Computer vision in LATAM retail: what works in 2026 (and what doesn't)

The one-minute version

Seven out of ten computer vision projects in LATAM retail never reach production. Not because the tech is bad — because it gets pointed at problems where the ROI doesn't pencil out. Here's the honest breakdown: which use cases are already paying for themselves at Walmex, Falabella and Cencosud, which never will, and why Amazon Go in Seattle doesn't work in a corner store in Coyoacán.

• Global retail shrink losses run about $112 billion in the U.S. alone per the NRF National Retail Security Survey. LATAM estimates run higher: 1.7–2.3% of revenue vs. 1.6% in the U.S., because external theft and organized retail crime are growing faster than the operational response.
• Computer vision in retail pays off in four scenarios: shelf monitoring (planogram compliance), shoplifting detection, queue analytics, and self-checkout + smart cart. It does not pay in three: facial recognition for demographic targeting, theft prediction "before the fact," and one-camera demo pilots.
• Live LATAM examples: Cencosud loss-prevention pilots in Chile, queue analytics at Walmex and Soriana, Scan&Go at Sam's Club Mexico, smart cart at Carrefour Brasil, visual search in Falabella's e-commerce.
• Cheapest entry point: shelf monitoring on the CCTV you already have. A defensible pilot costs $5–15k, not $300k.
• In 7 of 10 CV-in-retail projects, the failure isn't the algorithm. It's the ERP integration (Odoo, SAP, Oracle Retail) and operational change management.

Why now and why LATAM

LATAM retail has spent three years in "squeeze every basis point of margin" mode. E-commerce ate 6 to 11% of consumption depending on the country: Chile 11%, Mexico 8%, Peru 6%, per aggregated 2024–2025 data from local chambers of commerce and sector trackers. Physical retail didn't die — but margins compressed 200 to 400 basis points over five years. In parallel, three things flipped that make computer vision technically and economically viable right now, not in 2019.

First: edge computing got 10× cheaper. An NVIDIA Jetson Orin Nano costs $250 and runs real-time inference on 4 cameras at once. The 2022 equivalent ran $2.5k and needed a server rack in the back room. For LATAM, where in-store internet drops 2 to 6 times a month, edge isn't an optimization — it's the only sane option.

Second: vision foundation models matured. YOLOv8/v9, Grounding DINO, SAM 2 and DINOv2 fine-tunes deliver 92–96% accuracy on retail tasks out-of-the-box, no six-month dataset annotation marathon. What needed an 8-person ML team in 2020 now runs on one senior CV engineer with off-the-shelf pipelines. POC cost dropped from $250–400k to $30–80k.

Third: LATAM retailers grew up on data. Walmex, Falabella, Cencosud and Coppel built data teams of 50–200 people. Without that foundation, any CV project collapses because the model's output has nowhere to go. The top 20 retailers have the infrastructure now. PYME-scale chains (50–500 stores) have it partially — often on Odoo plus an analytics stack — but enough to run pilots.

Mature infrastructure is necessary, not sufficient. The next decision isn't how much CV — it's which CV.

What works: 4 use cases with positive ROI

The four scenarios below have public LATAM case material with defensible numbers. The order reflects ease of implementation, not necessarily the size of the payoff.

#1. Shelf monitoring and planogram compliance

The task: a camera above the shelf (or retrofitted onto existing CCTV) flags out-of-stocks, broken planograms, and facings that don't match the brand-partner agreement. The alert hits Telegram, Slack or an Odoo ticket for the replenishment crew.

Why it works in LATAM:

• Out-of-stock rates in LATAM supermarkets are chronically higher than in the U.S. or Europe. Walmex and Soriana have publicly discussed 8–12% OOS in FMCG categories; Walmart U.S. sits closer to 5%.
• Every percentage point of OOS reduction is worth 0.5 to 1% of category sales. In snacks and personal care that's $1–3M per year at a large store.
• Brand partners (Unilever, P&G, Nestlé, Colgate) pay CPG vendors like Trax, Pensa Systems and Focal Systems for audit data. Trax bills both sides — retailer and brand. That's a rare CV-vendor business model with two billing pockets.

Hard numbers: Trax pilots with major Brazil retailers, per the vendor's own published material, dropped OOS by 25–35% in dry goods. Coca-Cola FEMSA runs visual audits inside OXXO. Pensa Systems went into Walmart Mexico as a POC partner.

When it fits: stores larger than 500 m², more than 5,000 SKUs, frequent assortment rotation. If you run a 800-SKU mini-market with one juice shelf, use Excel.

#2. Loss prevention (shoplifting detection)

The task: floor and checkout cameras run CV models that flag suspicious patterns — an item dropped into a bag without scanning, basket-versus-receipt mismatches, return fraud, sweethearting (cashier waves merchandise through for friends).

Vendor landscape: Veesion (France, active in LATAM), Everseen (Ireland, global Walmart contract), Standard.ai (formerly Standard Cognition), and several local integrators in Mexico and Chile.

