A decision that was not on your board agenda eighteen months ago is now sitting on it, and the clock has started. On October 21, Walmart announced an Agent Gateway that lets approved AI agents, including ChatGPT and Gemini, place real orders against Walmart and Sam’s Club inventory. On the same day, by the same reporting, Amazon kept its door shut on outside agents. The two biggest retailers in the United States now carry opposite policies on the single question every ecommerce director will have to answer in the next twelve months.
I have watched this pattern before. In 1996 the question was whether to let a web browser read your catalogue at all. In 2008 it was whether to accept an iPhone-shaped shopper who scrolled with a thumb. Both times, the people who waited for proof paid for the wait in distribution. The AI agent question has the same shape with one new twist: the agent is a buyer that negotiates and substitutes. The “do nothing” position has a real cost, and that cost shows up as missed baskets you cannot see in analytics because your store never heard the request.
So I went and read what the people actually making this call are saying on the record, from Walmart, Amazon, Shopify, Google and Forrester. Six published positions. The disagreement is sharper than the hype coverage suggests, and the operational implications are not the ones trade press keeps summarising.
Named practitioners quoted below, each from a published source:
- Harley Finkelstein, President, Shopify, in a Modern Retail interview published January 22, 2026.
- Vanessa Lee, VP of Product, Shopify, in Shopify’s own newsroom post dated January 11, 2026.
- Ashish Gupta, VP and GM of Merchant Shopping, Google, in the same Shopify newsroom post.
- Daniel Danker, EVP of AI Acceleration, Product and Design, Walmart, speaking at the Morgan Stanley Technology, Media & Telecom Conference, March 2026, reported by PYMNTS.
- Andy Jassy, CEO, Amazon, on the Amazon earnings call reported by Modern Retail on October 31, 2025.
- Emily Pfeiffer, Principal Analyst, Forrester, in her “Predictions 2026” blog post published October 28, 2025.
If an agent asks your store a question, somebody is going to answer it
Shopify’s position is the most forward-leaning, and it comes with a receipt. Finkelstein told Modern Retail that Shopify’s merchants have seen “14x more orders coming from agents compared to last year at the same time.” That is a reported company figure, not a projection. If your competitor on Shopify has agent-sourced orders growing at that rate while your store does not even surface in agent catalogues, the gap is already open.
Walmart has moved from statements to infrastructure. The Agent Gateway announced at the Masters of Scale Summit on October 21, built on the Model Context Protocol, lets approved agents transact against Walmart and Sam’s Club inventory. TechCrunch paraphrased John Furner as expanding Walmart’s earlier work “to prepare for a world where AI agents become mainstream.”
Merchants have seen 14x more orders coming from agents compared to last year at the same time.
Harley Finkelstein, President, Shopify, in Modern Retail, January 22, 2026
What this group shares is a view that if the agent is there, you have to be legible to it, and the legibility is the lever. Vanessa Lee’s framing in the Shopify newsroom post is characteristic: “Agentic commerce has so much potential to redefine shopping.” Google’s Ashish Gupta, co-launching the Universal Commerce Protocol that same January day, said flatly that “the shift to agentic commerce will require a shared language across the ecosystem.” Both are company-side statements, both come with the caveat that the companies make money if the shift happens, and both carry an operational point: without a shared protocol, agents cannot transact against your catalogue at all.
If you let the agent in, decide what you are selling to it
The second position is less headline-grabbing and more interesting to a buyer. It is Walmart’s, through Danker, and it is about intermediation. Asked at the Morgan Stanley TMT Conference in March whether Walmart was giving up its customer relationship when an agent like Gemini handed off the basket to Walmart’s own Sparky assistant, Danker said, per PYMNTS: “I would argue it’s intermediation.” That is a one-line thesis about how the economics work. Walmart’s view is that an agent-mediated sale is still a sale with Walmart as the destination of the transaction, the fulfilment, the returns, and the long-term loyalty signal, even if the agent chose the SKU.
Finkelstein goes a step further. “If agentic commerce is going to take off, consumers need access to all their favorite brands, not just some,” he told Modern Retail. Shopify’s answer is to open its Catalog to non-Shopify brands, so an agent asking a Shopify-powered surface for a product can find it even where the brand sits off-platform. He also made the merchant-control point explicitly: “That’s the reason why, through the Shopify admin, you get to decide where your products show up in the agentic tools.” In procurement language: the first design question for any platform is who holds the switch that lets each agent in, and at what granularity.
If you do not let the agent in, say what you are protecting
Amazon is the counter-position, and it is a position worth taking seriously rather than reading as reactionary. On the Q3 2025 earnings call, as reported by Modern Retail on October 31, 2025, Jassy said: “We’re also having conversations with and expect, over time, to partner with third-party agents.” He added, in the same reporting: “But I do think that we will find ways to partner,” and “We’ve got to find a way to make the customer experience better and have the right exchange of value.” The language carries the position cleanly. Partnering is possible, but on terms that preserve the exchange of value, which in Amazon’s case includes the advertising surface and the recommendation layer that funds much of the business.
The Amazon stance is often reported as “blocking agents,” and operationally it is. The strategic point behind it is narrower. If an outside agent reads your catalogue, picks the SKU, and delivers the shopper to checkout, the margin from merchandising, sponsored listings and recommendation surfaces evaporates. For a store whose business model monetises attention inside the browse experience, letting agents in without a value exchange is a self-inflicted wound. For a store whose model is units sold at a margin, the calculus is the opposite.
