Prevention Review: Checking Every Void Just Got Cheap

Fourteen stores. Roughly 900 tickets a store on a weekday and half again as many on a Friday, which comes out near 4.6 million transactions across the portfolio in a year. Your loss-prevention review looks at about 900 of them: the top rows of the exception report, sorted by dollar value, on a Monday morning, in the ninety minutes your director of operations has before he leaves for the first store visit.

That is not a criticism of the DO. It is arithmetic. Nobody in multi-unit has ever had the hours to look at the other 4,599,100 lines, so the trade built a habit around sampling — pull the biggest voids, call the three general managers with the worst comp percentage, note it in the Monday recap, move on. Every operator I have sat with in an above-store office runs some version of that morning.

The rows below the fold

Open Delaget or Crunchtime or whatever your brand's approved above-store reporting is, and the exception report will hand you the usual families: manager voids after the ticket closed, employee-meal discounts rung on a guest order, refunds without a matching original ticket, no-sale drawer opens, order cancellations at the pickup window, and the third-party delivery adjustments where DoorDash or Uber Eats decided the guest was right about the missing side and charged it back to store 0417.

The dollar-sorted top of that list is where the fraud stories come from, so that is where everybody looks. But the money in a fourteen-store portfolio is almost never in one $340 void. It is in a $4.10 employee discount applied 60 times a week at one register by one closer who has figured out that nobody reads below row 20. It is in the 3 percent of delivery orders that get refunded at store 0417 and 0.4 percent at store 0422, a gap nobody has ever put a number on because the two reports live in different tabs.

Transaction-level review means every single ticket, not a sample, gets read against the rules you already have written down — and only the ones that break a rule reach a person.

What the other 4.6 million rows were hiding

The reason sampling survived this long is that a rule check is not really a rule check. To know whether a comp was legitimate you have to hold four things next to each other: the ticket itself, the manager PIN and reason code attached to it, the drive-thru timer entry for that same second, and whether the guest complaint that supposedly caused the comp shows up anywhere in the brand's guest-feedback tool. A person can do that in about ninety seconds. Ninety seconds times 138,000 exception lines a year is not a job anybody is going to fill.

So operators bought thresholds instead. Flag comps over $25. Flag any employee with voids more than two standard deviations above the district. Those rules catch the loud stuff and miss the patient stuff, and every crew member who has worked two brands knows roughly where the threshold sits.

The price of looking fell about tenfold

Here is the change that matters to you, and it is not a new feature — it is a price. Running a capable AI model over a piece of text cost roughly ten times more in 2025 than it does now in mid-2026, and for high-volume work that runs on your own hardware in the back office it is cheaper still, on the order of ninety percent below cloud pricing. Reading one ticket, with its reason code and its timer entry and its delivery adjustment, went from something you would only do to a sample to something you would do to everything.

In 2025 I talked several operators out of this. The honest math then was that checking every ticket in a fourteen-store group cost more than the shrink you would find in half of them. That sentence is no longer true, and the only thing that changed is the invoice.

flowchart TD
  A["Toast tickets, all 14 stores"] --> E{"Does the comp match a manager PIN and a real reason code?"}
  B["DoorDash and Uber Eats adjustments"] --> E
  C["Drive-thru timer, order to present"] --> E
  D["Guest complaints logged in the brand tool"] --> E
  E -->|"Everything lines up"| F["Filed, nobody is interrupted"]
  E -->|"Something does not"| G["Flagged with the ticket and the timer side by side"]
  G --> H["DO calls the GM before the 10am open"]

6:50 a.m. Tuesday, above-store office

The DO opens one page. Not an exception report — a short list, usually four to nine items across fourteen stores, each one written as a sentence. Store 0417, closing shift Sunday: eleven employee-meal discounts rung between 9:40 and 9:58 p.m. on tickets that also carried drive-thru timer entries, meaning the food went out the window, not to the break room. Store 0409: refund issued at 2:14 p.m. with a reason code of "wrong order" and no corresponding remake on the make line.

