Amazon quietly pulled Just Walk Out from most of its Fresh stores in 2024 — a good starting point for separating retail AI that actually works from the version in the marketing copy.
In April 2024, Amazon quietly pulled its "Just Walk Out" cashierless checkout technology out of most of its Amazon Fresh grocery stores. The system had been marketed since 2018 as a triumph of computer vision — cameras and shelf sensors tracking what you picked up, no scanning, no line, just walk out and get charged automatically. What actually shipped, according to reporting by The Information and Business Insider, leaned heavily on a review workforce in India watching video footage to verify a large share of transactions the vision models couldn't confidently resolve on their own. Amazon didn't kill the technology — it still runs in stadiums, airports, and some standalone Amazon Go stores — but it swapped Just Walk Out for cart-mounted "Dash Cart" scanners in most Fresh locations, a quieter, cheaper, more deterministic approach to the same problem.
That's a useful starting point for a piece about AI in retail, because it's the opposite of the usual narrative. Most coverage of AI in retail is either breathless (AI will "personalize everything") or dismissive (it's all chatbots and recommendation widgets). The real picture is a set of specific, load-bearing systems — some running in production at scale for a decade, some genuinely new and still finding their failure modes — sitting next to a few high-profile deployments that got walked back once the economics of always-on human review met the marketing copy about "AI-powered" autonomy.
Cashierless checkout gets a disproportionate share of retail-AI headlines because it's the most visible to a shopper, but it's arguably the least mature of the four areas covered here. Amazon isn't alone in the "harder than it looked" bucket — Zippin and AiFi both sell computer-vision checkout systems to smaller grocery and convenience chains, and both quietly emphasize "AI-assisted" rather than "fully autonomous" in more recent marketing than they did in 2019. Instacart acquired Caper Cart (smart shopping carts with built-in scales and cameras) in 2021 and has scaled that hardware out to hundreds of stores — notable because it doesn't try to eliminate scanning, it just moves it earlier and adds weight-sensor verification, which sidesteps the hardest part of the vision problem (occlusion — a shopper's hand or another item blocking the camera's view of what's being picked up).
The pattern across cashierless checkout: pure computer-vision attribution of "who picked up what" at open-shelf scale is still an unsolved problem when items are visually similar (two flavors of the same yogurt), when carts get crowded, or when a family shops together and hands products between people. Systems that route around the hardest cases — sensor fusion, weight verification, or a lightweight scan step — are shipping and staying; the vision-only, fully autonomous versions have mostly become vision-assisted, human-reviewed ones.
Where retail AI has been running the longest and causing the least controversy is demand forecasting and inventory optimization — because it's invisible to shoppers and the ROI math is straightforward: every stockout is a lost sale, every markdown is a margin hit, and both are measurable.
Walmart has used a proprietary forecasting system (internally referred to in earnings commentary and technical talks as "Element") for years to predict demand at the store-and-SKU level, feeding replenishment and labor scheduling. Zara's parent company Inditex pairs RFID tagging at the item level with machine-learning demand models to decide what to restock in-store versus fulfill from a distribution center — a big part of Zara's famously fast supply chain is knowing, within a day, which specific stores are selling through a style versus sitting on it. These systems predate the current generative-AI wave by years; what's changed recently is less the core forecasting technique (gradient-boosted trees and neural forecasting models have been standard here for a while) and more the ability to fold in messier signals — social trend data, weather, local events — as LLM-based feature extraction gets cheap enough to run at SKU-level scale.
The honest limitation here is the cold-start problem: forecasting works well for products with sales history and breaks down for genuinely new items, which is exactly the case that matters most for fast fashion and trend-driven retail. No amount of model sophistication fixes the fact that a brand-new SKU has zero history to learn from — retailers still fall back to analogous-product heuristics and human buyer judgment for true novelty, and that hasn't changed.
Recommendation engines are the oldest production AI system in retail — Amazon's item-to-item collaborative filtering paper is from 2003 — and remain among the most economically significant. The commonly cited claim that a third or more of Amazon's revenue comes from recommendations traces back to a McKinsey estimate that's been repeated so often it's treated as fact; it's directionally plausible but not something Amazon has itself precisely disclosed, so it's worth treating as a widely-cited estimate rather than a confirmed figure.
