Who Is Actually Shopping?

Inside the Court Battle That Will Decide the Agent Economy

In March 2026, a federal judge in San Francisco did something no court had ever done before: she ruled on whether an AI agent, acting with a customer's explicit permission and using that customer's own password, is allowed to shop on a website that doesn't want it there. The website was Amazon. The agent belonged to Perplexity, the AI search company whose Comet browser lets people say "buy me this" and walk away. Amazon won. The ruling sounds like a narrow, technical dispute about browser fingerprinting and terms of service. It is actually the opening skirmish in a fight that will determine who really controls commerce for the next decade - and the honest answer, for the first time in retail history, might not be a company or a person, but a piece of software neither the buyer nor the seller fully owns.

From "Help Me Find It" to "Just Buy It"

The idea of an AI agent doing your shopping stopped being science fiction in January 2025, when OpenAI demoed its Operator agent live: someone held up a handwritten grocery list to a camera, and within seconds the AI had read it, opened Instacart, built a cart, and scheduled delivery. It was a compelling glimpse of a world where the friction of online shopping - tabs, comparisons, abandoned carts - simply dissolves into a conversation. By September 2025, OpenAI turned that glimpse into a product: Instant Checkout, built on something called the Agentic Commerce Protocol, developed jointly with Stripe. For the first time, a ChatGPT user could buy something from an Etsy seller without ever leaving the chat window. Etsy's stock jumped 16 percent the day it was announced. Shopify's rose 6 percent. Within weeks, Walmart, Target, and Instacart had signed on too.

The mechanics matter, because they reveal what's actually changing. In traditional e-commerce, a retailer controls the storefront, the checkout, and the customer relationship, while the shopper does the browsing and clicking. In agentic commerce, an AI agent carries your identity, your stored payment method, and your specific intent directly into the transaction - it isn't just helping you decide, it's standing in for you at the register. OpenAI insists it doesn't take a cut of what you pay and doesn't rank sponsored products higher, collecting instead a small fee from the merchant on each completed sale. That fee, tiny per transaction, is also the first genuinely new revenue model e-commerce has seen since the invention of the shopping cart button.

Four Companies Building Four Different Toll Roads

Whoever controls the technical rails that agents and merchants use to complete a purchase stands to collect a small piece of nearly every transaction that flows through them - which is exactly why, within a single week in the spring of 2025, four of the most powerful companies in payments and technology raced to build competing versions of the same idea. Mastercard launched Agent Pay, which issues a tokenized "Agentic Token" scoped to one specific AI agent, one merchant, and one consent policy, so an assistant like ChatGPT or Microsoft Copilot can complete a purchase without ever touching your actual card number. Visa answered one day later with what became its Trusted Agent Protocol, built with security partners including Cloudflare and Akamai, designed to let a merchant tell the difference between a legitimate shopping agent and a scraping bot trying to slip past its defenses. Google built something more foundational still: the Agent Payments Protocol, an open standard both card networks eventually agreed to support alongside their own systems. And PayPal, moving faster than either card network expected, partnered with Perplexity to launch Instant Buy in November 2025, wiring its payment tokens directly into more than six thousand merchants and into Mastercard's own agentic framework.

Visa's own language about this moment is unusually blunt for a payments company: it has described the most recent holiday shopping season as the last one in which consumers will check out entirely alone. Whether or not that's premature, the direction is unmistakable. Mastercard has already gone a step further, launching a second system called Agent Pay for Machines specifically for micro-transactions between software systems that happen automatically, continuously, with no human present at all - the electricity meter and the subscription renewal, not the sweater you're deciding whether to buy.

Amazon's Impossible Position

No company illustrates the stakes of this shift more starkly than Amazon, which spent 2025 trying to do two contradictory things simultaneously: build its own agent while suing someone else's.

Amazon's own AI shopping assistant, Rufus, had been used by roughly 250 million customers by the end of 2025, and Amazon's internal data shows that shoppers who engage with Rufus are 60 percent more likely to complete a purchase than those who don't - a genuinely enormous number for any retail feature. Amazon also began testing "Buy For Me," letting Rufus purchase items from other companies' websites directly inside the Amazon app, and quietly updated its terms in March 2026 to formally require that any AI agent accessing its store identify itself honestly.

That last requirement is precisely what led to the lawsuit. In November 2025, Amazon sued Perplexity, alleging that Comet had been disguising its automated browsing sessions as ordinary human traffic from a Chrome browser rather than declaring itself as an agent - behavior Amazon says it warned Perplexity about repeatedly since November 2024, only to watch the startup push a software fix within 24 hours of Amazon's first technical block. Perplexity called the lawsuit bullying, arguing that an agent acting under a user's own login credentials should inherit that user's own permissions. In March 2026 the judge sided with Amazon, finding strong evidence that Comet had accessed password-protected accounts in a way Amazon never authorized - an early legal precedent suggesting that platforms can refuse entry to an agent even when the shopper behind it explicitly said yes.

