Sometime while you sleep tonight, a machine may spend your money โ legally, with your permission, using a card you enrolled. That stopped being science fiction on September 8, 2026, when Meta launched Muse, a personal AI agent that connects to your email, calendar, shopping and travel apps and completes purchases through Stripe’s Link checkout. To start, Meta requires one thing: a payment card on file.
Two days later, Visa published the most revealing consumer survey of the AI era. Roughly three quarters of Americans have already used an AI agent โ OpenClaw, Hermes, ChatGPT, Muse. But only 23% trust an agent with their money โ a figure that nearly triples to 61% when a trusted payment brand stands behind the purchase. People have handed AI their attention and are refusing, for now, to hand it their wallets.
“Your AI just bought you a couch. Are you okay with that?” asked Jack Forestell, Visa’s chief product and strategy executive. It is the right question, because the industry’s answer this month is a product blitz: Anthropic shipped an open commerce-agent blueprint on September 2, timed for the holidays. Google began rolling out Universal Commerce Protocol (UCP) checkout inside Search and Gemini. Mastercard projects ยฃ370 billion a year in AI-driven consumer spending by 2030. Ant International partnered with both card networks on agent verification โ and on September 12, OpenAI and Visa announced their own partnership to let AI agents shop and pay.
The counter-evidence is just as real. OpenAI’s flagship in-chat buy button, Instant Checkout, lived exactly 156 days before being quietly discontinued. Only about 3% of transactions at merchants testing agentic commerce actually involve an agent. And on the crypto rails built for machine payments, an independent audit found just 0.6%โ7.5% of the money plausibly came from AI agents โ the rest was ordinary scripts.
If you freelance or run a one-person business, you should care more than most: you are the procurement department, the CFO and the fraud department of your own company. This is the money story, not the tech story: what each platform does, what the trust data say, who eats the loss when your AI buys wrong, what it costs, and a nine-move playbook for using these tools without getting burned.
What an AI Shopping Agent Actually Is โ and What It Isn’t
An AI shopping agent carries a purchase instruction all the way through: it discovers products, reads structured price and inventory feeds, calls a checkout built for machines rather than browsers, and passes a fraud-and-identity layer that separates your authorized agent from a bot with a stolen card. A chatbot that recommends a laptop is not an agent. An agent buys the laptop.
In practice, 2026’s dominant pattern is more timid than the demos: analysts call it “discover in AI, buy on site.” Chatbots have gotten good at research and comparison, while final payment keeps drifting back to the merchant’s checkout page. None of the major consumer platforms currently allows a fully autonomous purchase without explicit confirmation โ the industry is shipping handbrakes on purpose. Underneath, a protocol war is deciding who owns the plumbing:
- ACP (Agentic Commerce Protocol) โ OpenAI + Stripe; ties merchant feeds and Stripe payments into ChatGPT’s surfaces.
- UCP (Universal Commerce Protocol) โ Google with Shopify, Etsy, Walmart and Target; agent-agnostic, spanning discovery through post-purchase.
- AP2 โ Google’s agentic payments protocol.
- Visa Intelligent Commerce โ with the Trusted Agent Protocol, an Agentic Directory and an “Agent Score” for verifying that an agent belongs to a real customer.
- Mastercard Agent Pay and Agent Connect โ tokenized agent payments plus merchant tools launched this month.
- x402 โ Coinbase’s crypto-native rail for per-request stablecoin payments; Amazon’s AgentCore Payments (May 2026) builds on it.
Every one solves the same problem: how does a merchant know the agent at its checkout API is a real customer’s real agent โ and who is liable when it is not?
The September 2026 Landscape: Five Ways an AI Can Spend Your Money Today
1. Meta Muse: the most aggressive consumer bet
Muse, announced September 8 for U.S. users, is the first mass-market agent treating spending as a headline feature. It connects โ one app at a time, opt-in โ to email, calendars, payments, health, smart home, dining, shopping and travel apps. It can send emails, book travel, “lower bills,” turn recipe videos into grocery orders and make purchases, with checkout via Link by Stripe, which carries purchase protections (Shop Pay and 1Password are coming). It runs on Meta’s Muse Spark model inside a dedicated “Muse Secure VM,” keeps working after you close the app, and lives at muse.ai, on iOS and Android, inside WhatsApp, and eventually on Meta’s AI glasses.
