Loss prevention spends most of its attention on the smallest controllable slice of the problem. In grocery and perishable-heavy retail the money is in product that was ordered wrong, rotated wrong, marked down too late, or held at the wrong temperature - all of which are forecastable, unlike theft.
The numbers, and where they come from
Brazil publishes unusually good data on this. Abrappe's 9th loss-prevention survey, conducted with Protiviti and presented in June 2026, reported R$42.1 billion in losses for 2025, equal to 1.65% of revenue, up from 1.51% the year before - a 9.27% rise in the loss index itself. Put in operational terms: for every R$61 sold, R$1 was lost.
The direction matters more than the level. Against sector revenue of R$2.55 trillion, losses grew 15.3% while revenue grew 6.4% - loss is compounding at well over twice the rate of sales, which means it gets worse every year you leave it alone. Meanwhile only about 13% of Brazilian retailers apply AI in operations. A large, accelerating, precisely quantified problem in a market that has barely picked up the obvious tools is an unusual combination.
The category detail is where it gets actionable. Pharmacies are being hit by theft of high-value GLP-1 medications such as Ozempic, a genuinely new loss vector that no historical model anticipated. Convenience stores saw their loss index climb from roughly 3% to 3.80% in a single year. And the survey found measurable positive impact from technology investment (RFID, analytics, monitoring) specifically in cash-and-carry, sporting goods, and regional chains - which is evidence that this is tractable, not just expensive.
On causes, that research attributes roughly 36-39% of loss to operational causes against roughly 17-20% to external theft, with expired product the single largest identified cause at around 41%, and perishables driving about 76% of loss impact.
A caution on US figures. We do not quote US shrink statistics with confidence, and you should be skeptical of anyone who does. The National Retail Federation discontinued its National Retail Security Survey in 2024 after 32 years, and retracted a widely repeated organized-retail-crime number in December 2023. Shrink has sat in a narrow band near 1.5-1.6% of sales for decades. Scope your project from your own markdown and disposal data, not from industry headlines.
What we actually build
- Shelf expiry detection. Reading date codes on shelf and flagging product approaching expiry while a markdown can still recover most of its value, instead of discovering it in the disposal bin.
- Condition and ripeness grading. Produce and bakery graded consistently, so rotation and markdown decisions stop depending on which employee is working.
- Markdown timing. The optimization question is not whether to discount but when, and by how much: too early gives away margin on product that would have sold, too late gives away all of it. This is a forecasting problem with a clean weekly feedback loop.
- Cold-chain monitoring. Catching temperature excursions and equipment degradation before a case of product is compromised.
- Order forecasting for short-life categories. The cheapest loss to prevent is the unit you never ordered - and short-life SKUs are where forecast error converts directly into disposal.
The measurement problem, and how we handle it
This is the part most vendors gloss over, and it decides whether performance-based pricing is honest here.
Total shrink is a terrible fee metric. It is a residual - the gap between book and counted inventory - calculated at physical inventory once or twice a year. It arrives 12 to 24 months after the work, it bundles theft, admin error, and spoilage into one ambiguous number, and it is computed by the client. Nobody should sign a contingency contract against it, including us.
Avoided markdown and disposal, by category, against matched control stores, is a good one. Those numbers already exist in your merchandising system weekly. Run the system in a set of stores, hold out a matched set, compare over the same window, and the seasonality and market movements that would poison a before-and-after comparison affect both groups equally. That is fast, attributable, computed from your data, and it is what we will write into the measurement agreement before building anything.
No cameras pointed at people
Everything above works on product, shelves, and equipment. No face recognition, no gait analysis, no biometric identifiers, in any market we operate in. That is deliberate on three grounds: biometric privacy litigation is an existential risk for a firm our size, LGPD and BIPA-style regimes make biometric data disproportionately expensive to hold, and camera systems that staff read as surveillance get quietly defeated by the people who work beside them. Loss prevention that the floor team supports is worth more than loss prevention it resents.
Who this is for
- Regional supermarket and atacarejo chains with enough stores to hold a control group - roughly ten or more.
- Retailers where perishables are a meaningful share of revenue.
- Operators who already track markdown and disposal by category, since that is the baseline we measure against.
If you run a handful of stores, or if your loss is concentrated in non-perishable theft, the honest answer is usually process and staffing rather than a model, and we will tell you that on the call rather than after the invoice.