Buried in Amazon's own court filings is the real explanation for why this fight matters so much, and it has nothing to do with security theater. Amazon told the court that agent traffic creates genuine problems for its advertising business, because automated visits have to be identified and filtered out before advertisers can be billed for the human eyeballs they're actually paying to reach. That advertising business generated roughly $68.6 billion for Amazon in the trailing year - and an agent that quietly buys a product without ever rendering a single sponsored listing on a human screen is, from that specific business's point of view, a customer who has stopped paying rent. Meanwhile, in a detail that captures just how tangled these corporate relationships have become, Amazon's own cloud division signed a $38 billion infrastructure deal with OpenAI one day before Amazon filed suit against a completely different AI company for doing something structurally similar to what OpenAI itself was building.

Even Amazon's public position keeps shifting under the pressure. CEO Andy Jassy has said openly that agentic commerce could be genuinely good for the industry, while also arguing that today's agents simply aren't good enough yet - no shopping history, no personalization, delivery estimates that are frequently wrong. Quietly, Amazon has let its own subsidiaries, including Zappos and Shopbop, remain open to outside AI agents, testing the model on smaller, lower-stakes storefronts before deciding whether to ever open the main event.

The Gap Between the Headlines and What Actually Works

The rush of announcements makes agentic commerce sound like a solved problem. The on-the-ground reality, as of early 2026, is considerably messier. Roughly a year after Instant Checkout launched with fanfare, only about a dozen of the more than one million Shopify merchants who were promised access had actually completed the technical integration required to use it. One industry analyst testing Perplexity's own Instant Buy tool tried repeatedly to purchase a simple sweater from a major retailer and gave up after the agent kept returning error messages, despite the item being clearly in stock. Weeks later, the same analyst asked ChatGPT to recommend a coffee machine, got a sensible suggestion, clicked through - and found himself looking at a photograph of a garden rake. OpenAI has since begun quietly redesigning its approach, moving away from checkout inside the chat window itself and toward dedicated retailer apps that route the final purchase back to the merchant's own site, trading some of the magic of frictionless buying for a transaction system that actually, reliably works. There was also, until quite recently, a genuine compliance gap nobody had fully solved: as of February 2026, OpenAI had not yet built a system for collecting and remitting U.S. state sales tax on transactions completed through its own checkout - a mundane but real regulatory problem hiding underneath the futuristic press releases.

None of this means the trend is overhyped so much as early. McKinsey projects agentic commerce could generate as much as one trillion dollars in U.S. retail revenue by 2030. Morgan Stanley expects nearly half of American shoppers to be using AI agents by then, adding as much as 115 billion dollars in incremental e-commerce spending. Adobe Analytics measured a 693 percent year-over-year increase in traffic reaching U.S. retail sites directly from generative AI tools during the 2025 holiday shopping season alone. Those numbers describe genuine consumer appetite colliding, for now, with genuinely unfinished infrastructure.

What This Means If You're the One Doing the Shopping

For an ordinary consumer, the appeal is obvious: describe what you want once, and let something else compare, negotiate, and complete the purchase. But the trade-offs are subtler than they first appear. An AI agent is, by design, considerably less impulsive than you are - one analyst pointed out that an agent sent to buy a laundry basket won't wander off and impulse-add a novel and a pair of earphones the way a distracted human browsing Amazon reliably does, which is either a genuine convenience or the quiet end of an entire category of marketing built around exactly that kind of distraction, depending on which side of the transaction you're standing on.

Security architecture has actually improved faster than most people realize: both Visa's and Mastercard's frameworks tokenize your payment credential specifically for one agent and one merchant, meaning your actual card number never passes through the AI system at all, which is a meaningfully stronger privacy position than typing your card into a browser yourself. But a genuinely new question has opened up that didn't exist before: whose interests does your shopping agent actually represent? The company that built your assistant collects a fee from the merchant on completed sales, which creates at least a structural incentive worth watching closely, even where a platform states clearly that its results aren't sponsored. And the reliability gap is real - an agent that occasionally sends you a garden rake instead of a coffee machine is a bigger problem when it's also holding your payment credentials and executing the purchase without a final human glance at the actual product.

What This Means If You're a Small Seller, a Freelancer, or a One-Person Shop

This is where the agent economy cuts in two genuinely opposite directions at once, and both are already visible.

The optimistic version: agentic commerce, at least in its stated design, promises a form of discovery that doesn't run through paid advertising or algorithmic feeds built to reward whoever spends the most. OpenAI has said explicitly that Instant Checkout results are ranked on relevance, not sponsorship - meaning a small Etsy maker with a genuinely well-reviewed, well-described product has, in theory, the same shot at being surfaced as a company with a seven-figure ad budget. For independent sellers who have spent years fighting algorithmic feeds engineered to favor whoever can outbid them, an assistant that judges products on merit rather than marketing spend is a real, structural opportunity - arguably the first one of its kind in the history of online retail.