The pricing is the story: a free tier with a usage meter, then Power at $20/month and Maximum at $100/month โ with a payment card required just to start. Context matters: Muse arrived less than two weeks after Meta agreed to an $18 billion multistate settlement (29 states) over social-media harms to children. Fortune’s verdict ran bluntly โ Meta’s new AI can shop, choose and check out, and most people don’t want it touching their money.
2. ChatGPT: the 156-day checkout that died
OpenAI moved first. Instant Checkout launched inside ChatGPT on September 29, 2025 โ tap “Buy,” confirm shipping, pay without leaving the conversation, on ACP rails with Stripe handling fraud review. On March 4, 2026 โ 156 days later โ OpenAI quietly discontinued it, moving the experience to its Apps model. ACP remains a live standard and ChatGPT keeps hundreds of millions of weekly users for discovery, but the retreat is the category’s most honest data point: in-chat autonomous checkout proved harder to sustain than any demo suggested. This month OpenAI re-entered via its September 12 Visa partnership.
3. Microsoft Copilot Checkout: the retailer-first play
Microsoft skipped the single buy button. Copilot Checkout is in U.S. testing on Copilot.com with a mixed payment stack (PayPal, Shopify, Stripe) and plans to extend across Bing, MSN and Edge. Its more interesting piece is Brand Agents: AI assistants Shopify merchants deploy on their own storefronts, trained on their catalog. The industry summary: ChatGPT has the users; Copilot has the retailers. Purchases require user confirmation.
4. Google Gemini and UCP: checkout inside Search
Google’s bet is distribution, not destination. UCP powers a new checkout on eligible product listings inside AI Mode in Search and the Gemini app (U.S. rollout began this month), with Shopify, Etsy, Walmart and Target as launch backers. Because UCP spans discovery through returns and is portable to any agent, it has the broadest scope of the competing standards โ and is the newest and least battle-tested.
5. Anthropic Claude Commerce Agents: the merchant-side blueprint
Anthropic arrived last with the opposite pitch. Released September 2, Claude Commerce Agents is an open blueprint (github.com/anthropics/commerce-agents) for retailers to build their own shopping and merchant-operations agents โ deliberately timed for holiday planning. There is no consumer buy button: Claude never places orders or charges cards; checkout, payment and fulfillment stay on the merchant’s infrastructure. The commercial case is striking โ retailers running shopping agents on Claude have seen carts up to 35% larger and shoppers 60% more likely to complete a purchase, with enterprise users including Shopify and Priceline, and Visa, Mastercard and Accenture as partners.
| Platform | Consumer buy button | Who owns checkout | Payment rails | Status (Sept 15, 2026) | Cost to you |
|---|---|---|---|---|---|
| Meta Muse | Yes โ purchases via Link by Stripe | Stripe Link (merchant fulfills) | Card on file required; Shop Pay + 1Password coming | Live, U.S.; WhatsApp/web/mobile; glasses next | Free tier; Power $20/mo; Maximum $100/mo |
| ChatGPT (OpenAI) | No โ Instant Checkout ended Mar 4, 2026 | Merchant site (“discover in AI, buy on site”) | ACP standard lives on; new Visa partnership | Discovery/comparison only inside chat | Free (within ChatGPT plans) |
| Copilot Checkout (Microsoft) | Yes, with confirmation โ U.S. testing | Merchant + Copilot surface | PayPal, Shopify, Stripe | Testing on Copilot.com; Bing/MSN/Edge planned | Free |
| Gemini / UCP (Google) | Yes, with confirmation โ eligible listings | Retailer, surfaced via Google | Google Pay-linked retailer rails; AP2 | Rolling out in Search AI Mode + Gemini app (U.S.) | Free |
| Claude Commerce Agents | None โ merchant-facing blueprint | Merchant, entirely | Merchant’s own stack | Open source; retail/travel/telecom/ticketing examples | N/A for shoppers |
๐ The 156-day lesson: Instant Checkout launched September 29, 2025 and died March 4, 2026 โ a full product cycle in five months. Independently tracked live AI-checkout integrations number roughly 30 retailers as of July 2026, against millions of merchants on Google Shopping or Amazon. When someone says AI shopping has arrived, ask which of those 30 stores they mean.
