AI Shopping Agents in 2026: What They Can Actually Buy

Every time I sit with a client who has read the agentic commerce headlines, the first question is some version of the same one. Should we rewrite our checkout for the AI agent, or is this the next bit of noise we quietly outlast. The honest answer is that nobody running a real store knows yet, and the people most publicly selling the idea have already pulled a feature back once.

I am writing for the person who signs the budget for 2027, not the person who writes the integration. The question is a procurement one. If AI shopping agents are going to shift how a brand gets found, cited and transacted, you need to know which parts of that are already running in production, which are a wager, and which are a demo that stopped working last March. The named practitioners making the case publicly disagree with each other more than any one piece of coverage lets on.

So I went and read what they actually said. Six voices across platform, payments, analyst and operator, on the record, with publication dates. The picture that comes out is neither the demo-stage utopia the keynote talks promised nor the bust the loudest skeptics predict. It is a market where the infrastructure is being poured faster than the behaviour it is meant to support, and the question for a buyer is where to spend first.

I asked whether AI shopping agents are actually buying anything a mid-market store would notice, and what the people building the plumbing would admit to on the record.

Who answered on the record

• Emily Pfeiffer, Principal Analyst, Forrester. Author of the firm’s Predictions 2026 for digital commerce, 28 October 2025.
• Anushree Verma, Senior Director Analyst, Gartner. Lead voice on the firm’s agentic AI project outlook, 25 June 2025, and the April 2026 hype-cycle work.
• Steve Dennis, President and Founder, SageBerry Consulting. Interviewed in Dallas Innovates, 5 February 2026.
• Will Gaybrick, President of Technology and Business, Stripe. On the Agentic Commerce Protocol launch, 29 September 2025.
• Vanessa Lee, VP of Product, Shopify. On the Universal Commerce Protocol launch, 11 January 2026.
• OpenAI, in a company statement reported by Storyboard18, 25 March 2026, on retiring Instant Checkout.

The platforms are building the pipes before the buyers arrive

The platform position is that this is already infrastructure, and the race is for the standard that other people adopt. Vanessa Lee of Shopify framed it in the January announcement: “Shopify has a history of building checkouts for millions of unique retail businesses,” and the Universal Commerce Protocol extends that history outward, so an AI agent can hit a merchant’s cart with discount codes, loyalty credentials, subscription preferences and payment all in one call. Twenty-plus retailers signed on at launch, including Monos, Gymshark, Everlane, Keen and Pura Vida.

Will Gaybrick of Stripe made the same bet at the payments layer four months earlier, with a quieter line. “Stripe is building the economic infrastructure for AI,” he said, launching the Agentic Commerce Protocol with OpenAI as an open standard. The claim is that if agents are going to transact, somebody has to run the rails, and Stripe would rather it be them.

Read these two announcements next to each other and the shape of the market becomes visible. Shopify’s standard is co-developed with Google. Stripe’s is co-developed with OpenAI. Neither sits inside the other. For a merchant that uses both a Shopify-shaped store and Stripe payments, the integration work is doubled, which is the usual outcome of a standards race before the market has picked a winner.

The pullback that nobody announced in the launch deck

Then, in March 2026, OpenAI reversed. The company retired the in-chat Instant Checkout it had launched the previous September and refocused ChatGPT shopping on discovery, comparison and reviews. The company statement, in its own blog post, was that the initial version “did not provide the level of flexibility” it had wanted, and reporting from The Information and Storyboard18 added that user uptake of the checkout feature had been thin.

That is a datum. One of the two most visible launches, from the company most people think of first when they hear “AI agent”, decided within six months that shoppers were not completing purchases inside the chat window at a rate that justified the product. Discovery is now the surface, and the merchant’s own site is the place the sale closes. The Agentic Commerce Protocol with Stripe survived; the “buy inside the chat” product did not, at least for now.

I would not read that as a verdict on the whole category. I would read it as a verdict on asking a consumer to type their shipping address into a chatbot before they have made a decision they feel sure about. There is a long literature on how little new behaviour online shoppers tolerate at the point of purchase, and this looks like another instance of it.

The analysts: a trust gap larger than the demos

Emily Pfeiffer of Forrester wrote the clearest-eyed piece of 2025 on this. Her 2026 predictions post, published 28 October 2025, forecast three things. Five major US or European brands will build unified agentic experiences that fold customer service, logistics, payments and recommendations into a single interaction. One-third of retail marketplace projects will be abandoned as answer engines steal traffic from them. And twenty percent of B2B sellers will see bot-to-bot quote negotiation. All three are hedged by one number from Forrester’s own Consumer Benchmark Survey: 24% of US online adults have used ChatGPT at all. A behaviour that most consumers have not even tried once is not yet a channel.