The less comfortable version: getting discovered by an agent requires being legible to one, and that legibility is itself a new kind of technical labor that not every small business has the time, money, or expertise to perform. The fact that only a small fraction of eligible Shopify merchants actually completed integration with Instant Checkout in its first year isn't really a story about disinterest - it's a story about how much backend work "just showing up" for an AI agent quietly requires: structured product data, clean inventory feeds, accurate real-time pricing, a checkout system built to accept a machine-initiated transaction. Traditional search-engine optimization is being joined, and in some categories replaced, by an entirely new discipline people are already calling generative engine optimization - the practice of structuring your product information so that an AI model reliably understands, trusts, and recommends it. For a solo seller already juggling production, packaging, and customer service, this is one more full-time skill to either learn or pay someone else to handle, and the sellers who move early are likely to enjoy a meaningful first-mover advantage over the ones who wait.

There's also a fee question hiding in the excitement. Every agentic checkout system charges the merchant something on each completed sale - small individually, but stacked on top of existing marketplace fees, payment processing costs, and shipping, on margins that are frequently thin to begin with for independent creators. Etsy notably agreed to absorb OpenAI's commission on behalf of its sellers early on, precisely to avoid scaring off the small businesses it depends on - a sign that the platforms themselves recognize how easily this new toll could squeeze exactly the entrepreneurs it claims to be leveling the playing field for.

What This Means If You Run, or Compete With, a Large Platform

For the Amazons, Walmarts, and big-box retailers of the world, the threat isn't abstract - it's the specific line item Amazon named in its own federal court complaint. Advertising has become one of the most profitable businesses in modern retail precisely because it depends on a human being physically present to see a sponsored listing and be persuaded by it. An agent that silently compares options, picks a winner based on price and specification, and completes the purchase without a human ever scrolling past a banner is a customer who has quietly opted out of the entire monetization layer built around their attention. Gartner analysts have already begun speculating about what they call "machine advertising" - a discipline built to persuade an algorithm rather than a person, whose feasibility, legality, and basic shape remain genuinely unresolved as of this writing.

This is precisely why the biggest players are pursuing what one analyst has called an embrace-and-defend strategy simultaneously: sign the partnership, get the press release, and quietly build every technical and legal guardrail available to keep the actual transaction, and the actual customer data, inside the corporate walls that generate the advertising and loyalty-program revenue in the first place. Walmart and Shopify both partnered with OpenAI on Instant Checkout while separately restricting what agents are permitted to see and do on their own platforms. Amazon is doing something similar but more dramatic: fighting a smaller AI company in federal court over unauthorized access while simultaneously deepening a multi-billion-dollar infrastructure partnership with the far larger AI company building the very technology it's suing over.

The deeper structural risk for large platforms is one that has already played out in other industries once a powerful intermediary inserted itself between a business and its customer: airlines and hotels didn't disappear when booking aggregators arrived, but they lost meaningful pricing power and a great deal of the direct customer relationship that had once been theirs. If AI agents become the primary interface through which people shop, the retailers behind them risk sliding into something closer to a fulfillment and logistics utility - excellent at getting a package to a door quickly, but no longer the entity that owns the moment of decision, the loyalty program, or the brand relationship that used to justify premium pricing in the first place.

Where This Moves Fastest, and Where It Doesn't

Not every category of shopping is equally exposed, and the dividing line is fairly intuitive once you see it. Routine, repeat, low-emotion purchases - the same brand of paper towels, a grocery reorder, a printer cartridge - are exactly the kind of decision an agent is built to handle well, because there's little genuine preference being expressed beyond price, availability, and speed of delivery. This is also where the biggest retailers' traditional advantages - scale, logistics, low prices - remain genuinely decisive, agent or no agent.

Purchases with real emotional weight - a wedding dress, a piece of art, a gift for someone specific, a luxury item bought partly for the experience of buying it - are far more resistant to full agentic handoff, and likely to stay that way for a long while. But even there, the first pass is already starting to move to software: an agent might do the initial screening of options across a dozen sites before a human ever steps in to make the final, emotionally-loaded decision themselves, which means simply getting shortlisted by an algorithm is quietly becoming as important as getting noticed by a human ever was.

The Fight That's Really Happening

Which brings the story back to a courtroom in San Francisco, and a company arguing that permission from a customer isn't the same thing as permission from the store. That distinction is going to keep resurfacing, in different forms, for years: not a fight between humans and machines, but a fight over who gets to stand between them, and who collects a fee every time they do. Consumers get real convenience and, so far, meaningfully better payment security than typing a card number into a browser - with a new and genuinely open question about whose interests their assistant actually serves. Independent sellers get a rare shot at being judged on the merits of their product rather than the size of their ad budget - if they can afford to become legible to the algorithm doing the judging. And the largest retailers on earth are discovering, in real time and in federal court, that the thing they built their businesses on - being the place people go to buy things - might no longer be theirs to control.