For as long as ecommerce has existed, the basic unit of a transaction has been a person clicking a button. Someone compares a few options, reads a review or two, picks a size, and types in a card number. Agentic browsing breaks that assumption. It refers to a new category of AI systems, often called agentic AI or AI shopping agents, that can search, compare, decide, and complete a purchase or booking with little or no human involvement at each step. The shopper sets a goal and a budget. The agent does the rest.
That shift sounds incremental until you watch it happen. In June 2026, Visa embedded its payment network directly into ChatGPT, allowing an AI agent to complete a purchase at any merchant that accepts Visa once a person links a card and sets spending limits. It is one of several moves, alongside OpenAI's Instant Checkout, Google's Agent Payments Protocol, and a wave of retailer-side assistants, that have turned agentic commerce from a thought experiment into something millions of people will encounter this year, whether they realize it or not.
This matters well beyond the checkout page. When an agent can shop, book, and transact on a person's behalf, it changes who the customer actually is, how trust gets established, who is liable when something goes wrong, and how businesses need to present themselves to be chosen by a piece of software rather than a person scrolling a page. For business leaders, marketers, and SMB owners, the practical question is no longer whether this is coming. It is how fast it will reach their category, and what needs to be true about their operations before it does.
From Recommendation to Execution: Why 2026 Is the Inflection Point
Agentic commerce did not arrive overnight. AI shopping assistants have existed for years, quietly summarizing reviews, suggesting products, and answering questions inside chat windows. What changed is the final step: execution. As one industry analysis put it, agentic commerce is delegated shopping, where a customer sets intent and guardrails and an AI agent compares options and completes the purchase on their behalf (nshift, 2026). The distinction matters because recommendation carries no financial risk. Execution does.
Consulting firm McKinsey frames this as a five-level “automation curve,” running from simple AI-assisted search at one end to fully autonomous, standing-goal agents at the other (McKinsey, 2026). At the lower levels, a person still approves each purchase; an agent might say “I found three flights under $400, which one?” Higher up the curve, a shopper authorizes rules rather than individual actions: “if groceries are under $120 and arrive by Friday evening, place the order.” At the top of the curve, agents operate against open-ended goals like maintaining airline status at the lowest cost over a year, or making sure a household never runs out of a particular item, stepping in for exceptions rather than every decision. McKinsey's research suggests adoption will not climb this curve uniformly. It will be shaped by ticket size, emotional stakes, and what the firm calls “regret risk”: the chance that a wrong call feels bad enough that a person wants to stay in the loop regardless of how capable the agent is.
The infrastructure behind this shift has matured quickly. A cluster of open protocols, including Google and Shopify's Universal Commerce Protocol, Google's Agent Payments Protocol, OpenAI's Agentic Commerce Protocol, and Anthropic's Model Context Protocol, are establishing a shared technical language so that agents do not need a custom integration with every merchant they might transact with (Charle, 2026; McKinsey, 2026). Google's Agent Payments Protocol, launched in September 2025 with backing from Mastercard, PayPal, American Express, Adobe, and Alibaba, uses cryptographically signed “mandates” that link a person's intent, the resulting cart, and the payment details into a single verifiable record. That detail is not a footnote. It is the mechanism that is supposed to let a merchant, a card network, and a customer all agree on what an agent was actually authorized to do, after the fact, if there is ever a dispute.
The scale of the projected shift varies depending on who is counting and what they are counting. McKinsey estimates agentic commerce could generate between three and five trillion dollars in orchestrated global retail revenue by 2030, with up to a trillion of that in the United States alone. Morgan Stanley puts agent-driven share of US ecommerce at ten to twenty percent by the same year, while Bain forecasts the US market specifically at three hundred to five hundred billion dollars. Gartner, focused on business purchasing rather than consumer retail, predicts that ninety percent of B2B purchases will be intermediated by AI agents by 2028, representing more than fifteen trillion dollars in spending (Digital Commerce 360, 2025). These numbers will not all prove right, and they measure different things: one counts checkout that happens entirely inside an AI platform, another counts the full value chain an agent touches. The consistent thread across every estimate is the direction, not the precise size.
What Actually Changes for the Business
The most immediate, practical change is who, or what, a business is actually selling to. A product page written to persuade a human browsing on a Tuesday night is not the same artifact that helps an AI agent decide whether to add an item to a cart. Agents do not respond to emotional copy, aspirational photography, or scarcity banners the way people do. They parse structured data: price, availability, delivery window, return policy, and product attributes. According to McKinsey's European retail research, when delivery windows, shipping costs, and return terms are unclear or inconsistent, an agent can simply skip the offer without a human ever seeing it (McKinsey, 2026). This has produced a new discipline some are calling generative engine optimization, or GEO, treated by McKinsey as a sibling to traditional SEO and, in their view, already a foundational layer of agentic commerce.