The Trust Gap: 23% Won’t Hand Over the Card โ 61% Will if Visa Stands Behind It
The Visa Trust Index survey of 2,065 U.S. consumers is the cleanest measurement yet of where the public stands. Almost three quarters have used an AI agent. Only 23% trust one to spend their money. Attach a trusted payment brand and acceptance jumps to 61%. That 38-point swing is not about technology โ it works โ but about who consumers want standing behind them when their agent has just bought a $50,000 sofa.
“AI has the potential to fundamentally reshape how people discover, buy and pay for goods and services, much like e-commerce and mobile commerce did before it,” Visa Group President Oliver Jenkyn said with the release. “Trust will be foundational.” Asked what consumers need first, Jenkyn named three things: an authenticated, tokenized payment credential, genuine personalization, and clear controls and permissions โ when it can buy, how much it can spend, where it can shop. “Yes in travel, no in healthcare. Yes below a certain threshold, no above it.”
Other measurements point the same way, with a consistent price ceiling: Worldpay finds only 14% of consumers trust AI recommendations without manual verification, and trust drops sharply once digital goods exceed $50. An Elon University / Washington Post poll found 27% of Americans bring personal or emotional questions to chatbots โ AI as confidant is accepted; AI as spender is not. And Mastercard’s European research (26,000 people โ 13,000 teens paired with 13,000 parents across 13 countries) projects one in ten consumers routinely using AI to shop and pay by 2030.
The generational split predicts your future customer base. Teen adoption of AI shopping assistants runs at 27% versus 16% for adults; 18% of teens use an assistant weekly to compare products, versus 9% of parents; and a third of teens would hand full control โ choosing and paying โ to a machine. If you sell anything to under-25s (see our read on Gen Z’s soft-saving money culture), their purchasing interface is becoming an agent, not a storefront.
| Metric | Number | Source (2026) |
|---|---|---|
| Americans who have used an AI agent | ~75% | Visa Trust Index (n=2,065 U.S.) |
| Who trust an agent to spend their money | 23% | Visa Trust Index |
| Trust when a major card network stands behind it | 61% | Visa Trust Index |
| Trust AI recommendations without verifying | 14% | Worldpay |
| Trust cliff for AI-bought digital goods | above $50 | Worldpay |
| Teen vs. adult AI shopping-assistant adoption (Europe) | 27% vs 16% | Mastercard / Forkast |
| Teens who would give an AI full choose-and-pay control | ~33% | Mastercard / Opinium |
| Americans using chatbots for personal/emotional queries | 27% | Elon University / Washington Post |
๐ The teen signal: Mastercard’s paired teen-parent survey is the industry’s clearest admission that adoption will not be linear. A third of teenagers are ready to delegate both the choosing and the paying. Futurologist Magnus Lindkvist, one of four contributing to the report, compares agentic commerce to the washing machine: an enormous giver-back of time on utilitarian tasks.
Reality Check: Only 3% of Transactions Actually Involve an Agent
The most sobering number in the space comes from payments processor Checkout.com: 42% of merchants are testing agentic commerce โ but only 3% of actual transactions involve an agent. Testing is cheap; trusting a machine with checkout is not. Forkast’s analysis notes 89% of companies say they are preparing for the shift โ an infrastructure gold rush running far ahead of consumer behavior.
The crypto rails tell the same story with harder numbers. TRM Labs audited x402 โ the protocol Coinbase launched in 2025 letting agents pay per-request in stablecoins (the same USDC rails many remote workers now get paid on). Since May 2025, x402 facilitators processed 198.9 million settlements worth $52.7 million โ USDC is 99.6% of settled value. After stripping self-payments and wash-like flows, only 0.6%โ7.5% of the remaining $25.62 million looked like genuine AI-agent activity rather than ordinary scripts. “The rail already works,” TRM concluded. “What is needed is accurate registration, counterparty reputation an agent can check on its own, and monitoring built for volume.” The registries that would assign responsibility are voluntary โ “and currently not utilized by the majority of participants.”