Anushree Verma of Gartner put it harder in the firm’s agentic AI project outlook. Her position: “Most agentic AI propositions lack significant value or return on investment (ROI), as current models don’t have the maturity and agency to autonomously achieve complex business goals or follow nuanced instructions over time. Many use cases positioned as agentic today don’t require agentic implementations.” Gartner’s forecast, attached to that reading, is that more than 40% of agentic AI projects will be cancelled by the end of 2027. The 2026 hype cycle for agentic AI, published in April, sits the category at the Peak of Inflated Expectations.

Two different firms, two different registers, same underlying claim. The infrastructure is real. The behaviour it is meant to serve is not yet widespread, and some of what is being sold as agentic is a thin wrapper on a workflow that did not need an agent in the first place.

“Most agentic AI propositions lack significant value or return on investment… many use cases positioned as agentic today don’t require agentic implementations.”

Anushree Verma, Senior Director Analyst, Gartner

The pragmatic middle: real, with the usual casualties

Steve Dennis of SageBerry Consulting, writing in Dallas Innovates in February 2026, took the position I find most useful for a buyer. He is explicit that failed experiments are inevitable and that the industry “may well find ourselves teetering on the edge of the trough of disillusionment soon enough”. He then names why he does not treat this as another metaverse cycle: “Unlike the metaverse, this feels orders of magnitude more real.” The pace of adoption, in his reading, is “breathtaking”, with significant capital behind it.

That is the register to copy. Say plainly which bets will fail, and still make the bets that look like they are tied to a real shift in behaviour. The honest buyer position on this topic in October 2026 is some version of: part of it will stick, part of it will embarrass somebody, and the way to tell which is which is to watch what people actually do, not what they are shown doing in a demo.

What a WooCommerce store should actually build this quarter

Here is where I earn the slot. If you run a mid-market store on WordPress and WooCommerce, the question is not which of these protocols to adopt. It is which of the preparations you make for them pay off even if the protocols never win.

The data model first. The thing an agent and a human shopper both want is a product record that answers questions in structured fields rather than paragraphs of marketing copy. In WooCommerce that means attributes modelled as real taxonomies, variations with explicit option values, dimensions and weight populated, and the price, stock, sku and shipping class fields clean and current. The Store API, which ships with WooCommerce Blocks, already exposes most of that cleanly. If an agent cannot read your catalogue through Store API today, it is not going to read it through UCP or ACP tomorrow.

The structured data layer is the next one. Product schema with offers, availability, price, priceValidUntil, sku, gtin or mpn, aggregated review and brand. Yoast and Rank Math both emit most of this; the plugin is not the hard part. The hard part is that most catalogues emit it incompletely, with missing GTINs, stale availability, and review markup that duplicates fields the merchant did not realise were being populated by two different sources. A merge-and-prune pass against a real crawl is a one-week project on a 1,500-SKU store, and it is the single highest-leverage piece of agent readiness I know.

The feed layer sits alongside that. Google Merchant Center, Microsoft Shopping, Meta’s catalogue and Pinterest all accept the same core feed shape, and when Shopify’s UCP shipped the retailers on the launch list were brands with clean, up-to-date Merchant Center feeds already. The feed is also where agents will read inventory first, because it refreshes faster than a crawl of the storefront. If your feed is the twice-weekly XML your SEO plugin writes to disk, that is where to start, not with another protocol to adopt.

The checkout, finally. OpenAI’s pullback is the signal that for most consumer purchases the chat window is not where the sale closes, and that is good news for a WooCommerce merchant. It means the agent will likely hand the shopper to your checkout for the foreseeable future, and the work there is the work that was always there: fewer form fields, a return policy a human can read without three clicks, abandoned-cart recovery that does not look like a stalking campaign. Shipping and returns are also where agents will concentrate their questions on your behalf, so the policy pages that are often the thinnest part of a site are now also the part an answer engine is most likely to cite.

Rough shape of effort on a typical 1,000-SKU store with WooCommerce and the standard SEO and feed plugins already running: a focused month of work, one senior engineer plus one content lead, lands all four layers at a level a reasonable agent integration will read. Carry on past that and you are optimising for a protocol that has not been picked yet. Stop short of it and the first agentic feature you turn on will trip on fields that should have been filled in two years ago.

What nobody on the record has solved

The unresolved part is attribution. When a shopper reaches your checkout after a conversation they had with an AI agent that pulled your product from a feed and recommended it against three competitors, which part of your marketing budget caused the sale. None of the six voices above has answered that cleanly, and the measurement tooling that most teams use today runs on JavaScript in the browser, which an agent is not executing. Jodi Cerretani of WordPress VIP made that point explicitly in a webinar last month, and she is right: the tools measure the human, and the agent is not one.

So the honest 2027 position for a mid-market board is that the catalogue, the schema, the feed and the checkout are a bet that pays off either way, and the agent protocols are a bet that pays off only if the protocol you picked is the one that wins. We would do the first set of work for any client in Q4, and we would hold the protocol commitment for one more quarter, because the lesson of September 2025 to March 2026 is that the loudest launch is not always the one that survives. If you want to talk through what this looks like on your estate, that is the conversation to pick up.