This is not a small adjustment to a metadata field. It means rethinking product catalogs as machine-readable assets: structured attributes, current inventory, accurate pricing, and consistent taxonomy, ideally aligned to open standards like Schema.org so an agent can reliably compare a listing against a competitor's. Businesses with incomplete or inconsistent product data are not just ranking lower in search results; they may never be considered by an agent doing the comparing at all.
The second change is in what “winning a customer” means. In a level four or five world on McKinsey's automation curve, the competitive battle is not for a single sale but for a standing place inside an agent's ongoing plan, the recurring decision logic that governs how household essentials get restocked or how a loyalty program gets optimized over a year. That reframes loyalty programs, subscription mechanics, and substitution policies as competitive infrastructure rather than back-office detail, because an agent reasoning through trade-offs needs those policies to be visible, predictable, and ideally exposed in a machine-readable way.
Third, the operational risk profile of a “customer” changes. Payment and fraud systems were built around a human-in-the-loop model in which identity, intent, and authorization are explicit and observable at the moment of purchase, according to McKinsey's analysis of the commerce opportunity (McKinsey, 2026). An agent acting on someone's behalf disrupts that model directly. The customer is now software, executing under a delegated authorization that has to be checked, scoped, and revocable, rather than a person typing a CVV code. This is forcing a new layer of identity and compliance work that the industry has started calling “know your agent,” an extension of know-your-customer standards built for a world where the buyer at the other end of a transaction might not be human at all.
Travel, Hospitality, and the Multi-Step Booking Problem
Retail tends to dominate the agentic commerce conversation, but travel and hospitality face a distinct version of the same shift, and arguably a harder one. Keith Mercier, Microsoft's vice president of worldwide retail and consumer goods, has noted that while retail may see the most immediate impact from agentic AI, travel is “close behind,” because agents are well suited to manage planning, pricing changes, and multi-step bookings that unfold over time rather than resolving in a single transaction (FinTech Futures, 2026).
That multi-step quality is exactly what makes travel both promising and risky as an agentic use case. A flight booking, hotel reservation, and ground transportation plan are not independent purchases; they are interdependent commitments where a delay in one cascades into the others. An agent that books a connecting flight without accounting for a hotel's cancellation window, or that selects a property in a way that overlooks safety or accessibility considerations a person would have caught instinctively, raises the liability question in its sharpest form. Researchers examining the ethics of agentic AI have used almost exactly this scenario to illustrate the underlying problem: if an agent autonomously books a non-refundable flight or selects a hotel in a high-risk area and something goes wrong, the question of who bears responsibility, the user who set the goal, the company that built the agent, or the platform that executed the transaction, does not yet have a settled answer (arXiv, 2025). Informed consent becomes genuinely difficult to define when the reasoning behind a multi-step booking decision is not fully visible to the person who authorized it.
This is precisely why McKinsey's research finds that consumer comfort with delegation tracks not just convenience but reversibility. People are far more willing to hand over a decision when the outcome is easy to undo, clearly authorized in advance, and simple to audit afterward, and far less willing when the action is hard to reverse or opaque about how it happened (McKinsey, 2026). Travel bookings sit at the harder end of that spectrum precisely because so many of them are non-refundable by design. Businesses in travel and hospitality that want to be agent-ready will need cancellation, modification, and substitution policies that are explicit enough for software to reason about safely, not just polished enough for a human to read on a confirmation email.
Trust Is the Real Bottleneck, Not Capability
If the technology to let agents transact already exists and is improving quickly, what is actually holding adoption back is trust, and the data on this is strikingly consistent across independent surveys. Adyen's research found that in the United States, the share of shoppers using AI shopping assistants more than doubled from twelve to thirty-five percent in a year, and just over half said they would let an agent handle an entire purchase process, including the final payment. Yet a separate analysis citing Checkout.com data found that forty-two percent of consumers worry about losing control over what an agent buys, and twenty-eight percent cite a lack of transparency into how it made its decision (Mohammed Shehu, 2026). Trust is highest among younger consumers and drops sharply for transactional steps: one estimate puts overall US trust in AI to place an order on someone's behalf at only fourteen percent, even though trust in AI for research and comparison is far higher (commercetools, 2026).
There is also a clear preference for who runs the agent. A Bain survey cited in industry research found that consumers trust retailer-owned agents roughly three times more than third-party agents to complete a transaction. That is a meaningful strategic signal: a brand's own AI assistant, built on its own data and accountable to its own policies, may have a structural trust advantage over a general-purpose agent acting across many merchants, even if the general-purpose agent is more capable.