Merchants are not passive either: Amazon has spent a year blocking third-party agents from its catalog, a fight with Perplexity that has produced lawsuits. And forecasts diverge wildly โ Mastercard says one in ten shoppers by 2030 (roughly 300 million people, ยฃ370 billion a year); TheStreet reports Morgan Stanley’s view that it could be half. When serious models disagree by 5x, build habits that are robust either way. One wrinkle for crypto earners: machine-payment rails do not escape the tax net โ CARF and Europe’s DAC8 push automated reporting of crypto transactions to tax authorities, micropayments included.
Who Pays When Your AI Buys the Wrong Thing?
This is the question the industry cannot yet answer cleanly, and Visa’s own group president says so. Fraud detection “built for humans clicking buttons doesn’t necessarily transfer to AI transacting at machine speed,” Oliver Jenkyn told Forbes. “Neither does dispute resolution: if your agent buys the wrong thing, or the right thing at the wrong price, the chargeback framework that protects you today isn’t obviously fully applicable.”
That is the crux of consumer risk in 2026. Card-network zero-liability rules protect you against unauthorized transactions โ a thief, a cloned card. But an agent purchase is, by design, authorized: you enrolled the card, you set the agent loose, the merchant received a valid credential. If your agent misreads the instruction and buys the $900 wrong mattress, that is not fraud in the network’s eyes. It is you, shopping.
| What went wrong | Who likely eats it today | What actually protects you |
|---|---|---|
| Agent buys the wrong item, size or variant | You โ an authorized purchase under your return rights | Merchant return window; confirmation-before-pay settings; start small |
| Agent pays a glitched or hallucinated price | Usually the merchant cancels; messy if it ships | Screenshot the listing the agent acted on; keep the approval email |
| Agent session hijacked; unauthorized purchases | Card issuer โ zero-liability rules should apply | Virtual cards with hard caps limit the blast radius; report fast |
| Agent tricked by a fake listing or scam merchant | You โ authorized payment to a fraudster is the hardest chargeback there is | Merchant allowlists; never let agents pay unfamiliar sellers |
| Agent books nonrefundable travel on wrong dates | You, unless platform purchase protection covers it | Human confirmation on all travel; read protections before delegating |
| Item never arrives from a real merchant | Issuer via “goods not received” dispute โ winnable with a paper trail | Keep agent order confirmations like receipts |
The platforms know this is the adoption blocker. Muse routes checkout through Link by Stripe precisely because it carries purchase protections. Jenkyn’s answer is that clarity “has to be designed into the experience”: consumers must know what they approved, what protections apply, and who steps in when the agent errs. Visa is building that layer โ Trusted Agent Protocol, Agentic Directory, Agent Score โ while Mastercard ships Agent Pay tokenization and Agent Connect. Ant International joined both networks on September 10 to verify AI shopping agents, and Reuters reports India is planning an AI registry before rolling agentic payments onto UPI. The rails of accountability are being poured right now โ which is why you should not assume they exist yet. Until they do, the backstop is personal: a properly sized emergency fund that can absorb a $900 agent mistake without becoming a credit-card-debt spiral, and a credit file clean enough that one disputed charge never snowballs.
The New Fraud Surface: Fake Agents, Hijacked Sessions and Poisoned Listings
“The flipside of all this incredible technology is that the bad guys also have access to it,” Jenkyn says. Agent commerce creates three attack classes that did not exist when fraud models were tuned on humans clicking buttons.
1. Fake agents. A bot with a stolen card now has a legitimate-sounding costume: “I’m not card-testing, I’m the customer’s authorized shopping agent.” Merchants must tell your real agent from an impostor in milliseconds โ the exact problem Visa’s Trusted Agent Protocol and Agent Score, Mastercard’s Agent Pay verification, and the Visa-Mastercard-Ant partnership exist to solve. None of it is mature.