What's more verifiable is the shift in how personalization gets built. Stitch Fix built its entire business model around algorithmic styling paired with human stylists reviewing the algorithm's picks before they shipped — a hybrid that the company has talked about publicly for years. That hybrid model matters because it's an implicit admission that pure algorithmic personalization has a ceiling: recommendation systems are good at "more like this" and bad at "this, but for an event next month with a specific budget and body type," which is exactly the kind of judgment call Stitch Fix pays humans to make on top of the model's shortlist.
The newer wrinkle is LLM-based conversational personalization — an assistant that asks clarifying questions instead of relying purely on past purchase history — which promises to fix the cold-start problem for new shoppers (no purchase history to work from) at the cost of needing the shopper to actually engage in a multi-turn conversation, which most people don't want to do for a $12 purchase. It's a genuinely better fit for high-consideration categories (furniture, electronics) than for routine reorders, where the old collaborative-filtering approach still wins on speed.
The newest category is AI agents that can complete a purchase on a shopper's behalf rather than just recommending one. OpenAI shipped Instant Checkout inside ChatGPT in 2025, initially integrated with Etsy and expanding to Shopify merchants, letting a user complete a purchase without leaving the chat interface. Perplexity has pursued a similar "answer, then buy" flow. Google has been extending Shopping into its AI Mode search experience, using Gemini to synthesize product comparisons and, increasingly, to hand off directly into checkout flows.
This is architecturally the same class of problem MCP and tool-calling solve on the developer side — an agent needs a structured, authenticated way to call "add to cart" and "checkout" on a merchant's behalf rather than screen-scraping a storefront, and merchant-side adoption is currently the bottleneck, not model capability. Shopify's and Etsy's early integrations exist because those platforms built the API surface for it; a long tail of retailers simply hasn't, which is why agentic checkout today covers a specific set of partnered merchants rather than "any store on the internet." The fraud and chargeback questions are also unresolved: when an agent completes a purchase on a stored card without a human clicking "buy" in real time, existing card-network dispute processes weren't built with that flow in mind, and issuers are still working out where liability sits.
| Application area | Maturity | What's actually AI vs. hype | Primary blocker |
|---|---|---|---|
| Checkout automation | Early / walked back in places | Real computer vision, but often paired with human review at scale | Occlusion and attribution accuracy at open-shelf scale |
| Demand forecasting & inventory | Mature, long-running | Genuinely automated, running for years at major retailers | Cold-start problem for new/trend products |
| Personalization & recommendations | Mature, evolving | Real revenue driver; LLM layer is additive, not a replacement | Ceiling on nuanced, high-consideration purchases |
| Agentic / conversational commerce | Early | Real capability, narrow merchant coverage | Payment/liability infrastructure, not model quality |
Two problems get far less attention than they deserve. First, biometric and video-based checkout systems have run into real regulatory friction — Illinois' Biometric Information Privacy Act (BIPA) has already produced litigation against retailers over palm- and face-based payment systems, and any vision-based checkout system that can plausibly identify individual shoppers by gait or face inherits that legal exposure regardless of whether identification was the intended use case. Second, generative AI has created a listing-quality problem that retailers are still fighting: AI-generated product descriptions and even AI-generated customer reviews have shown up at scale on major marketplaces, degrading the exact trust signals that recommendation and search systems rely on to work well. The tools that make it cheap to generate a plausible product listing are the same ones flooding the input side of the systems retailers use to rank and recommend those listings.
The takeaway: the retail AI that's actually working is the kind you don't notice — forecasting and recommendation systems that have quietly run for years and keep getting incrementally better. The kind that makes headlines — autonomous checkout, agentic shopping — is real technology solving a real problem, but it's also the newest and least proven, and the honest cases (Amazon's own walkback included) show it's still cheaper and more reliable in many settings to keep a human, or a simpler sensor, in the loop than to bet fully on vision-only autonomy.