Agentic commerce just drew its first visible fault line. CNBC reports that OpenAI is phasing out Instant Checkout after Walmart found that purchases completed inside ChatGPT converted at one-third the rate of transactions where users clicked through to Walmart.com. At the same moment that retreat was announced, Walmart confirmed it was embedding Sparky — its proprietary AI shopping assistant — directly into ChatGPT, routing discovery through the platform while keeping the transaction on its own infrastructure. The split is now structural: the conversation belongs to the platform, the checkout belongs to the brand, and fashion retailers who have not built assets on both sides of that line will find themselves welcome in neither.

What Walmart’s data actually proved

Walmart tested Instant Checkout across approximately 200,000 products. Daniel Danker, Walmart’s EVP of Product and Design, described the in-chat purchase experience as “unsatisfying” — a word that names a structural problem, not a design one. A conversational interface optimized for open-ended exploration does not naturally resolve into the confirmations, substitutions, and corrections that retail checkout requires. The Sparky-via-ChatGPT architecture that replaced it — ChatGPT handles the query, Sparky handles search and selection, Walmart processes the order — now converts at roughly 70% of Walmart.com’s own rate. That is still a penalty. It is not a disqualifying one, and the difference between one-third and seventy percent is the difference between owning your checkout layer and ceding it.

OpenAI’s move is not a retreat from commerce. The forthcoming model treats purchases as merchant-run apps callable from within ChatGPT, with the retailer’s own systems handling the transaction. What OpenAI is abandoning is the position of being the checkout infrastructure. What it is keeping — and building toward — is being the conversation infrastructure through which shoppers discover and intend to purchase. Fashion retailers who read this as “AI shopping failed” are misreading the signal entirely.

Meta is building a third model that fashion brands may not have noticed

While OpenAI and Walmart were separating the conversation layer from the checkout layer, Meta was constructing a scenario in which neither distinction applies to its platform. At Shoptalk 2026, TechCrunch reports, Meta unveiled AI-generated product summaries — a scannable overview of pros, cons, and use cases, potential discounts, brand context — appearing as pop-up overlays when Instagram or Facebook users click ads. Meta also announced testing of a single-tap checkout built on Stripe and PayPal, with Adyen and Shopify integrations planned. The platform is not routing discovery through a third-party AI assistant. It is building the intelligence layer, the checkout layer, and the distribution channel as a single vertically integrated surface. For fashion brands that have built Instagram as their primary acquisition channel, this announcement warrants attention that it has not yet received. Meta is not adding a feature; it is closing a loop.

The comparison TechCrunch draws — to Amazon’s AI-generated review summaries introduced in 2023 — understates the ambition. Amazon built those summaries inside a commerce environment where it already owned checkout. Meta is building them inside a social environment where brand spend drives the discovery, then using that context to capture the transaction it previously handed off to external shops. Fashion brands that pay for Instagram acquisition are now funding a system that converts within Instagram.

The infrastructure gap that makes all of this academic for most retailers

The European analysis sharpens the supply-side picture. A February 2026 piece by Dr. Frank Becker of Kyndryl Germany in etailment.de cites the Kyndryl Retail Readiness Report 2025, which surveyed 114 global retail decision-makers and found that only 15% have integrated omnichannel systems. Fifty-six percent report data locked in legacy infrastructure they cannot access. The minimum viable requirement for participating in any agent-mediated commerce — Sparky-on-ChatGPT, Meta’s pop-up intelligence layer, or the next variant — is the ability to surface clean, real-time product data, inventory, and pricing to an external system on demand. Most of the industry cannot do that.

An Omnisend survey from summer 2025, cited by Becker, found that 59% of US consumers already use generative AI tools for shopping, and that hesitation toward AI-assisted transactions fell from 66% to 32%. Consumer adoption is not the lagging variable. Retailer infrastructure is.

The counter-argument, and where it runs out

Fashion is not grocery. The objection is reasonable: discovery in apparel is exploratory and emotional where staple repurchase is systematic and specification-driven. A shopper asking ChatGPT for a winter coat for a coastal wedding is doing something fundamentally different from reordering paper towels. If conversational discovery is where fashion shoppers naturally operate, the argument runs, then the Walmart conversion penalty would be smaller in apparel than in general merchandise — and the case for investing in the conversation layer grows stronger, not weaker.

This is not wrong. It is also not an argument against the thesis. If conversation is where fashion discovery happens, fashion brands need agent-native assets — inventory-connected, brand-controlled AI presences — available inside the platforms that own the conversation. That is precisely what Walmart built with Sparky, and what Amazon has declined to build by staying entirely out of third-party AI assistant integrations. That is a defensible choice for a retailer with Amazon’s direct traffic. Fashion brands without comparable owned audiences do not have the same option.

The position to take

Walmart’s architecture — own the transaction, rent the distribution — is now a tested template. If an AI platform generates discovery at scale but converts poorly when it also handles checkout, the answer is not to abandon the platform. The answer is to embed a branded agent within it that routes checkout back to your own infrastructure. Walmart in discussions with Anthropic about bringing Sparky to Claude suggests this is a deliberate multi-platform strategy, not a one-off integration.

Fashion brands have two near-term decisions. The first is whether to build or license agent assets capable of operating inside third-party conversation platforms — the Sparky model. The second is whether to treat Meta’s new commerce intelligence layer as a feature update or a fundamental change to how Instagram acquisition economics work. Neither decision requires predicting which AI platform wins the conversation layer. Both decisions require acknowledging that the conversation layer and the transaction layer are now distinct infrastructure problems, and that being absent from either one has visible costs.