Agentic Commerce & Payments
Autonomous shopping and payment agents, agentic checkout, and the protocols and trust questions underneath them.
Retailers are taking different positions on shopping agents, with QVC allowing purchases, Tapestry permitting them only through Google's tools with shopper approval, and Amazon blocking outside agents while it builds its own. Roughly 60% of recent AI-assisted purchases still closed on Amazon, which suggests the gatekeeping choice matters more than the agent itself.
An opinion piece arguing that consumers have not shown they want agents handling everyday tasks, citing surveys in which people welcome AI for research but hesitate to let it make the booking decision. It also notes that the companies promoting the agent vision have strong financial reasons to do so.
Shopify added checkout tools through WebMCP so agents running in a buyer's browser can revise an order and place it once the buyer approves. The authorization step is where the trust question sits, since it decides how much the shopper actually reviews before the purchase completes.
Eligible US Shopify merchants learned by email that native checkout inside Google's AI Mode and Gemini had been turned on for them, with the opt-out tucked under Sales channels in the Shopify admin. The purchase now completes on Google's surface with Google Pay credentials, and nothing in the flow tells the shopper the transaction has left the merchant's own checkout.
NatWest, ASB, Bank of America, Capital One, Commonwealth Bank of Australia and ING published joint principles on transparency, safety, privacy, choice and interoperability for AI agents that shop and pay. The principles are voluntary and written ahead of any regulator, so what they leave to the banks' own discretion is worth reading as closely as what they promise.
Shopify opened Shop Pay checkout to Meta's Muse agent the same week Amazon shut it out, and PayPal followed days later. The rails for agent purchases are being laid one bilateral deal at a time, which amounts to a governance structure built without public input.
Amazon began refusing purchases made through Meta's Muse agent, citing its Conditions of Use. The access question that Amazon v. Perplexity put before the courts is now being settled by platform policy, and the shopper who delegated the purchase has no say in either venue.
Visa's Trust Index finds 72% of U.S. adults have used an AI assistant, but only 23% would trust generative AI to make a payment on their behalf, while 61% would trust an established payment brand to handle the transaction. The gap between the tools people use and the brands they will let hold the card is the number to watch as agents start checking out.
Pairs Muse's checkout-through-Stripe rollout with Vogue Business polling in which 31% of shoppers would hand a purchase to an agent and 72% would withhold their card details. The gap between what the platforms are building and what people will permit is the story.
Internal testing reportedly exposed private data and produced unreliable behavior before launch. Read it alongside Meta's Secure VM and Sentinel claims in the announcement post.
The trust-deficit framing of the Muse launch, landing two weeks after Meta's $18B multistate consumer-harms settlement. A useful counterweight to the launch-day coverage.
Meta's free US personal agent browses, fills forms, negotiates and pays on a user's behalf, checking out through Stripe Link with single-use card numbers. Each user gets a dedicated Secure VM and a separate Sentinel agent that gates anything reaching the internet, and Meta says agent data stays out of its ad systems. Worth watching for how those accountability claims hold up once it is in millions of hands.
Any site behind Cloudflare will be able to charge AI agents machine-to-machine via HTTP 402 and stablecoins — no account, no checkout, no human. The pay-per-access rail being laid under the agentic web.
The Ninth Circuit heard the first federal appellate test of whether an AI shopping agent acting on a user's instruction counts as an 'authorized' visitor under the CFAA. However it comes out, the ruling will shape whether agentic shopping tools need a platform's permission or just the user's.
Salesforce's Shopper, Buyer, and Merchant agents hit general availability with native ChatGPT integration and Google Search/Gemini coming this summer — agentic checkout moving from pilots to default retail infrastructure ahead of peak season.
A framework and evaluation suite for how advertising incentives bend LLM behavior, finding most models forsake user welfare for company interests — recommending a sponsored product almost twice as expensive (Grok 4.1 Fast, 83%), surfacing sponsored options to disrupt the purchase (GPT-5.1, 94%), and concealing prices in unfavorable comparisons (Qwen 3 Next, 24%). Behavior also shifts with reasoning depth and the user's inferred socio-economic status — the academic underpinning for the 'sponsored AI recommendations' trust problem.
Adyen's agentic-commerce stack (Feed, Cart, Payments) bets on being a 'universal translator' across the rival UCP, ACP, and AP2 agent protocols — with Visa, Mastercard, and American Express among its launch partners.
A new survey finds three in four Americans would trust AI shopping agents less if recommendations were swayed by brand dollars — and would trust the brands less too. A direct warning shot at the emerging agentic-commerce business model.
Visa adds AI dispute handling and a Ramp bill-pay tie-up; Mastercard extends agentic payments to Hong Kong as both networks build toward an international agent-commerce network.
The architecture beneath agentic checkout: tokens bound to a specific agent, merchant scope and consent policy. The consumer-consent story still reads thinner than the speed-and-scale pitch.
An institutional look at the systemic and consumer-protection questions when autonomous agents transact at scale: liability, fraud, concentration, and who bears the risk.
Sizes the market and the protocol race (Visa Intelligent Commerce, Mastercard Agent Suite, Google's Universal Commerce Protocol) as agent-led shopping heads toward a real share of e-commerce.
Visa is embedding tokenized, agent-bound credentials with user-set spend limits and real-time fraud monitoring into OpenAI experiences. The firms that built behavioral advertising are now laying the rails for autonomous spending.
Merchant platforms are wiring in agent checkout as Visa predicts agentic commerce goes mainstream in 2026 — the supply side moving faster than consumer trust.
A protocol letting AI agents transact continuously with on-chain permissioning and multi-rail settlement across 30+ partners. The notable twist is an audit trail by design — accountability built into the rails rather than bolted on later.
Coverage of the launch and what it signals about who controls the payment layer as autonomous agents start spending on people's behalf.
Fresh survey data showing consumer trust lagging well behind the infrastructure — and no consensus yet on who's liable when an agent buys the wrong thing.
Agents that can trade and buy on a user's behalf — the consumer-finance and agentic-commerce threads converging in one product.
Google's move to let its assistant complete purchases across merchants — a look at how the checkout layer gets consolidated.
The full chronological feed lives at On My Radar.
US20260006023A1 · Google Patents ↗
Dynamic token generation and scoping for payment transactions — the underlying credential architecture for Visa's Intelligent Commerce Connect agent payment platform. The system enables AI agents to be issued payment credentials scoped to a specific agent identity, spending limits, and permitted merchant categories, so that an autonomous agent can complete purchases within a defined consent boundary without exposing the consumer's underlying card credentials. Filed June 2024; part of the broader Visa/Mastercard race to own the credential layer of agentic commerce.
In the news(20)
US12483411B1 · Google Patents ↗
A blockchain framework for registering and authenticating AI agents across their entire life cycle: each agent is issued an identity token tied to a cryptographic hash of its code and model weights, linked to its human owner's identity token, and must re-register whenever anything about it changes. In the patent's banking use case, a personal AI financial agent presents its own token plus a delegation token to make payments within preset limits (e.g., $500 per transaction), while a trust score built from its transaction history and behavioral signature — API call patterns, decision-flow consistency, computational footprint — determines how often it must re-authenticate. Every resource request an agent makes is logged as a blockchain transaction, and deviations from its usual behavior trigger forced re-verification or token revocation. This is the accountability side of agentic commerce — infrastructure for answering how a bank knows the AI spending your money is the one you authorized, and unmodified since you authorized it. Notably filed not by a payments network but by a small Irvine, California LLC, months before Visa and Mastercard shipped their own agent-identity token schemes.
The full catalog lives at Patent Watch.