2. Hijacked and injected agents. Your agent reads the web. A poisoned product page โ hidden text instructing it to reroute payment or exfiltrate data โ is prompt injection with a price tag. Reuters ran wire coverage this month of OpenAI tracking malign activity involving rogue agents. And read the fine print: early testers of one assistant, Instinct, found it demanded a “perpetual and irrevocable” license to access, store, modify and use their materials, including for model training.
3. Your agent as a single point of failure. A consumer agent that works as designed holds your email, calendar, app credentials and a payment token. For a freelancer that is not a gadget โ it is the keys to the business. A compromised Muse session is a compromised company, and no travel insurance policy covers the downtime. Futurologist Patrick Dixon expects new cyber-attack risks to keep laggard consumers wary even as the U.S., U.K., China and South Korea lead adoption.
What AI Shopping Actually Costs You
Three of the five platforms are free to consumers because the merchant pays โ integration engineering, feed maintenance, processor fees. Meta is the outlier, and its pricing tells you how it sees Muse: not a shopping feature but a subscription product that happens to shop.
Do the freelancer math before enrolling. Muse Maximum costs $1,200 a year; Power costs $240. At a $50/hour billable rate, Maximum must hand you back 24 hours a year โ two hours a month of grocery lists, price tracking and form-filling โ just to break even, before counting the risk premium of letting it near a card. For most solo operators, the free discovery tiers of ChatGPT or Gemini plus your own checkout clicks win.
The sneakier cost is behavioral. Frictionless purchasing is lifestyle creep with a turbocharger: an agent that learns your tastes, watches prices and can buy inside a chat window removes every pause that ever talked you out of an impulse purchase. The no-buy challenge and moneymaxxing trends exist precisely because friction is a feature. And audit any agent subscription monthly โ an agent pitched on “lowering your bills” will not, by default, audit itself.
The Freelancer & Digital Nomad Playbook: 9 Moves to Use AI Shopping Agents Without Getting Burned
None of this is a reason to sit out. Agents are genuinely good at the tedious half of commerce โ comparison, monitoring, reordering, form-filling โ and Jenkyn’s prediction is that the first proving ground is “procurement, invoicing and reconciliation: B2B workflows that are rules-based, permissioned and boring.” That is a description of a freelancer’s week.
Move 1: Delegate discovery, keep the buy button
Use the pattern the industry converged on โ discover in AI, buy on site. Let agents research, compare specs, track prices and summarize reviews; complete payment yourself on the merchant’s checkout. You lose nothing (it is how ~97% of “agentic” transactions work anyway) and keep every consumer protection intact.
Move 2: Never link your primary card โ use capped virtual cards
This is Visa’s own first requirement (a tokenized credential, not your real card number) and you can enforce it today. Issue a dedicated virtual card for agent use with a hard per-transaction cap and low monthly ceiling, locked to allowed merchants. The neobanks that serve expats โ Revolut, Wise and friends โ generate disposable virtual cards in seconds; several of the top nomad credit cards do the same. If the agent is hijacked or hallucinates a merchant, your loss is capped at the virtual card’s balance, not your credit line.
Move 3: Write the guardrails before you write the prompt
Borrow Jenkyn’s rulebook literally: yes in travel, no in healthcare; yes below $X, ask me above it; yes at merchants I have used, no at new sellers. Set category allowlists, per-purchase caps and a monthly budget inside whatever controls the platform exposes โ then screenshot them. Documented permissions are permissions you can prove when a dispute turns on “what did the user approve?”
Move 4: Set your personal confirmation floor at $50
Worldpay’s data shows trust falling off a cliff above $50, and the platforms themselves still require confirmation for every purchase. Make it a hard personal rule: anything above $50 โ and anything nonrefundable, at any price โ comes back to you. An agent that cannot respect a $50 floor is an agent to uninstall.
Move 5: Start with boring replenishment, not big decisions
The highest-value, lowest-risk delegation for a solo business is the recurring and rule-bound: software renewals, coworking day passes, standard supplies, price-watching on gear you already want. It composes with our AI-proofing playbook: let the machine run procurement; keep judgment work human.