Why it pays: LATAM shrink, per crossed estimates from several retail consultants and reports from local supermarket associations, runs 1.7 to 2.3% of revenue. For a large retailer that's $80–200M a year. Cutting it 15 to 25% leaves $12–50M in EBITDA. With $1–3M of annual CV spend, ROI lands at 6× to 25×.

Walmart U.S. expanded checkout CV in 2023. In Chile, Cencosud (Jumbo, Paris, Easy) ran loss-prevention pilots with local integrators. Large Mexican retailers are testing too — no one talks publicly, and you can guess why.

#3. Queue analytics and footfall

The task: cameras count people in the store and in checkout queues, measure dwell time by category, and track entered-vs-purchased conversion. A real-time dashboard tells the store manager when to open a register, where to move a cashier, and where to relocate the promo end-cap.

Vendor landscape: RetailNext (U.S.), V-Count (Turkey, expanding into LATAM), local integrators in Mexico and Colombia, and Vodafone Business Retail for large-format chains.

Why it works: people-counting doesn't demand ultra-precise CV — 95% accuracy is out-of-the-box. The benefit is measurable: NPS up 5 to 10 points after rolling out dynamic cashier deployment; per-category conversion up 3 to 7% once dwell time gets optimized.

Coppel Mexico, Walmex and Falabella Chile have publicly discussed people-counter rollouts — though they don't always clarify whether the system is full CV or a simpler IR sensor.

When it doesn't fit: small stores with low traffic (under 200 visits/day). The instrumentation overhead eats the effect. If you're a corner store in Lima, this isn't for you.

#4. Self-checkout and smart cart

The task: the customer drops items into a smart cart (or simply walks out in "just walk out" format) and CV builds the receipt automatically.

Where it actually stands in 2026:

• Amazon Just Walk Out: Amazon pulled Just Walk Out from Amazon Fresh in the U.S. in 2024, keeping the tech for Whole Foods Smart Cart and third-party deployments (stadiums, airports). Details on About Amazon.
• AiFi: dozens of stores deployed globally, including the Aldi partnership in Europe and Carrefour pilots.
• Carrefour Brasil publicly piloted Scan&Go and smart-cart formats.
• Sam's Club (Walmex) has run app-based Scan&Go for years — not pure CV, but pointing the same way.

Honest caveat: Just Walk Out is the most capital-intensive use case. Public vendor pricing: $300–800k per mid-sized store. Not for a typical LATAM PYME. Makes sense for premium format, high-traffic convenience, or mall stores.

What doesn't work (or works poorly)

The three scenarios below circulate in every pitch deck and almost never close on ROI. Worth naming them before they get sold to you.

Facial recognition for demographic marketing

"Camera at the entrance flags the visitor's age and gender and the in-store screen shows contextual ads." Sounds great. In practice, three problems:

1. LGPD (Brazil), Law 25.326 (Argentina), Habeas Data (Colombia, Peru) and LFPDPPP (Mexico) require explicit consent for biometric data. A door sticker with fine print "by entering you accept facial recognition" doesn't qualify as consent in most jurisdictions. Brazil's ANPD has already issued fines on that exact scenario.
2. Age and gender accuracy drops in LATAM versus the U.S. or EU because training datasets are skewed. On Mexican, Peruvian or Bolivian populations, off-the-shelf pre-trained APIs miss 8 to 15% more often. The bias is documented — MIT Media Lab and the Algorithmic Justice League have papers on it.
3. The commercial effect is unproven. Pernod Ricard and L'Oréal ran pilots like this in Europe and published no ROI numbers afterward — which is a signal in itself.

Verdict: don't do it. If you need demographic segmentation for marketing, pull it from loyalty-program data, not from face recognition.

Theft prediction "before the fact"

"The model predicts that this person is going to steal based on micro-movements and pose estimation, and the guard intercepts." This is a bias factory. Any model trained on historical detention records inherits the bias in those records — and in LATAM that bias is typically racial and socioeconomic. The legal risk (CIDH precedent, local habeas data) and the reputational risk both exceed any ROI.

Several vendors sold this in 2021–2022; most quietly renamed the product to "behavior analytics" or "anomaly detection." If someone pitches it, ask hard questions: training data, third-party audit, independent validation.

Cheap "pilots" on a single camera

"Let's drop one camera into one store and see." That's not a pilot — that's a demo. It won't give representative data, won't let you measure ROI, and won't cover real operations (guard rotation, peak hours, seasonality, real-time traffic). A "pilot" like that for $5–10k will close in three months with the conclusion "doesn't work" — when the truth is that the design was never going to work.

Defensible minimum pilot: 3 stores, 8 weeks of baseline plus 8 weeks of pilot, 20 to 40 cameras, a fixed hypothesis and a defined success metric. That's $50–150k. Below that, there's no signal.

Common implementation mistakes

When a CV project fails in LATAM, the root cause almost always lives in one of these five. If you're hiring an integrator or evaluating a vendor, use this list as a filter during the pitch.

