Somebody is stealing from you. That is the premise of almost everything written on this subject, and the number attached to it is always the same: employee theft costs restaurants 4% of sales and accounts for 75% of inventory shrinkage. You will find that sentence in consultancy brochures, loss prevention whitepapers and the marketing of every company selling you a camera. It is attributed, without exception, to the National Restaurant Association. It is also the number a supplier quotes at you while pricing a camera package, or a stock management module, against a loss nobody in your building has measured.
I went looking for the primary source and could not find one. The annual cost attached to the same claim is $8.5 billion in some retellings and $20 billion in others, a difference of more than double, which is not the sort of thing that happens to a figure with a methodology behind it. Other versions of the same attribution say 7% rather than 4%. So before anything else: the number you have been quoted at trade shows for twenty years has no traceable study underneath it, and if a supplier leads with it, ask them which report, which year and which sample. What follows uses only figures I could trace to a source, which turns out to be a smaller and much more interesting set. It also points somewhere other than the camera, because most of the usable evidence is already inside your restaurant POS, in the void, discount and reprint trail that nobody reads.
What has actually been measured
There is real research here, and it is unusually good, because a point of sale distributor handed academics seven years of transaction data. In 2021 Manufacturing and Service Operations Management published a study by Tat Chan, Yijun Chen, Lamar Pierce and Daniel Snow covering 83,153 servers at 1,049 restaurants across 34 casual dining chains in 46 US states, built on more than 5.7 million transactions.
Their headline descriptive figure is the one worth knowing: 56% of servers in that data set committed identifiable theft at least once.
Read that carefully before you use it. "Identifiable theft" means transactions the monitoring software classified as theft, which is not the same as theft proven to a legal standard, and a single flagged incident over a long period is a very different thing from habitual stealing. The data is US casual dining, where servers are tipped and handle cash. But it is a real measurement from a real sample, and it says something more useful than 4% of sales. It says that a majority of servers will, at some point, do something the system would flag, and that a policy built on the assumption that theft is rare and confined to a few bad people is built on a false premise. Most of your losses are not one villain. They are a thin, wide film.
The money is not in the theft you recover
The same research group published the study that should actually change how you think about this. In Management Science in 2015, Lamar Pierce, Daniel Snow and Andrew McAfee examined 392 restaurant locations across five casual dining firms that installed transaction monitoring software, using difference in differences models with staggered adoption dates so that locations which had not yet switched on acted as controls.
Theft went down. The direct saving from theft alerts was $108 per restaurant per week. That is real money and it is also, by restaurant standards, unremarkable.
Weekly revenue per restaurant went up by $2,982, roughly 7%.
The revenue effect was about twenty seven times the theft recovered. And it was not caused by sacking people: the authors found the effect came from changed behaviour among the same staff rather than from turnover. The most plausible mechanism, and the one the authors examine, is that servers who could no longer make money by skimming made it the legitimate way instead, by selling. A second beer, a dessert, a better bottle. Their tips went up with the till.
Two caveats, because this result is strong enough to be worth stating honestly. These were firms that chose to buy theft monitoring, so they are not a random sample of restaurants, and the authors are explicit that they are estimating the effect on adopters. And the mechanism runs through tipping. In a Spanish or French dining room where staff are salaried and any service charge is pooled or absent, the private incentive to sell harder is much weaker, so I would not expect a European venue to see 7%. What transfers is the direction and the logic: the reason to look at your own transaction data is not the cash you claw back, it is what happens to selling behaviour when people know the data is read. If you want the other half of that, we wrote separately about increasing average check.
It spreads, and your new starters are the vector
The peer effects study is the one that should change your onboarding rather than your camera budget.
Servers were more likely to steal when they worked alongside high theft colleagues. The spillover coefficient was modest, about 4% on theft count, but the authors then show how small coefficients compound in a shared workplace. Their estimates indicate that doubling one worker's average theft raises total theft in an average restaurant by 76%, and doubling every worker's raises it by 550%. A single persistent thief is not just their own losses. They are a teacher.
And the contagion has a window. The positive peer effect from high theft colleagues existed only for workers in their first three to five months, and then disappeared. New staff arrive without a settled sense of what normal is here, learn it from whoever is next to them, and then hold it. Which means the most cost effective loss prevention available to you is deciding who trains your new starters, and it costs nothing. Put the person you trust most on the first fortnight of every new hire, and do not put a new server next to the one you have doubts about. Our guides to staff training and hiring and retention cover the surrounding process.
There is one more finding in that paper, and it is the strangest and most revealing thing in this whole literature. On days when their colleagues stole more, servers stole less. And that negative correlation was stronger in restaurants running the monitoring software. They were not copying each other's behaviour on the day; they were managing shared detection risk, easing off when the shift already looked suspicious. People are modelling your surveillance. Which tells you they can tell whether anybody is reading it.
Your best detector is a person, not a report
For losses at the larger, slower end, the best data comes from the Association of Certified Fraud Examiners, whose 2024 Report to the Nations covers 1,921 cases across 138 countries. The overall median loss was $145,000 and the median case ran for twelve months before it was found. For food service and hospitality specifically the median loss was $100,000, twice the retail figure of $48,000.