This trust gap is not just a consumer phenomenon. It shows up just as starkly inside the businesses building these systems. McKinsey's 2026 AI Trust Maturity Survey, conducted across roughly five hundred organizations between December 2025 and January 2026, found that the average maturity score for responsible AI practices rose only modestly, to 2.3 out of a possible five-point scale, up from 2.0 the year before (McKinsey, 2026). A separate Harvard Business Review Analytic Services study, surveying more than six hundred business and technology leaders, found that only six percent of companies fully trust AI agents to autonomously run their core business processes, with the largest share restricting agents to limited or routine operational tasks (HBR Analytic Services, 2025). Kim Huffman, chief information officer at Workiva, observed in that research that the change management and reskilling required across organizations adopting agentic AI has been broadly underestimated. The pattern across both studies is the same: confidence in what the technology can do is rising faster than confidence in the organization's own ability to govern it.
McKinsey's survey points to a specific and telling imbalance. Nearly two-thirds of respondents cited security and risk concerns as the top barrier to scaling agentic AI, well ahead of regulatory uncertainty or technical limitations, which suggests companies are less worried about whether agents can perform the task than about whether they can be trusted to perform it safely at scale (McKinsey, 2026). That distinction, between capability and governance, is likely to define which businesses move quickly and which stall in pilot mode for another year or two.
The Accountability Question Nobody Has Fully Answered
The trust gap is not irrational. It reflects a genuinely unresolved question: when an autonomous agent buys, books, or transacts and something goes wrong, who is responsible?
Traditional consumer protection law assumes a person made the decision being disputed. An agent complicates that at every step. Legal researchers writing on agentic AI accountability describe this as a “responsibility gap,” a term coined to describe situations where an autonomous system makes a decision that even its own developers could not fully predict, which makes it difficult to assign moral or legal responsibility through the usual channels (arXiv, 2025). A McKinsey partner, quoted in industry coverage of the firm's trust research, framed the underlying shift sharply: agency is not simply a new feature, it is a transfer of decision rights, which means the central governance question changes from asking whether a model gave an accurate answer to asking who is accountable when the system takes an action (AgentMarketCap, 2026).
Regulators are beginning to respond, unevenly and from different angles. Colorado's AI Act, which takes effect in 2026, requires companies deploying high-risk AI systems to conduct annual impact assessments and maintain risk management programs. California's Civil Rights Council finalized rules on automated decision systems that extend record retention requirements and require employers to ensure such systems do not discriminate. The EU AI Act's rules for general-purpose AI models, which underpin many commercial agents, are already in force. None of these frameworks were written with agentic commerce specifically in mind, and legal commentary on the topic notes that liability exposure already exists in practice, regardless of whether dedicated regulation has caught up (Baker Botts, 2026).
This is where the technical infrastructure described earlier, the signed mandates, audit trails, and scoped payment tokens built into protocols like AP2, is doing double duty. It is not only a fraud-prevention mechanism. It is the closest thing the industry currently has to an accountability record: a verifiable trail showing what a person authorized, what an agent decided, and what a merchant executed, which becomes the evidence base if a dispute ever needs resolving. Businesses adopting agent-facing commerce tools should treat the quality of that audit trail as a governance requirement, not an engineering afterthought.
What Businesses Should Actually Be Doing Now
Given the gap between hype and operational readiness, a few priorities stand out as genuinely useful rather than aspirational.
The first is data hygiene, treated as a commercial priority rather than a technical one. Structured, accurate, real-time product and policy data is now a precondition for being considered by an agent at all, not just a way to rank better in a search result. This includes pricing, inventory, delivery timing, and crucially, the kind of policy detail, return windows, cancellation terms, substitution rules, that agents need to reason about trade-offs on a customer's behalf.
The second is building or selecting agent-facing commerce infrastructure with auditability built in from the start, rather than retrofitted after an incident. Authorization that can be limited by budget, merchant, or category; activity logs that can reconstruct what an agent did and why; and mechanisms to reverse or override an action quickly are no longer optional extras. McKinsey's research is explicit that these three properties, limitable authorization, auditable activity, and reversible action, are exactly what determines whether consumers feel comfortable delegating further (McKinsey, 2026).
The third is being honest about where an organization actually sits on governance maturity before expanding what agents are allowed to do. The McKinsey research on agentic governance found that companies investing twenty-five million dollars or more in responsible AI initiatives reported meaningfully higher maturity scores and a far higher likelihood of seeing real EBIT impact from their AI investments, suggesting governance spending functions as an enabler of value rather than a tax on it (McKinsey, 2026). For an SMB that cannot realistically match that level of investment, the more proportionate move is narrowing the scope of what any agent is authorized to do, rather than under-investing in oversight while still expanding agent permissions.
Finally, businesses should expect this to roll out unevenly by category, not as a single wave. McKinsey's category-level research found wide variation in current AI-assisted shopping adoption, with categories where physical fit and in-person assessment matter, eyewear was their specific example, lagging well behind categories that resolve cleanly through comparison and specification. Planning for “agentic commerce” as a single undifferentiated trend risks missing exactly where in a product line the shift will actually bite first.
