Some retailers are already seeing 15 to 20 percent of their referral traffic arrive not from Google, not from social media, but from AI chat interfaces. That number is moving fast. By 2030, Morgan Stanley estimates that AI agents handling purchases on behalf of consumers could represent $190 billion to $385 billion in U.S. e-commerce spending alone — with groceries and consumer packaged goods leading the way. In September 2025, ChatGPT users started buying products without ever visiting a retailer’s website. Google, Amazon, and Microsoft followed within months. This isn’t a new checkout button. It’s a restructuring of how commerce works — and understanding what’s actually changing requires mapping a system that looks almost nothing like the one it’s replacing.
The Parties in the System
For most of the internet era, a purchase involved two parties and a network between them. A person wanted something, a merchant sold it, and a payment rail connected the two. Simple enough that we stopped thinking about it.
Agentic commerce breaks that open.
The consumer is still there, but their role has changed. They set the intent — a budget, a preference, a constraint — and then step back. The actual work of finding, comparing, and buying is handed off. That handoff goes to a buy-side agent: software that acts on the consumer’s behalf, moving across merchant systems, reading product data, evaluating options, and executing the transaction. The consumer may never see a product page.
On the other side of that transaction, the merchant has deployed their own agent. A sell-side agent manages inventory signals, pricing, and how the brand presents itself to incoming buy-side agents. As J.P. Morgan has noted, businesses will need sell-side AI that can effectively market their products directly to other AI — not to humans browsing a homepage. The pitch deck and the hero image are irrelevant. The structured data feed is everything.
Sitting beneath all of this is the payment network — verifying that the agent is who it claims to be, that the authorization is genuine, that the transaction clears and settles correctly. A system designed to authenticate humans now has to authenticate software acting on their behalf.
And the merchant remains the merchant of record throughout — legally responsible for what’s sold and how it’s fulfilled — while potentially never interacting with the buyer at all. The human relationship at the center of commerce for centuries has been quietly removed from the loop.
What Has to Change
Four things need to shift across the commerce ecosystem before agentic transactions can scale. They are not equally urgent.
The most immediate is product data. When an AI agent is the buyer, it doesn’t browse — it reads. Brands whose product information isn’t machine-readable, cleanly structured, and semantically rich simply won’t appear in an agent’s consideration set. McKinsey describes this as a shift that transcends traditional SEO: companies need to align with the data structures and decision logic of AI agents, not the keyword patterns of search crawlers. The discipline already has a name — answer engine optimization — and for brands that have invested heavily in visual merchandising and paid search, the pivot is significant. The storefront has moved, and most brands haven’t updated their address.
The second shift is in advertising, and it’s the one the industry is least prepared for. For fully autonomous agents doing product discovery, the traditional ad unit is invisible — agents don’t see banner ads, and they don’t scroll past sponsored placements. Retail media networks — a multi-billion dollar industry built on onsite ad formats and promoted listings — face a structural challenge if discovery moves entirely to AI-driven interfaces. But a more immediate disruption may be less dramatic: Stripe’s just-announced ACP-powered checkout inside Facebook ads points to a hybrid model where human discovery and agentic execution coexist. The consumer still sees the ad; the AI handles the transaction. In that model, the valuable real estate shifts — not from ads to nothing, but from click-through rates to checkout infrastructure connectivity. Whether a brand’s backend is wired into the platforms where discovery happens matters more than how much they spend on placement. The entire economics of brand visibility online is being renegotiated, and the terms aren’t fully written yet.
Third, consumer trust has to be earned category by category, and it won’t happen uniformly. Research suggests roughly half of consumers remain cautious about fully autonomous purchasing even as a growing share already use AI for product research and comparison. The adoption curve will follow risk tolerance: subscription renewals and grocery replenishment before electronics, travel, or anything requiring judgment. This is not a technology problem. It’s a human one, and it will resolve on its own timeline regardless of how good the protocols get.
Regulation is the fourth gap, and it’s widening. The Center for Data Innovation has documented how frameworks like Regulation E provide no clear mechanism for handling disputes in agentic purchases, and how Sarbanes-Oxley compliance requirements were never designed for AI-generated audit trails. Neither the executive branch nor Congress has proposed concrete changes. The EU AI Act begins enforcement in August 2026 — the first major framework that will touch agentic systems at scale — but contains no specific provisions for autonomous purchasing. The companies building this infrastructure are, for now, writing the rules themselves.
The Execution Question
Beneath the protocols and the policy debates, there is a more fundamental problem that doesn’t get much attention. These agents need to operate reliably in environments that were not designed for them — interacting with merchant systems, handling payment credentials, managing the full sequence of a transaction without human oversight. When something goes wrong mid-transaction, the system needs to know how to recover cleanly, keep accurate records, and avoid outcomes like duplicate charges or orders stuck in ambiguous states. The infrastructure that makes safe, autonomous transactions possible at scale is still being built. The protocol layer tells agents what they’re allowed to do. The execution layer determines whether they can actually do it without breaking things.
Where This Is Heading
The companies that will have an advantage in this transition are not necessarily the largest ones. They’re the ones that treat their product data as infrastructure, rethink how they’re discoverable to non-human buyers, and understand that the new parties in this system — the buy-side agent, the sell-side agent, the execution layer between them — are now as important as the consumer relationship they’ve spent years optimizing. The ones waiting for regulatory clarity or a dominant protocol to emerge may find that the architecture was set while they were watching.
