Agentic Commerce: How AI Agents Will Shop, Book, and Buy on Your Behalf
For years, AI chatbots could help you shop but not finish the job — compare headphones, find a flight, then hand you back to a checkout page to actually pay. That last step just closed. As of mid-2026, an AI agent can complete the purchase itself: find the product, negotiate within your constraints, and pay — with your approval encoded in advance rather than clicked in the moment.
This is agentic commerce, and 2026 has been the year it moved from experimental to infrastructural. eMarketer forecasts AI platforms will drive $20.9 billion in retail spending in 2026 — nearly four times 2025's figure. This article explains what actually launched, what already failed and why, the protocol landscape every retailer now has to navigate, and what agentic commerce means for businesses on both sides of the transaction — buying and selling. Written for product leaders, e-commerce teams, and technical decision-makers assessing where to invest.
What Changed in 2026 — From Assistant to Buyer
The mechanical shift is simple to state: the agent now holds the wallet. You link a payment method, set spend caps and merchant limits, and the agent completes purchases within those bounds — with approval prompts for anything outside them.
On 10 June 2026, Visa integrated its payment network directly into ChatGPT, letting agents complete purchases at any Visa-accepting merchant using tokenised credentials, real-time authorisation, and fraud monitoring bound by user-set limits. Mastercard followed with Agent Pay for Machines (AP4M) — a network letting agents and connected systems initiate and settle transactions at machine speed, down to fractional-cent microtransactions, settled across cards, bank accounts, and even stablecoins. Google, meanwhile, has been shipping checkout directly inside AI Mode and Gemini, letting purchases complete from eligible retailers via Google Pay without leaving the chat interface — built on Google's own Universal Commerce Protocol (UCP), developed with Shopify, Etsy, Walmart, and Target, and endorsed by over 20 companies including Amex and Visa.
The pattern across all three: payment networks, not AI labs, are becoming the trust layer. That detail matters more than it first appears — see the failure story below.
The Honest Part of the Story — What Already Failed
No serious account of agentic commerce in 2026 skips this: OpenAI's first attempt, Instant Checkout, launched with Etsy in September 2025 and expanded toward Shopify's merchant base — but struggled badly. It charged merchants a roughly 4% transaction fee on top of standard payment processing (on a $100 order, that meant close to $7.20 in combined platform and processing fees), and adoption stayed minimal — fewer than 30 of Shopify's millions of merchants went live before OpenAI scaled the feature back in March 2026.
The lesson is instructive for anyone evaluating this space: AI-native checkout is a genuinely hard integration problem, not a thin wrapper over existing payment rails. Messy, fragmented product data across merchant systems has been repeatedly cited as the practical bottleneck slowing AI-driven commerce — the unglamorous plumbing problem, not the AI, is what stalls deployments.
The Visa–OpenAI deal in June 2026 is best read as the industry's correction: routing agent payments through the card networks merchants already trust and already integrate with, rather than a new fee layer bolted on top. That structural difference — familiar rails versus a novel toll — is a large part of why analysts expect broader adoption this time.