Move 6: Keep the approval trail like it is a tax document
For a freelancer, it eventually feeds one. Save the agent’s order confirmations, the listing it acted on and your permission settings where your receipts live (accounting software automates this). Two payoffs: dispute evidence when liability turns on what you approved, and clean deductions when agent-bought expenses land in your quarterly estimates.
Move 7: Budget the subscription like a nomad, not a tourist
If you pay for Muse Power or Maximum, log it in your monthly nomad budget next to coworking and insurance, and run the break-even honestly: at $100/month the agent must return two billable hours a week. Pair it with a budgeting app that flags new recurring charges, and cancel the first month it fails the test. Zombie subscriptions are the lifestyle creep of the agent era.
Move 8: Mind the tax tail on machine money
If your agent spends in stablecoin micropayments (x402-style), every disposal is a reportable event in most jurisdictions, and CARF/DAC8 reporting means tax authorities get the data whether you track it or not. And for U.S. persons abroad running purchases through FEIE-excluded income: the exclusion covers earned income, not record-keeping โ agent-made business expenses still need documentation in USD terms.
Move 9: Add the nomad layer โ currency, borders and time zones
Agents booking travel across borders inherit every trap in our hidden-fee guide: dynamic currency conversion, cross-border fees, and a machine that cannot eyeball the “pay in USD or EUR?” prompt. Instruct agents in writing to always charge in local currency, fund them from a multi-currency account, and remember: Muse, Copilot Checkout and UCP checkout are all U.S.-first โ your agent’s powers change when your passport stamp does.
๐ The playbook in one line: let the agent research everything, decide little, and touch money only through a capped virtual card, with a $50 confirmation floor, a written permission list, and every approval archived like a receipt. That sentence implements all three of Visa’s conditions โ tokenized credentials, personalization, explicit controls โ with tools you already have.
What Comes Next: Holiday 2026, Protocol Wars and the 2030 Forecast
The next ninety days are the category’s first real exam. Anthropic timed its blueprint for holiday planning; Microsoft’s Copilot Checkout expansion targets the same Q4 window; Google’s UCP checkout is mid-rollout; Muse will eventually reach Meta’s glasses. Expect the ACP-versus-UCP standard fight to sharpen โ Shopify is already hedging by supporting both โ and the accountability layer (registries, scores, directories, India’s planned UPI registry) to advance faster than the buy buttons, because that is the order trust demands.
| Date | Event |
|---|---|
| Sept 29, 2025 | OpenAI launches Instant Checkout in ChatGPT (ACP + Stripe) |
| Late 2025 โ early 2026 | Microsoft begins Copilot Checkout testing (U.S.) |
| Early 2026 | Google formalizes UCP with Shopify, Etsy, Walmart, Target |
| Mar 4, 2026 | Instant Checkout discontinued โ 156 days after launch |
| May 2026 | Amazon AgentCore Payments with Coinbase and Stripe (stablecoin agent rails) |
| June 2026 | Mastercard launches Agent Pay for machines; Checkout.com counts 42% of merchants testing, 3% of transactions agentic |
| Sept 2, 2026 | Anthropic releases open Claude Commerce Agents blueprint ahead of holidays |
| Sept 7โ8, 2026 | Mastercard’s ยฃ370bn-by-2030 research unveiled in London; Meta launches Muse ($20/$100 tiers) |
| Sept 9โ12, 2026 | Visa Trust Index (23%/61%); Ant International joins Visa and Mastercard on agent verification; OpenAI-Visa partnership reported; TRM Labs finds 0.6%โ7.5% of x402 volume is genuine agent activity |
| By 2030 | Mastercard: 1 in 10 shoppers routinely delegate; ~300M people; ยฃ370B annual AI-driven spend. Morgan Stanley (via TheStreet): possibly half of shoppers |
The long-run shape is easier to see than the timing. Futurologist Katja Forbes expects agents to network and negotiate at scale โ your agent haggling with a merchant’s agent โ and to migrate into wearables like glasses. Lindkvist’s caveat is the one to tape above your desk: “The vision of each major technological advancement is never its reality.” Agentic commerce will arrive, but in the permissioned, human-in-the-loop form the 3% transaction share already previews. The people who set their guardrails now profit from it later.
