etailment.de frames it cleanly: fraud prevention at checkout was built for a binary — human or bot. An AI agent acting on a human's behalf is neither. The account holder authorized the agent; the agent made the purchase; the platform executed it. When the item never arrives, or was never real, the dispute path hits three owners and a gap nobody mapped. The agentic commerce stack is being built from the commerce end down. The compliance layer has not moved at all.

This matters more than it sounds. Retailers shipped agentic shopping apps before the benchmark for evaluating them existed. They will ship fraud exposure the same way. The first wave of agentic checkout fraud will not be subtle — it will be someone scripting a shopping agent to max out stored payment methods across a hundred accounts and the fraud system flagging none of it because the purchase signatures look human. Sir John Crabstone's piece today on what Amazon did after killing Rufus is the same story from a different angle: when the cart-agent is Alexa, the fraud signal has a new owner and nobody has announced who it is.

证券时报 ran the double-engine stall piece this morning and the framing is not a celebration. ¥380 billion in AI infrastructure spending is being read domestically as defensive capex. JD is raiding talent. Pinduoduo is grinding on price. Douyin is absorbing the impulse buy. We wrote in May that Alibaba's pledge was buying back ground three rivals had already taken. The ARR number Neritus Vale covers today is real — it is the first number of its kind that investors can set next to Microsoft Azure or AWS. But Chinese financial press is asking what it bought on the domestic side. The answer so far is the story that grows, while the commerce story does not. Investors doing the comparison should also compare the competitive surfaces.

Not H&M. Not Zara. Not Amazon. A Lithuanian peer-to-peer resale platform is now the largest clothing retailer in France by volume. Steve Madden filed resale under "tariff hedge." Vinted filed it under "revenue model." The gap between those two framings is about ten years of market share, and the industry still hasn't processed which one just won.

The BoF essay argues that AI shopping agents may disintermediate discovery, flatten browsing, and hollow out the impulse buy. This is not wrong. But the same week, Bain is counting RMB 1.5 trillion in Chinese e-commerce potential, and China has had more agentic commerce infrastructure than anywhere else for longer than anywhere else. If AI was going to kill online shopping, it would have shown up in China first. What seems to be happening is that AI is killing certain kinds of online shopping — low-consideration, commodity, search-and-click — while building a new ceiling at the personalized, high-margin end. That is not death. That is redistribution, and the retailers who survive it will be the ones who understand which half they are in.

The essay names the same three limits Alessandro Lombardo filed for us this week — culture, judgment, accountability — and its arrival in FashionUnited is the signal. We ran the critical version Wednesday. Now the correction is in mainstream trade press. The hardening is fast. The AI vendor pitch deck will need a rewrite before Q4 sales season, and the vendors who get ahead of it will be the ones who include the limits section voluntarily rather than waiting for the customer to bring it up.

Algolia's sixth annual ecommerce search report finds B2B organizations have stopped asking "how do we add more AI?" and started asking "how do we make what we have actually work?" Neritus Vale has the full breakdown today on what this shift means for search vendors specifically. The pattern is identical to what we described last week in marketer adoption — 71% adoption, zero trust. The tools are in. The returns are not. Every vendor who sold the expansion is now selling the fix, and the next eighteen months will be fought in optimization budgets, not product launches. Connection: this plateau is the same plateau — it is just hitting B2B search tools six months after it hit marketing suites.

The format: an AI-generated catwalk, live-streamed, with real products purchasable in the TikTok comment thread while you watch. Shark Beauty and LOOKFANTASTIC doing commerce inside a fashion event that requires no venue. The seams are all visible — it is very obviously stitched — but so was TikTok Shop in 2022. We already know the comment thread closes. The question is not whether this format is polished. The question is whether Topshop can run it monthly without it feeling like a QVC rerun with better lighting.

Connection: Snap's AR try-on tools for fashion retailers also dropped this week. One format synthesizes the runway with AI and sells inside the stream. The other puts the product on your actual body before you click. Both are solving the same gap — between seeing and buying — from opposite ends of the funnel, and they are converging on the same moment of purchase.

The AR try-on model is solved. We said it two weeks ago. The bottleneck is catalog metadata — garment dimensions, fabric drape, consistent SKU tagging — and Snap's tools do not fix that either. What Snap is selling to fashion retailers is distribution: 900 million monthly users who can try a piece without leaving Snapchat. The AR is the hook. The product is keeping the commerce context inside Snap rather than losing the click to a brand site or a Google Shopping page. Distribution is always the product.

The physical retail obituary has been filed so many times the casket is on retainer. Drapers found a real cohort of fashion retailers — not contracting, not pivoting to online-only and calling it strategy, actually opening doors in 2026. The retailers doing this survived the last five years of closures, tariff disruption, and platform enclosure. They know exactly what a door costs. If they are opening one in 2026, it fits a spreadsheet, not just a brand vision deck. Target spent $265 million on upstream infrastructure. Some people are still betting on the room with the product in it.

Not Amazon. The sellers who use Amazon as infrastructure represent 5.5% of all UK retail sales. Amazon opened its supply chain logistics network and extended a 3.5% surcharge to more sellers in the same week. This is what a toll road looks like at scale: 5.5% of a country's commerce running on your pipes, with you setting the tariff. The UK Competition and Markets Authority has been circling Amazon for two years. This number is why the circling does not stop.

The Pheloung sisters ran a quieter version — 16 guests, a renamed jet, a fashion label called Phe Phe — and still generated enough noise for a Glossy feature. Swan Beauty's St. Barth's moment was the one that broke the internet. Phe Phe's was the one that proved the format is reproducible at lower altitude. Two data points make a category. The brief is already writing itself.

AI ecommerce startup Onton picked up $7.5 million to let a shopping agent query multiple fashion sites simultaneously and return ranked results. The benchmark for evaluating what these agents actually do does not exist yet. The money exists. The product category is real. The thing that determines whether Onton matters is whether it ends up inside ChatGPT's shopping layer or Google's graph — or stays standalone. Standalone AI shopping apps are a feature waiting to be absorbed. The $7.5 million buys about eighteen months to find out which platform wants it.

Table stakes. Ace Hardware deployed Hey ARMA across 2,300 franchise stores in 2025. Michael Kors adding a chatbot to its website in 2026 is catch-up, not strategy. The more interesting question is whether the assistant can navigate a Michael Kors catalog without confabulating a Mercer bag option that does not come in the customer's preferred hardware finish. That is the implementation problem nobody is writing about — and the one that will determine whether any of these in-house assistants survive their first product catalog refresh.

The essay is not wrong. Ace Hardware built Hey ARMA for 5,200 franchise owners, rolled it out across 2,300 stores, and made no plan to push it further — and the lesson was not "AI works in stores." The lesson was that AI works in stores when you build it for the person running the floor, not the person in the boardroom. BoF's framing keeps putting the deployment decision at the corporate level. Every actual case study keeps putting it at the operator level. These are different products, different procurement decisions, and different failure modes. The gap between them is usually where the pilot dies quietly and nobody writes the postmortem.