The gradients are what you would expect and are worth saying out loud anyway. Frauds by ordinary employees had a median loss of $60,000, by managers $184,000, and by owners and executives $500,000. Cases involving people with more than ten years of service were the costliest at a median of $250,000, while almost half of all cases involved someone with one to five years of tenure. Your controls almost certainly point at the people with the least authority and the shortest service, which is the opposite of where the money goes.
Then the finding that should reorder your spending. Tips from people were how 43% of cases came to light, more than three times the next most common method. Internal audit found 14% and management review 13%. Of the tips, 52% came from employees, 21% from customers and 11% from vendors.
Nothing in your point of sale will outperform a member of staff who is willing to tell you something. Most restaurants have no mechanism for that at all beyond "tell your manager", which fails precisely when the manager is involved, and asks a twenty two year old on an hourly wage to accuse a colleague to their face. A named person outside the venue's own line of management, an email address that a general manager cannot read, and a stated position that the person who reports keeps their job, costs nothing and is the single best value control on this list. Our note on the employee handbook is where the policy belongs.
What to look at in the till data
I should be plain about the conflict of interest here. We sell a restaurant POS, so treat a POS company telling you the answer is better POS reporting with the scepticism it deserves. The honest version is narrower: your transaction log is the cheapest evidence you own, you are almost certainly not reading it, and reading it has legal advantages over the alternatives that most articles on this subject never mention.
Every pattern below is a question, not a finding. Each has an innocent explanation that is usually the correct one.
Voids and deletions, by who and by when
The number that matters is not your void total, it is the distribution across staff for comparable shifts, and whether the item was removed before or after it was sent to the kitchen. An item deleted before it prints is a correction. An item served and then voided is a question, because the food left the building and the money may not have arrived. Reason codes are what make this readable at all, which we go into in voids, comps and discounts.
Comps and discounts as a rate, not a total
Compare discount value as a percentage of each person's own sales. One server comping twice as much as everybody else per euro sold is worth a conversation even if their absolute total is small. Staff discounts applied outside staff hours belong in the same review.
Reopened and transferred checks
A closed check being reopened, an item moved between tables, a check transferred to another server late in a shift. All legitimate operations. All also the mechanics of moving a paid item onto an unpaid bill.
No sale drawer opens
Opening the till without a transaction has ordinary uses, making change among them, and it is also the simplest possible way to remove cash. Count them per person per shift and look at the outliers rather than the total.
Card receipts that reappear
The same card authorisation, or the same unusual amount, closing more than one table across a shift is the pattern behind the oldest trick in table service: settle a cash table against a card slip from an earlier table and keep the notes. It shows up in the data as duplication that has no service explanation.
Refunds without an original
Refunds and post payment adjustments should reconcile to a specific earlier transaction. Ones that do not are worth explaining before your daily close is signed off, not a fortnight later.
The variance the till cannot see
A generous pour, a steak that walks out of the back door and a delivery signed for short never touch the point of sale at all. That is a counting problem, and the only thing that finds it is stock management with real counts on a fixed cycle, plus the pour cost arithmetic on your highest volume lines. Our inventory management guide covers the count discipline.
What European law actually lets you do
Almost all writing on restaurant theft is American, and its advice, install cameras and watch your staff, will get a European operator fined, lose them a tribunal case, or both. This is the part worth reading twice.
The six questions
In Bărbulescu v. Romania, decided by the Grand Chamber of the European Court of Human Rights on 5 September 2017, the court set out the factors national authorities must weigh when an employer monitors staff. Whether the employee had prior notice of the possibility and the nature of the monitoring. The extent of the monitoring and how intrusive it was. Whether the employer had legitimate reasons for it. Whether a less intrusive method would have achieved the same aim. The consequences for the employee. And whether adequate safeguards existed.
Those six questions are the test, and they are worth writing down before you buy anything, because the answers are what you will be asked for later. Notice that "we suspected theft" answers only one of them.
What the cameras case actually decided
Operators who have heard of López Ribalda and Others v. Spain usually remember it as permission for hidden cameras. The Grand Chamber judgment of 17 October 2019 is narrower and more instructive than that.
A supermarket found discrepancies between stock and sales amounting to around €80,000 over five months. It installed cameras, told staff about the visible ones and not about those pointed at the tills. Ten days later it had its evidence, stopped filming and dismissed fourteen people. Five went to court. The Grand Chamber held, by fourteen votes to three, that there was no violation of Article 8, overturning a Chamber judgment that had gone the other way the previous year.
Every element of why it survived is a constraint on you. There was a reasonable suspicion of serious misconduct, based on quantified losses, before the cameras went up. The suspicion pointed at concerted action by several people. The cameras covered only the checkout area, the place where the suspected loss occurred. It ran for ten days and stopped as soon as the individuals were identified. The footage was used for nothing except proving the thefts. A speculative camera left running over your bar for a year because theft is generally a risk has none of those features, and the same court said in the same judgment that a departure from the duty to inform staff needs a weighty justification.