FAQ: AI Shopping Agents in 2026
Is it safe to let an AI agent use my card?
Only with containment. Card zero-liability rules cover unauthorized (hacked) transactions, but an agent buying the wrong thing with your permission is an authorized purchase with weak dispute rights โ Visa’s own leadership concedes the chargeback framework “isn’t obviously fully applicable.” Use a capped virtual card, set explicit permissions, require confirmation above small amounts.
Does ChatGPT still have a buy button?
No. OpenAI discontinued Instant Checkout on March 4, 2026, 156 days after launch, shifting shopping toward its Apps model. ChatGPT remains useful for discovery, with purchases completing on merchant sites. The September 2026 Visa partnership signals a future re-entry on verified-agent rails.
What does Meta Muse cost?
A free tier with a usage meter, then Power at $20/month or Maximum at $100/month. A payment card is required at signup, and checkout runs through Link by Stripe, which includes purchase protections.
Who is liable if my AI agent gets scammed or buys wrong?
In most scenarios today, you are. Authorized agent purchases fall under merchant return policies and your own documentation, not fraud protections โ the exceptions being hijacked sessions (issuer zero-liability should apply) and goods that never arrive (disputable with a paper trail). The accountability layer โ Visa’s Trusted Agent Protocol, Mastercard’s Agent Pay, the Visa-Mastercard-Ant verification partnership โ is being built now, but assume it is unfinished until you have read the specific terms.
Sources
- Visa Trust Index โ corporate.visa.com, “Trust powers agentic commerce” (n=2,065: 23% / 61%); Forbes, John Koetsier, “23% Of Us Trust AI Agents To Spend Our Money. Visa Has A Plan,” Sept 9, 2026 (Jenkyn interview)
- TechCrunch, “Meta debuts its Muse AI agent. Will consumers trust it?” Sept 8, 2026 (Muse launch details, pricing, Link by Stripe, settlement context, Instinct license concerns)
- Anthropic, “Building commerce agents with Claude” โ claude.com/blog/claude-for-commerce-agents, Sept 2, 2026 (+35% carts, +60% completion, partners, Forestell quote)
- Tech Insider, “ChatGPT vs Copilot vs Gemini: AI Shopping Agents 2026,” Sept 9, 2026 (Instant Checkout dates, ACP vs UCP, ~30 live integrations per the AI Checkout Tracker)
- Mastercard agentic AI research โ Retail Technology Innovation Hub, Sept 8, 2026 (Opinium: 26,000 respondents, 13 nations; 1-in-10 by 2030; ยฃ370bn; 18% vs 9%; one third of teens; futurologist contributions); Forkast.news, Sept 13, 2026 (27% vs 16%; 89% preparing; 300M shoppers); TheStreet (Morgan Stanley contrast)
- Checkout.com, “Agentic Commerce 2026” (42% testing; 3% agentic transactions); Worldpay, “Agentic commerce takes flight” (14% unverified trust; $50 cliff)
- TRM Labs, “Who’s actually paying? Measuring AI agent payments onchain” โ via Decrypt, Sept 13, 2026 (x402 volumes and 0.6%โ7.5% agent share)
- CNBC, “Ant International partners with Visa, Mastercard on developing AI payments,” Sept 10, 2026; Stocktwits, “OpenAI And Visa Partner To Let AI Agents Shop And Pay For You,” Sept 12, 2026; Reuters (India UPI agent registry; rogue-agent coverage; AmazonโPerplexity litigation)
- Elon University / The Washington Post poll, Sept 2026 (27% use chatbots for personal, emotional or social queries)
Disclaimer: This article is for informational purposes only and does not constitute financial, legal, or tax advice. The agentic-commerce landscape is changing weekly; platform features, pricing and protections reflect the sources cited as of September 15, 2026. Read each platform’s terms โ especially purchase-protection and liability clauses โ before linking any payment method, and consult a licensed professional for your specific situation.