Each item shows the underlying ticket, so the DO is not taking anything on faith. He forwards two of them to the GMs with a one-line note, keeps one for the store visit on Thursday, and dismisses the rest as explainable. Total time: eleven minutes. The other seventy-nine minutes go to the thing he was actually hired for, which is standing in a store during a lunch rush.

The second-order effect is the one operators underrate. When the review is complete rather than sampled, you stop needing thresholds, and crew stop being able to work under one. That deterrent is worth more than the recoveries.

The arithmetic, with the assumptions on the table

This is an illustration, not a benchmark from any brand's system. Assume fourteen units at a $2.0 million average unit volume, so $28 million in portfolio sales. Assume 4.6 million tickets a year. Assume checking one ticket against its reason code, timer entry and delivery adjustment costs about six-tenths of a cent at 2025 pricing and about six-hundredths of a cent now.

Now the other side. Assume discount and void abuse across the portfolio runs three-tenths of one percent of sales, which is $84,000 a year on $28 million, and assume complete review plus the deterrent effect recovers half of it in year one. That is $42,000 against $2,760 of checking and maybe forty hours of setup. Even if you think three-tenths is double the real number, the trade is still obvious — and unlike most AI arithmetic, this one is provable inside a quarter by comparing comp percentage at the same fourteen stores before and after.

What this will not tell you, and what still needs a store visit

It will not tell you why. A pattern of late-night employee discounts at one store is equally consistent with theft, with a GM who told the closers to help themselves because the fryer went down, and with a broken reason-code list nobody has updated since the last menu change. I have watched an operator fire the wrong person off a report. Every flagged item should be treated as a question for the GM, not a finding.

It will also not survive a brand standards problem dressed up as a loss problem. If store 0417 is comping heavily because the drive-thru headset has been failing for three weeks and orders keep coming out wrong, the answer is a work order in ServiceChannel, not a conversation about integrity. The report cannot see the headset. Your DO can.

And keep the human on anything that touches an employee's record. Coaching, write-ups and terminations stay with the person who runs the org chart — a flag is evidence to go look, never a verdict.

Frequently asked questions

Does this replace my loss-prevention vendor?

No, and if you have Delaget Guard or a similar service under contract, keep it. What changes is that the exception report stops being the finish line. The reports still surface categories and rankings; the new capability reads the individual tickets underneath them, which is the part your people were never going to get to.

My franchisor mandates the reporting tools. Can I even do this?

Read your franchise agreement's technology clause before you connect anything. Most brands restrict what can be installed on the point-of-sale itself and are far more relaxed about what you do with your own sales data once it lands in your above-store reporting. Ask your franchise business consultant in writing; several brands have started publishing guidance because enough operators asked in the last year.

What does it cost to run for a fourteen-store group?

The checking itself, at the volumes above, is a few thousand dollars a year — genuinely less than one store's annual pest-control contract. The real cost is the four to six weeks of getting your reason codes cleaned up so the rules mean something, which is work you have been meaning to do anyway.

Will crew figure out how to beat it?

Some will try, and the honest answer is that a rule you write down can be worked around. The difference is that there is no longer a row 20 to hide below, so the workaround has to be small enough to be pointless.

The first hour on Monday

Do not start with fourteen stores. Pick the two with the widest gap in comp percentage, export ninety days of ticket-level detail for both, and have every line read against the reason codes your brand already publishes. You will get one of two results: a list of specific nights to ask about, or confirmation that the gap is menu mix and not behavior. Either answer is worth having, and you can get it before your next period close.

One last note that sits next to this rather than inside it. A good chunk of what shows up as a comp or a refund starts as a phone call nobody answered — the guest who could not reach store 0417 about a wrong catering order and went to the brand's guest-feedback form instead, which lands on your desk as a make-good. CallSphere builds AI voice and chat agents that answer the store line and the web chat around the clock, take the order details, and book the callback so the complaint gets handled at the store instead of becoming a chargeback two days later. It does not read your tickets — that is the work above — but it does cut down on how many of them need reading.