The Protocol Landscape — What Retailers Actually Need to Support
By early 2026 an emerging technical consensus had formed, reported by The Register as: "MCP for tool integration, A2A for agent communication, UCP + AP2 for e-commerce." In practice, several protocols now matter for any commerce business:
| Protocol | Backed by | Purpose |
| ACP (Agentic Commerce Protocol) | OpenAI + Stripe | Open protocol for conversational discovery and checkout inside ChatGPT |
| UCP (Universal Commerce Protocol) | Google, with Shopify, Etsy, Walmart, Target | Spans discovery through checkout to post-purchase, portable across any agent that implements it |
| AP2 (Agent Payments Protocol) | Payment-authorisation layer beneath UCP; already adopted by providers like Walmart's OnePay | |
| AP4M (Agent Pay for Machines) | Mastercard | Machine-speed transaction initiation and settlement, including microtransactions |
| Visa Intelligent Commerce | Visa | Tokenised payments, Agent Score trust scoring, an Agentic Registry, and stablecoin settlement |
What Agent-Ready Commerce Actually Requires
Supporting agentic checkout is an integration project with concrete technical requirements, not a marketing toggle:
A structured, current product feed. ACP-style integration means pushing a compressed product feed (title, description, price with ISO currency code, availability, images, eligibility flags) to the agent platform on a defined update cadence — typically daily. Stale or incomplete feeds are the single most common reason agentic integrations underperform; an agent cannot sell what it cannot see accurately. A checkout API the agent can call. Beyond product discovery, the agent needs an API surface to create orders, apply the authorised payment method, and receive confirmation — structurally similar to the AppFunctions pattern we've covered for Android, applied to commerce. Payment integration with the relevant network. Whether that's Stripe (ACP), Google Pay (UCP/AP2), or direct Visa/Mastercard agent rails depends on which agent ecosystems you are targeting. Fraud and authorisation logic built for non-human buyers. Visa's Agent Score — a trust-scoring capability for agents themselves — signals where this is heading: fraud systems will need to reason about agent behaviour patterns, not just human ones, and existing fraud detection architectures (see our fraud detection guide) will need new features entirely.The Trust and Liability Question Nobody Has Fully Answered
Every serious account of agentic commerce in 2026 flags the same open question: who is liable when an agent buys the wrong thing? Wrong size, wrong quantity, a subscription the user didn't intend to renew, a purchase that exceeded the spirit if not the letter of a spend cap.
The current answer is procedural rather than legal: spend caps, merchant category limits, and approval prompts are configurable and revocable by the user — the guardrail pattern our own agent architecture guide describes as essential to any consequential autonomous action. But procedural guardrails are not the same as a settled liability framework, and expect this to be an active area of commercial dispute, regulatory attention, and contract-law development over the next several years — not a solved problem in 2026.
What This Means for Different Businesses
If you sell direct-to-consumer: UCP and ACP integration can let you bypass marketplace gatekeepers entirely, routing high-intent agent-driven traffic straight to your own checkout — reducing platform dependence rather than increasing it. If you sell on Amazon primarily: the near-term lever is optimising for Amazon's own AI shopping assistant (Rufus) — reviews, Q&A depth, and price competitiveness — while watching for Amazon's own agent-partnership terms to formalise. If you sell across multiple channels: budget for maintaining distinct product feeds, checkout flows, and attribution models per agent ecosystem. This fragmentation is, as one analysis put it, simply the cost of omnichannel commerce in 2026 — and it is real engineering work, not configuration. If you are evaluating whether to build at all: the Instant Checkout story is the cautionary data point. A thin, fee-heavy integration bolted onto an existing storefront underperformed badly. A properly engineered integration — clean structured data, a real checkout API, and payment rails buyers already trust — is what the second wave of deployments is being built on.A Readiness Checklist
- Product feed can be generated in structured, machine-readable format on a daily cadence
- Checkout can be exposed as an API, not only a rendered web page
- A position taken on which protocol(s) matter for your customer base — UCP, ACP, or both
- Fraud detection reviewed for agent-initiated transaction patterns specifically
- Legal review of spend-cap, approval, and dispute handling for agent-authorised purchases
- A realistic view of integration cost — this is systems engineering, not a plugin install
Conclusion
Agentic commerce crossed from experiment to infrastructure in 2026 — not because the AI got dramatically smarter, but because payment networks built the trust rails that AI-native checkout was missing on its first attempt. The failure of Instant Checkout and the rapid correction that followed is the most useful data point in the whole story: this is a genuine systems-integration problem, with real winners and real cautionary tales already on the record within a single year.
For businesses deciding whether and how to participate, the work is concrete — structured product data, real checkout APIs, protocol selection, and fraud and liability design for non-human buyers. None of it is exotic engineering, but all of it needs to be done properly the first time.
If your organisation is evaluating agentic commerce integration — or building the broader AI agent capabilities that make it possible — NetConsulate designs and builds agent-ready commerce architecture, from product feed and checkout API design to fraud detection tuned for agent-initiated transactions.
Evaluating an agentic commerce integration? Submit a proposal request and our team will respond with a tailored approach within 2 business days.