Agentic commerce had a split-screen day on November 17, 2026. In the morning, retailers shipped agents: Zalando announced Zalando Extended, a service that will have its AI shop and buy from other retailers on the customer’s behalf, and Commerce.com rolled out Shopping Agent for BigCommerce storefronts ahead of the holiday season. The same day, the payments industry shipped armor for a threat consumers already fear more than they want the technology: Spreedly and Riskified announced a partnership that puts fraud decisioning in front of every agent-placed payment, and published survey data explaining exactly why the product exists.

Read together, the three announcements are the most honest picture of agentic commerce yet. The supply side has committed: agents will browse, choose, and buy. The fraud stack has committed: agent transactions will be screened before money moves. The consumer has not committed at all, and the reason is not capability. It is trust, with numbers now attached to every layer of it.

The demand side dug in: 52 percent name fraud as the blocker

Spreedly’s State of Checkout 2026 survey, conducted by Talker Research among 2,000 U.S. adults, put a hard number on what has been anecdote for two years. Fifty-two percent say fear of fraud is the top reason they will not let AI shop for them. Not privacy, not job displacement, not “I enjoy shopping”: fraud. The person fears being deceived by an invisible actor at the moment money leaves their account.

The second number is quieter and more corrosive. Thirty-one percent of respondents do not know who is responsible when an AI purchase goes wrong. If the agent buys the wrong item, or buys from a fraudulent seller, or buys a product whose five-star rating was manufactured, who eats the loss? The merchant? The platform? The agent operator? The card network? The consumer? Nearly a third of adults have no model of accountability, and they are not wrong to be confused, because the industry has not answered the question either.

Riskified’s own research sharpens the point from the merchant side of the table: only 16 percent of consumers are comfortable with an AI agent spending their money autonomously, and 54 percent believe AI increases fraud risk. Those two figures are nearly mirror images of each other. The belief that AI expands the attack surface and the refusal to grant spending authority are the same instinct expressed twice, and no amount of agent capability changes it.

For context, this tracks with what we documented in our analysis of the first agentic Black Friday datasets: agent traffic is exploding at the browsing layer while checkout conversion stays in the low single digits. The funnel narrows to almost nothing exactly where trust has to change hands.

Cyber Five already showed what agent-scale fraud pressure looks like

The fraud fear is not hypothetical, and Spreedly’s own transaction data is the evidence. Across aggregated, anonymized order flow from its platform, declines flagged for suspected fraud averaged 31 percent higher per day during the 2025 Cyber Five weekend than on a normal November day, while payment attempts rose 20 percent.

Sit with that ratio. Fraud-flagged declines grew half again as fast as payment volume during the highest-stakes retail week of the year. That is the signature of a fraud ecosystem that scales with transaction volume faster than ordinary commerce does, and the 2026 holiday season arrives with AI agents in the funnel for the first time at scale. Every agent that can buy is, from the merchant’s perspective, a buyer whose behavior patterns look nothing like a human’s: faster, more parallel, less predictable. Riskified built AI Agent Intelligence for precisely this, giving merchants visibility into which orders were placed by AI shopping agents and a policy builder to decide how those agents get treated: approved, declined, or stepped up to additional verification.

The Spreedly-Riskified integration puts that decision inside Spreedly’s Composer orchestration flow, before any payment is processed, without merchants changing their payment providers. It is a genuinely sensible architecture. It answers one question extremely well: is this transaction legitimate?

Zalando’s countermove: become the gateway, not the shelf

While consumers named fraud as the barrier, Zalando announced a strategy that treats trust as its home-field advantage. Zalando Extended inverts the retailer’s traditional role: instead of only selling what it stocks, Zalando’s AI agents will complete purchases from other retailers on the customer’s behalf. The first partner is Westwing, the premium home and living retailer. Westwing products already surface in Zalando search results, but today the customer clicks through and buys on Westwing’s site. From 2027, Zalando’s agent will buy the sofa or the lamp itself, using the customer’s existing Zalando account and payment details, without leaving the Zalando app.

“We’ll buy it on your behalf,” chief technology officer Graeme Smith told Vogue Business at Web Summit in Lisbon.

The mechanics matter for anyone building in this space. Retailers join through an integration Smith describes as barely more than an affiliate product feed plus a commercial agreement. No marketplace listing, no inventory integration, no API rebuild. Zalando is betting that in the agent era, the entity that owns customer identity, payment credentials, and preference data becomes the gateway through which all other retail flows. The company brings real assets to that bet: more than 63 million active customers, a personalization stack it calls Lifestyle AI, over 10,000 size charts fed by tens of thousands of body measurements per week, and 22 million customers who engaged with Zalando Assistant this year. It is also hedging across the whole board, with planned checkout inside Google’s AI Mode and the Gemini app when agentic commerce launches in Europe, and advertising tests inside ChatGPT in the UK.

Smith’s defense of the trust position is straightforward: an 18-year-old European brand operating under European data protection standards knows the customer in ways a general-purpose agent cannot. “No third-party agent has got that,” he said of the fit data.