Counterarguments and Real Limits
It is worth taking seriously the case that agentic commerce is being overbuilt relative to where trust and reliability actually stand today. Gartner has predicted that more than forty percent of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls, a notably blunt assessment from a firm that is otherwise bullish on the category's long-term scale (Gartner, 2025). Gartner analyst Anushree Verma has argued that most agentic AI projects underway right now are early-stage experiments driven more by hype than by a clear-eyed read of deployment cost and complexity.
There is also a meaningful behavioral counterpoint to the idea that delegation will simply keep expanding. McKinsey's own automation curve research stresses that delegation does not move uniformly upward. People are demonstrably more willing to hand over the cognitive work of comparing options than the actual authority to spend money, and for many categories, that ceiling on delegation is durable rather than temporary; it reflects the fact that human involvement is sometimes intrinsic to the value of the purchase itself, not just friction waiting to be automated away. A wedding dress, a first car, a gift for a parent: these are not simply transactions waiting for a more efficient interface.
Finally, the fragmentation of competing protocols, ACP, UCP, AP2, A2A, MCP, each backed by different combinations of technology and payment companies, mirrors the early, messy years of other infrastructure shifts, where standards consolidation took years longer than the underlying technology took to mature. Businesses betting heavily on any single protocol today are making a bet on which coalition wins that consolidation, not just a bet on agentic commerce itself.
Where This Goes From Here
The realistic near-term picture is not a sudden mass handover of purchasing authority to software. It is a gradual, uneven climb up McKinsey's automation curve, fastest in categories with low emotional stakes and high comparability, such as commodity household goods and routine subscriptions, and much slower in categories defined by identity, trust, or irreversible commitment, such as major travel bookings or significant purchases. Layered on top of that gradual climb is a parallel, slower-moving build-out of the legal, governance, and technical trust infrastructure that will determine how much authority people are actually willing to extend, and how quickly businesses can extend it back to them with confidence.
For business leaders, the practical takeaway is not to chase agentic commerce as a feature to bolt on, but to treat it as a forcing function for two things that were already overdue: making product and policy data genuinely machine-readable, and building governance that can survive an audit before an incident forces the question. The agents are already capable of shopping, booking, and transacting. Whether businesses, and the people they serve, are ready to trust them with that authority is the part still being worked out, deal by deal, category by category, through the rest of 2026 and beyond.
Frequently Asked Questions (FAQ)
Agentic commerce describes AI agents that can search for products or services, compare options, and complete a purchase or booking on a person's behalf, often within budget and policy limits the person sets in advance, rather than simply offering recommendations for a human to act on.
It is becoming safer through mechanisms like scoped payment tokens and cryptographically signed authorization records, which limit what an agent can spend and create an audit trail of what it did. That said, consumer trust remains low for fully autonomous purchasing, and most current systems still involve some form of human approval at checkout.
Retail and B2B procurement are currently furthest along, followed closely by travel and hospitality. Categories involving routine, low-stakes, easily compared purchases are moving faster than categories involving high emotional stakes, physical fit, or irreversible commitments.
This is still legally unsettled. Liability could fall on the consumer who authorized the agent, the company that built it, or the platform that executed the transaction, depending on the circumstances, the terms of service involved, and how courts and regulators eventually interpret emerging AI accountability frameworks.
Prioritize accurate, structured, and current product data, including pricing, inventory, and policy details such as returns and cancellations, since agents skip listings they cannot confidently evaluate. Businesses should also consider what level of autonomous authorization they are willing to extend to agents acting on a customer's behalf, and ensure those interactions are auditable.
Most research suggests not in the near term. Forecasts generally place agent-driven share of ecommerce in the ten to twenty-five percent range by 2030, implying a majority of shopping will continue to involve direct human browsing and decision-making, particularly for high-consideration or identity-driven purchases.
Conclusion
Agentic browsing is no longer a speculative scenario confined to product demos. With Visa's payment network now embedded directly into conversational AI, open standards like AP2 and UCP gaining real backing from major payment and retail players, and trillions of dollars in projected commerce flowing through autonomous agents within the decade, the shift from human clicks to AI-mediated transactions is already underway. What separates the businesses that benefit from this transition from those that get left behind is not how quickly they adopt agentic commerce technology, but how seriously they treat the trust, data, and governance work that has to happen underneath it. The agents can already shop, book, and transact on your behalf. The open question, for businesses and consumers alike, is how much of that authority anyone is actually ready to hand over.
Aadarsh Senapati
AI enthusiast · Writer · Developer
Bhubaneswar, Odisha, India
Rate This Article
Leave a Comment