#1. Buying a model, not an operational process

"We bought the shoplifting detection system — it handles it." No. CV produces a signal. After the signal you need a guard, a reaction-time SLA, an escalation flow, a false-positive handler, and reporting for law enforcement. Without the process, CV is an expensive logger.

#2. Not pricing the cost of a false positive

The model fires "possible theft," the guard approaches, the customer isn't a thief. If this happens twice an hour at each store, customer NPS breaks before the shrink savings cover the investment. Measure precision and recall weighted by business impact — not bare accuracy.

#3. Not integrating with the ERP and POS

The OOS alert lands in Slack. And who turns it into a purchase order to the supplier? Replenishment lives in SAP, Odoo, or Oracle Retail. If the alert doesn't become a task with a deadline and an owner, it's useless. Integrating CV with inventory and procurement (the sale + stock + purchase modules in Odoo, for example) is 30 to 40% of the scope of any serious retail CV project. Not 10%, not "we'll do it later."

#4. Picking a vendor on the pitch deck, not on retraining cost

CV models drift. New SKUs ship every month in FMCG, packaging changes, store lighting changes (new LEDs → different color balance). Retraining cost at the 6-month mark can exceed initial deployment. The vendor should bake it into the price — or you should bake it into your TCO.

#5. Skipping baseline measurement

"We launched the pilot and sales went up 4%." Versus what? Versus the non-pilot stores? Controlling for seasonality? Without 8 weeks of baseline and a control group, any ROI claim is theater, not measurement. Per McKinsey Retail, 60% of "successful" retail-AI pilots don't reproduce at scale — and the cause is usually weak experimental design.

Anonymous case: pricing intelligence + visual catalog in beauty retail

A mid-size beauty-retail group in the Andean region, ~120 stores, 8 local plus 3 international brands. Objective: dynamic pricing driven by competitor monitoring, plus compliance with promo conditions agreed with the brands.

Before the project: a 4-person pricing team compared competitor prices manually every two weeks via screenshots. Brand promo conditions (e.g., "product X must be at eye-level for the full September campaign") were checked through random audit once a month. Non-compliance: ~23% of stores, which triggered fines from brands and lost promo-rebates.

What got built:

1. A web-scraping pipeline against 12 competitor e-commerce sites plus visual search across product images to match SKUs across sites with different SKU IDs (visually the same lip liner).
2. A visual merchandising audit on top of the stores' existing CCTV: the model checks that on the right day and in the right zone, the right product is staged per planogram. If not, a photo plus an automatic Odoo ticket for the owner.
3. A dynamic pricing engine in ML, recommending price changes based on competitor positioning and per-category elasticity.

Six-month results:

• Promo compliance: from 77% to 94%.
• Recovered brand rebates: $480k in the half.
• Margin uplift in priced categories: +1.8 percentage points.
• ROI: 6.2× on infrastructure investment plus 18-month opex.

This isn't magic. It's combining CV (visual matching, planogram check) with standard retail analytics and operational process. Magic is when a vendor promises all four use cases at once for $20k. If they promise that, it isn't going to work.

How to size ROI before you start

Before you sign the vendor SOW, run the back-of-the-napkin. If the numbers don't pencil out here, they're not going to pencil out in production.

1. Baseline problem cost. Shrink × store revenue (loss prevention). OOS-rate × lost sales (shelf). Wait-time × abandonment rate (queue).
2. Achievable percentage reduction. What public case studies show as realistic: 15–30% on shrink, 20–40% on OOS, 10–25% on abandonment. Take the lower bound.
3. 3-year system TCO. Hardware (cameras, edge) + software license + integration + retraining + operations. And change management: 15–25% of cost, never zero.
4. Payback period. Past 18 months: revisit scope. Between 6 and 12: workable. Promised 3 months: the vendor is lying.

If your retail operation is still building the data foundation, don't lead with CV — lead with an Odoo/ERP audit. Without clean inventory, sales and promotion tables, the CV signal never becomes action.

Close: when to start and how

Computer vision in LATAM retail has become a serious technology. But it's not magic and it's not a lift-and-shift from the U.S. What pays: shelf monitoring, loss prevention, queue analytics and smart cart — roughly in that order of complexity. What doesn't: demographic face recognition, theft prediction, single-camera demos.

If you're launching a project this quarter, pick one use case with the clearest ROI (for most PYMEs that's shelf monitoring), run an honest pilot in 3+ stores with baseline and control, and integrate with the ERP from day one — if the signal doesn't become action, you've bought an expensive recorder. For concrete architectures of computer-vision pilots in retail and restaurants, the services page lists public cases with numbers.

Want a second opinion on a concrete CV project for your retail operation? Write to us and we'll walk through whether there's real ROI or the vendor is selling a mirage. And if it's better to first audit your current Odoo/ERP, do that: it usually shows where CV adds value and where the data foundation needs fixing first.

FAQ