BigCommerce put the agent on the storefront itself

The same day, Commerce.com (Nasdaq: CMRC) took the opposite architectural position. Shopping Agent for BigCommerce is a customizable, AI-powered storefront assistant that merchants brand and control, now rolling out to eligible merchants before the holidays. It personalizes discovery, supports comparisons, carries context across follow-up questions, and drives “conversational add-to-cart and checkout guidance.” Behind it, a new conversational storefront search combining keyword relevance with semantic understanding is now available to all paid BigCommerce plans. A B2B Agent is in early access with retailers including AS Colour, automating purchase-order workflows with support for negotiated pricing and complex catalogs.

“Shopping Agent gives BigCommerce merchants a way to bring that guidance into their storefront through an assistant they can make their own,” said Sharon Gee, senior vice president of product for AI at Commerce, “while also delivering rich first-party signal to the merchants from agentic conversations.”

Note the design center. Zalando’s agent and BigCommerce’s assistant share a goal: conversion on the merchant’s own terms. The storefront assistant exists to turn conversations into purchases on that storefront, and its byproduct is first-party signal for the merchant. That is a legitimate business design. It is also, structurally, a referee with a jersey in the team’s colors, a point this blog has made about every sponsored-answer architecture from Alexa to AI Mode.

What the November 17 stack solves, and what it does not

Line up the day’s launches and you can see the emerging agent-commerce architecture with unusual clarity:

  • Discovery and choice: Zalando Extended, Shopping Agent, and the general-purpose agents behind them
  • Transaction legitimacy: Riskified AI Agent Intelligence plus Spreedly orchestration, screening every agent order before payment
  • Payment execution: stored credentials at Zalando, wallets at the platform layer
  • Accountability: unanswered, by 31 percent of consumers’ own admission

The transaction layer is being armored at serious speed, and it deserves credit: fraud decisioning for agent orders is a real problem solved by real infrastructure. But look at what every layer consumes and none of them verifies: the product itself. When Zalando’s agent buys a Westwing lamp on your behalf in 2027, the product facts it reasons over come from the affiliate feed and the merchant’s own listing. When a BigCommerce Shopping Agent recommends the merchant’s SKUs, the “relevant product details” come from the merchant. When a general-purpose agent summarizes reviews before checkout, it summarizes whatever reviews exist, authentic or purchased.

This is the exact failure mode the FTC’s fake-review enforcement era has documented for years: listings whose ratings are manufactured, review pools seeded with incentives, average scores that measure marketing spend rather than product quality. An agent that consults a manipulated listing before buying does not launder the manipulation into truth because a payment fraud engine approved the transaction afterward. The fraud stack answers “was this order placed by a legitimate buyer?” No component of the November 17 stack answers “was this product worth buying?”

The responsibility vacuum is a product-trust vacuum

The 31 percent who cannot say who is responsible when an AI purchase goes wrong are pointing at the real architectural hole. Fraud has an answer being built this week. Wrong product has nothing: no independent signal between the agent’s recommendation and the buy button, no persistent quality record the agent can consult that the seller cannot edit.

That is the layer GoBuy exists to be. An independent product-evidence checkpoint that any agent can call before money moves:

  • Smart Score 0 to 100, computed from the quality of review evidence rather than the quantity, so a thousand incentivized five-stars cannot outvote a pattern of verified failure reports
  • Fake-review filtering applied before scoring, so manipulated input never reaches the output the agent consumes
  • A shortlist, not a flood: the top 7 products per category, because an agent’s value is curation, and a trust layer that returns a thousand SKUs has returned nothing
  • GoBuy Verified, a badge reserved for products holding a Smart Score above 80 across 90 days, which converts “trust” from an adjective into a checkable state
  • An MCP endpoint at gobuy.ai/api/mcp, so shopping agents consult GoBuy as a standard tool call during evaluation, the same way they now consult payment rails at checkout
  • A Chrome extension that injects the trust panel directly onto Amazon product pages for the human who still shops the old way

The pattern of this week generalizes: every time agentic commerce ships a capability, it immediately requires a verification layer that does not ship with it. Payments shipped, so fraud decisioning had to be built. Agents will now choose products, so product evidence has to be built. The only question is whether it gets built independently, or whether the same platforms that sell the products get to grade them.

The season that decides it

Black Friday is nine days out. Zalando is right that gateway trust compounds, Riskified is right that agent orders need screening, and BigCommerce is right that merchants want agent conversations on their own storefront. All three are building for a consumer their own data says is not ready: 52 percent afraid of fraud, 84 percent unwilling to let an agent spend autonomously.

The fastest way to close that gap is not a better agent. It is giving the agent something to check against that no seller controls. If you run shopping agents, wire product evidence into their evaluation loop before this holiday season starts: agent documentation is here. If you are a human who still reads your own reviews, the trust panel is here.

The payment layer got armored this week. The product layer is next, or the 31 percent who do not know who is responsible are about to find out the hard way.