National rules then add layers on top. In France the CNIL's position is that cameras must not film employees at their workstations except in particular circumstances such as handling money, and even then the framing should favour the till over the cashier. Constant and permanent surveillance of a person is excessive. Footage should in principle be kept no more than a month, with the relevant sequence extracted and logged if a procedure begins. Employees must be informed individually, the works council consulted before the system goes live in businesses of fifty or more, and cameras covering areas open to the public need préfecture authorisation. In Spain, Article 89 of the data protection act allows image processing for the employer's supervisory powers but requires prior, express, clear and concise information to workers and their representatives, forbids cameras entirely in changing rooms, toilets, canteens and other rest areas, and treats audio recording as exceptional. The only softening is for a flagrant unlawful act caught on camera, where the signage alone can satisfy the duty to inform.
Why your transaction log is legally different from a camera
Here is the practical asymmetry that almost nobody tells operators about, and it is the strongest argument for looking at data rather than filming people.
Italy is the clearest case. Article 4 of the Workers' Statute says that audiovisual equipment and other instruments from which remote monitoring of workers' activity may result can be used only for organisational and production needs, work safety or the protection of company assets, and can be installed only after a collective agreement with the union representatives or, failing that, authorisation from the national labour inspectorate. That is a serious procedural hurdle for a camera. But the second paragraph of the same article exempts instruments the worker uses to perform the work, and systems that record access and attendance. A till that a waiter operates in order to take orders is, on its face, a work tool. The information duty and data protection rules still apply in full, and you should take Italian advice on your specific configuration rather than my reading of it, but the union agreement or inspectorate authorisation that a camera triggers does not obviously attach to reading the transaction history of your own point of sale. The inspectorate issued guidance as recently as 26 May 2025 on how multi site businesses should file those camera applications, which tells you how live the process is.
Germany runs the other way and is stricter about the software. Under section 87(1)(6) of the Works Constitution Act, a works council has full co-determination over the introduction and use of technical devices suitable for monitoring the behaviour or performance of employees, and a monitoring report on individual servers is squarely that. If you have a works council, you need an agreement, and this is not a formality you can retrofit.
The general point holds across all five markets: a report showing that one server's void rate is four times everybody else's is targeted, proportionate, keeps no images of anybody, and answers the "less intrusive method" question in Bărbulescu in the way you want it answered. It is also, unlike footage, something you can show the person and ask them to explain.
When you find something
A pattern is not proof, and the most expensive mistake at this stage is treating a report as a confession.
German law is instructive here even outside Germany, because it has thought hardest about the problem. A dismissal on suspicion, the Verdachtskündigung, is a recognised and separate ground from dismissal for a proven act, but the Federal Labour Court's requirements are strict: strong grounds resting on concrete objective facts, an urgent suspicion meaning a high probability that it is true, circumstances not equally well explained by something innocent, and, critically, the employer must have made every reasonable effort to establish the facts, which specifically includes giving the employee the opportunity to respond. That hearing is a condition of validity, not good practice. It has to put a concrete allegation, bounded in time and place, so the person can actually rebut it, and confronting somebody with a vague accusation while withholding what you have does not count. There is also a clock: a summary dismissal has to be issued within two weeks of the employer learning the decisive facts, so an investigation that drifts can destroy the remedy.
The German courts have also been clear that data protection failures do not immunise thieves. In a judgment of 29 June 2023 the Federal Labour Court held that findings from open video surveillance intended to show a deliberate breach of duty can generally be used in dismissal proceedings even where the surveillance itself did not fully comply with data protection law, on the reasoning that the right to informational self determination cannot be invoked for the sole purpose of escaping responsibility for intentional wrongdoing. That is a genuine protection for employers, and it is not a licence: it applied to open surveillance, and serious breaches of fundamental rights can still exclude the evidence.
The workable sequence, in any of these markets, is the same. Establish the pattern from records you were entitled to hold. Quantify it. Check the innocent explanations yourself before you speak to anyone, because most of them are innocent. Then hold a documented conversation in which you put the specific facts and listen to the answer. Take local employment advice before you dismiss, every time, because none of these five countries treats this the same way and the cost of getting it wrong is a reinstatement or a compensation award on top of the money you already lost.
What to do this month
Read one week of exception reports. Voids after the item was sent, comps as a percentage of each person's own sales, no sale drawer opens, reopened checks, refunds without an original. You are not hunting anybody; you are finding out what your own baseline looks like, which you probably do not know.
Set up a way for staff to tell you something without going through their line manager. Forty three percent of cases in the ACFE data surfaced that way. Nothing else you do this month has that hit rate.
Decide who trains new starters, and treat it as a control rather than a rota convenience. The contagion window is the first three to five months and it only runs in one direction.
Put a count cycle on the categories the till cannot see, starting with spirits and the three highest value items in your walk in.
Tell your staff that transaction reporting is reviewed, in writing, before you start reviewing it. It satisfies the notice limb of the legal test, it is the honest thing to do, and on the best evidence available it is also the part that makes money, because the measured gain from monitoring came from people selling more once they knew somebody was looking.
And if a supplier quotes you 4% of sales, ask for the study.
Read next: voids, comps and discounts, the daily close and Z reports, and the shift handover.