Agent legibility is upstream of agent policy
The quiet point Finkelstein makes, and that gets lost in the open-door-versus-closed-door framing, is that neither position matters until your catalogue is readable by machines in the first place. In his Modern Retail interview he put it bluntly: “Merchandising is a real thing. You can’t just scrape information and expect a great experience.” And: “That is the way agentic commerce will take off, not by just focusing on a simple transaction.” Translated: a product record with a decent title, a good image and a price is not enough. The agent needs attributes, variations, availability, substitutions, policies, and a credible rating signal, in a shape it can parse without inferring.
This is the part a buyer should test before any platform promise lands. Ask the vendor to pull one of your SKUs through a working agentic surface, with your attributes and policies, and see what the agent renders back. Nine times out of ten the gap is not the protocol, it is the product data model underneath it.
True autonomy is further out than the demo suggests
The sceptical voice in this symposium is Forrester’s Emily Pfeiffer, and she deserves her own hearing. In her Predictions 2026 post, published October 28, 2025, she wrote: “True autonomy and broad use are still in the distance.” She also wrote that “some of these initiatives will pave the way to a completely new way of interacting,” which is the fair concession a serious sceptic makes. Her forecasts land somewhere between the two camps: “five major US or European brands will unify agentic commerce experiences” in 2026, “one-third of retail marketplace projects will be deserted as answer engines steal traffic,” and “twenty percent of B2B sellers will be forced to engage in agent-led quote negotiations.” Her B2B framing is worth the quote in full: “Buyer bots will negotiate prices and terms, establish replenishment cadence, and confirm compliance.”
The honest reading is that the agent story is real in the direction it points and overstated in the pace at which it is arriving. B2B is further along than consumer, because B2B carries contracts and quote cycles that reward exactly the kind of patient negotiation a bot can run. Consumer still has the trust problem: Forrester’s 2025 benchmark put ChatGPT usage at only 24% of US online adults. A 2027 roadmap should let both halves of that reading sit side by side.
What this looks like inside a real WooCommerce store
Here is the build view, because the policy debate above is theatre until something is shipped. A WooCommerce catalogue that wants to be legible to an AI shopping agent in 2027 needs five concrete pieces of work done, in roughly this order.
Product data model that resolves cleanly. Every variable product needs attributes modelled as attributes, not as free text in the description. A navy medium t-shirt has to come back as colour=navy and size=M, not as a line in a paragraph. The default Woo attribute system handles this; the work is going back through the catalogue and cleaning records written for human eyes. Budget this at roughly fifteen minutes per SKU for a mid-market catalogue, less with AI-assisted bulk edits. Nine times out of ten this is the highest-leverage piece.
Structured data that matches the catalogue. Product schema with offers, aggregateRating, brand, GTIN or MPN where you have it, availability, and shippingDetails. The schema sits in the page HTML and is what both search crawlers and agent crawlers read first. On WooCommerce, the honest answer is that the out-of-the-box schema from WooCommerce core plus a well-maintained SEO plugin covers eighty percent of the ground; the last twenty percent is the properties your category needs and the generic plugin does not emit. That piece is custom work.
A product feed worth reading. Google Merchant Center feed as the baseline, maintained to the current spec, every attribute populated, no “misc” categories. If you plan to participate in the Universal Commerce Protocol surface, this feed is the raw material. A clean feed also feeds the ad surface, so there is a day-one return even before any agent reads it.
A policy layer the agent can honour. Returns policy, shipping policy, availability by region, and MAP constraints have to live somewhere the agent can read. On a store we build, these live as structured facts on the product and in site-wide schema, not as PDFs the agent cannot parse. If an agent sells your product to a buyer in a country you do not ship to, the complaint lands on your customer service team, not the agent’s.
An agent switch, per surface, under your control. This is the one Finkelstein named. Whatever platform choice you make, insist that the admin can enable or disable each agent surface individually, with a log of what each agent saw and bought. If the only control is “on” or “off” across the board, you have given up the ability to negotiate terms per agent, and the first time ChatGPT changes its commission structure you will wish you had not.
For a mid-market WooCommerce catalogue of two to ten thousand SKUs, we would scope this at four to six weeks of focused work, roughly half on product data cleanup and half on the schema, feed and policy layer. The agent switch is a smaller piece of configuration, assuming the platform exposes it. If it does not, that is the first conversation to have with the vendor.
The question that is still open
None of the six voices above has an answer to the hardest question, which is attribution. When an AI agent delivers a basket, who gets credit, and against what budget? The agent took the fan-out query, filtered the shortlist, matched the policy and ran the checkout. The traditional attribution stack sees “direct” traffic and shrugs. The ad budget that fed the shortlist does not light up. The content budget that trained the agent on your brand does not light up either. Shopify’s 14x figure is a count, not an attribution.
Our view, at the risk of taking a position before the industry has: open the door, cleanly and conditionally. Participate in the protocols that are published and verifiable. Build the data model and the policy layer that makes the catalogue legible to any agent, because that work pays even if agents arrive slower than Shopify’s figure suggests. Keep the per-agent switch under your own control. Treat the attribution problem as the next build, not a blocker on the first. If you are an ecommerce director weighing this for 2027 and want a second opinion from people who ship the data model on real WooCommerce estates, that is a conversation we welcome.
Last modified: October 22, 2026
