Agentic Commerce Has Arrived, But Not Without Complications

At this summer's Commerce Next in New York, retail's biggest conference, agentic AI wasn't a buzzword anymore; it was the conversation. Not "should we think about this," but "how do we get ready?" A new report from ICSC and McKinsey & Company projects that agentic commerce in the U.S. business-to-consumer retail market could reach $1 trillion in revenue by 2030. Stripe co-founder and President John Collison was even more blunt: consumers will soon hand off their most boring shopping tasks to AI, the same way they once handed off trip planning to Expedia or product search to Google. But beneath the excitement, a more complicated picture is emerging. The Hype Has a Foundation The momentum behind agentic commerce isn't appearing from nowhere. Nearly 68% of consumers used at least one AI tool in the past three months as part of their shopping journey, and 62% have used AI to compare brands, models, prices, or reviews, according to the ICSC/McKinsey findings. The behavior shift is already underway; the question is how fast it will scale, and which brands will benefit. Collison frames agentic commerce as the natural endpoint of two decades of friction-reduction in online shopping. "When you find the product at the very end, do you really then want to go and be filling out all these web form fields?" he asked in a recent Bloomberg interview. His answer: consumers will increasingly say "buy it for me", especially for low-stakes, repetitive purchases like household staples, recipe ingredients, or travel accessories. He also makes a distinction that gets to the heart of the strategic challenge: not all shopping is the same. People want AI to handle the mundane. They don't want it to replace scrolling through fashion, planning a vacation, or discovering something unexpected. That line between "automate this" and "let me do this myself" will define which categories thrive in an agentic world, and which ones resist it. David Ruidor, CEO of GenLayer, frames the underlying shift in similar terms, but pushes it further. "We are humans and we have 24 hours a day," he told International Business Times. "All the agentic vision goes around the concept of having thousands of agents doing actions on our behalf while we sleep. Humans will be able to 10x themselves thanks to this approach. And it won't only be a matter of hours but also quality." Retail Is Already Ahead of the Curve, But Not Evenly For companies actually building AI systems, retail is turning out to be fertile ground. AI engineering company Solvd found that retail AI projects are significantly more likely to report high positive ROI than other industries, 26.4% versus 18.6% in the broader market. That gap suggests retail has real structural advantages, such as high transaction volume, rich customer data, and clear, measurable outcomes. But those advantages don't distribute evenly. To understand why, Solvd CTO Skylar Roebuck offers a diagnostic he calls a five-axis framework, a way to test whether a given category is suited for agentic commerce, and if so, which type. The first axis is what Roebuck calls "specifiability of intent": can a customer's desire be expressed as machine-evaluable constraints? "'Coffee filters, size 4, the brand I always buy' is fully specifiable," he told International Business Times. "'A sofa that feels right in my living room' is not." The more a buying decision is articulated in terms of criteria rather than taste residing in the customer's head, the more an agent can transact autonomously today. The second axis is repeat cadence and loyalty: agentic value compounds through a learning loop, and categories with high purchase frequency let the agent learn preferences over time and earn delegation. Ruidor describes almost exactly how that trust gets built in practice: "This will happen little by little, first delegating decisions worth very little, in money or impact, and when the experience proves to be good, users will keep delegating more and more." Delegation, in other words, isn't a switch consumers flip once; it's a habit an agent has to earn one small purchase at a time. Third is the consequence of a wrong choice: low-stakes, reversible purchases get delegated early; high-stakes or identity-centric ones stay human-managed. Two final axes shape the competitive dimension: whether a category differentiates beyond price, and whether it has a high-intent replenishment loop. "If your category competes mainly on price, an agent collapses the decision to a price comparison, and you may win the transaction and destroy your margin," Roebuck warns. Categories where brand, curation, service, or fit genuinely differentiate are better positioned to survive agent mediation. High-intent replenishment categories, the ones that look like Amazon's automatic purchasing, are ready for autonomous transactions now. Taste-driven discovery categories like fashion, home décor, and beauty aren't unfit; they're a different kind of fit. "They're a fit for agent-assisted curation and powerful multimodal experiences," Roebuck says, "which is exactly where your intelligence layer is the moat rather than a checkbox." The Readiness Gap Nobody's Talking About In terms of where business leaders should start if they want to be agentic-commerce-ready in 18 months, Roebuck's answer is counterintuitive. "I would tell them explicitly not to start with the agent," he says. "The most common 2026 failure is bolting a chat or checkout surface onto a business that can't transact, fulfill, or learn legibly underneath it." His roadmap begins with what he calls making the substrate "LLM-legible". Catalog, pricing, inventory, and fulfillment exposed as structured, real-time, trustworthy data. "If an agent can't get a reliable 'can this arrive Thursday' answer, you simply don't appear in the consideration set," he explains. The second step is instrumenting an "intent flywheel" to capture the signals needed for the system to learn. The third, and most organizationally challenging, is building a path to actual decision authority. "An agent that routes everything back to a human for approval isn't agentic; it's a recommendation engine with extra latency." The most underestimated prerequisite, in his view, is the governance work required to make autonomy safe enough to actually ship, the guardrails and decision-rights architecture that allow a system to act without a human in the loop for every interaction. "Most retailers can produce a clean catalog for a demo," he says. "Very few can guarantee inventory and fulfillment accuracy under production load, and have done the governance and risk-tolerance work to let a system act without a human in the loop for every interaction." Ruidor points to what that governance work actually looks like in practice: "To give trust to the users, agents will have harnesses such as max expenditure limits, man-in-the-middle validations, allowance in certain websites or concepts, etc." This isn't a theoretical example. Recently, Nisum's Global AI Leader Guillermo Delgado and their SVP of Growth and Corporate Development, Martin Lewitt, hosted a webinar on agentic commerce to show exactly what this looks like in practice. They ran a live demo in which a chain of autonomous agents shopped for a pair of running shoes end to end: one searched retailers, another weighed reviews and ran a sentiment analysis to decide, a third handled payment, and a fourth coordinated delivery around the buyer's calendar. The exercise revealed that when retailers blocked data access or failed to return product specs, the agents simply dropped them from consideration, regardless of how well known the retailer is. "For the agents, this is really a reason to simply skip you and go to the next retailer," Delgado said. "It automatically bypasses you because you're not the only retailer out there." For Lewitt, that reordering rewrites the competitive question itself. "It's no longer how to attract clients, but how to ensure that agents find us and select us," he added. The consumer trust question is just as relevant. Research cited by PYMNTS found that 95% of consumers have at least one concern about agentic commerce, ranging from buying the wrong item to more serious fears around identity theft and data misuse. At scale, a single high-profile failure — an agent that over-ordered, bought the wrong variant, or exposed payment data — could set back adoption in an entire category. The "human-in-the-loop" question is therefore not just a UX decision. It's a trust-building strategy, especially in the early adoption window. Where the Real Opportunity Sits Given all of this, where should retail leaders actually focus? Roebuck's answer may surprise those expecting a list of obvious quick wins. "The opportunity is narrower and harder than 'automate the simple stuff,'" he says. "It's how you preserve brand and relationship trust with the consumer while delivering an experience that's both genuinely frictionless and powerful enough that they want to delegate to you." He outlines a few concrete moves for brands serious about getting ahead. The first is aggressive product data enrichment, not just for protocol compliance, but rich enough that an agent can actually reason about it. "Product data is a product now in the world of LLMs," he says, "and this will require ongoing maintenance, not a one-time cleanup." The second is a willingness to pioneer use cases rather than wait for a settled playbook. "The playbook is being written right now, and the brands writing it get the advantage. Don't wait or blame messy data as a reason not to act." The third, and perhaps most strategic, is a deliberate decision about AI's role in the organization. "Are you bolting on point solutions, or investing in an intelligence layer that lets you maximize personalization across the entire organization?" Roebuck asks. "Those are two very different companies." He advocates for running a portfolio of concurrent AI projects rather than a single pilot, spanning quick wins, staged bets, and longer-term game changers. An underlying enabler, he notes, is agentic software development itself, which accelerates the pace of building and changes the build-versus-buy calculus in ways most retailers haven't yet accounted for. Ruidor offers a useful way to think about where that line between "automate" and "keep human" will actually sit for individual consumers, not just categories. "If you only need to buy a house once, and it's a very important decision for your life, you won't delegate it to an AI," he says. "But if it's something that's very repetitive, low value, or highly mathematical — a simple comparison of prices or similar — you will. At the beginning it will be only a small subset of purchases, but little by little this subset will start increasing." It's the same boundary Roebuck describes with his five axes, distilled into a single, memorable test: would you trust an agent to buy this more than once? A $1 Trillion Market, With Fine Print The ICSC/McKinsey forecast is ambitious, but it comes with nuance that often gets dropped in the headline. Physical retail isn't going away; nearly 40% of Gen Z and millennials surveyed expressed a preference for experiential retail, and younger consumers are increasingly using digital channels to research before completing purchases in-store. Agentic commerce may handle the transactional layer while stores evolve to deliver the emotional, experiential layer that agents can't replicate. McKinsey partner Colleen Baum put it plainly: AI isn't eliminating the store, it's raising the bar for what stores need to deliver. That framing points to a more strategic question for retail leaders: rather than asking "should we invest in agentic commerce?" the more useful question may be "which parts of our customer journey should an agent own, and which parts should a person own?"