Field report · updated 24 September 2026

Nine agents now reach for your card.

Every announced agentic-commerce checkout, in one place: what the experience actually looks like, who it serves, when it shipped, and the markets where it is legally allowed to take your money — read from the network layer underneath it.

Open the player console Where Visa sits
● 7 live or in beta · 1 pivoted · 1 in litigation ● 9 protocols · 6 banks now writing rules ● 6 markets with in-agent checkout
This week

The size of the prize

A rounding error today, a trillion-dollar flow by 2030 — and the second-largest band is not retail.

eMarketer US retail forecast for 2026 · Adobe Digital Insights, Q1 2026 · McKinsey orchestrated-spend estimate for 2030. Definitions vary wildly — Grand View puts the 2026 global market at $7.7B, Mordor at $60.4B, depending on whether you count the assistant, the order, or the whole retail stack behind it. Treat any single number with suspicion; treat the gradient between them as the real signal.

Where the growth comes from

US orchestrated spend, 2025 → 2030

Consumer retail is the headline band, but B2B procurement is the fastest-compounding one — because the buyer actively wants the purchase automated, the approval chain already lives in software, and nobody has to be persuaded that a machine should place the order.

By 2030 the two together are roughly nine tenths of the flow. If you are building the authorisation layer, the commercial credential matters as much as the consumer one.

Illustrative build to McKinsey's ~$1T US orchestrated-spend estimate for 2030, anchored on eMarketer's $20.6B run-rate for 2026 and split by segment. The endpoints are sourced; the path between them is a model, not a forecast.

The landscape

Two questions decide everything: who completes the checkout, and how many people can reach it.

The interesting axis is not model quality. It is whether the agent finishes the purchase itself or hands you back to the merchant — and that single choice is where the money, the liability and the customer relationship land.

Read it this way

Move right and the agent holds the card. Move up and it is in front of hundreds of millions of people. Alexa for Shopping is the only agent that currently sits top-right: it has Amazon's entire signed-in base and it will complete a purchase on third-party sites for you.

ChatGPT sits far left on purpose. It has the reach and it deliberately gave the checkout back.

The network layer · where Visa sits

Visa shipped agentic infrastructure six months before any assistant took a payment.

Intelligent Commerce launched on 30 April 2025. Instant Checkout — the first consumer agentic checkout of any scale — launched on 29 September. The rails were built before there was anything to run on them, which is exactly why the 2026 retreat on in-chat checkout did not slow the payment side down at all.

≈5B
Visa credentials
175M+
Merchant locations
200+
Countries & territories
257.5B
VisaNet transactions, FY25

The Visa agentic stack

Six pieces, each solving a different part of the trust problem

30 Apr 2025

Intelligent Commerce

The umbrella programme

Tokenised credentials, spend controls and authentication for agent-initiated transactions, built to work across whichever assistant the consumer picks.

  • Launch partners: Anthropic, Microsoft, Mistral AI, OpenAI, Perplexity, Samsung, Stripe
  • Extended to Visa's own B2B payments automation platform for corporate bill pay
14 Oct 2025

Trusted Agent Protocol

Agent identity · built with Cloudflare

The piece that lets a merchant tell a legitimate agent from a scraper. Published as an open spec on the Visa Developer Center and GitHub.

  • Agent signatures carry a timestamp, session id, key id and algorithm id
  • Cryptographically bound to a specific merchant domain and operation
  • Secure agent↔merchant communication at every step of a transaction
Live

Tokenized Agent Payments

Credentials scoped to the agent

A token issued to a specific agent, not just a card — authorised for particular transaction types under particular spend limits.

  • The cardholder keeps the underlying card and their dispute rights
  • If an agent goes rogue, revoke the agent token — no card reissue
  • Multi-rail: settles on Visa fiat rails or on the stablecoin programme
8 Apr 2026

Intelligent Commerce Connect

The acceptance on-ramp

One integration with Visa Acceptance Platform and a merchant accepts agent-initiated payments from any of the major protocols — including on non-Visa cards.

  • Covers Trusted Agent Protocol, MPP, ACP and UCP through a single integration
  • Protocol- and token-vault-agnostic; handles PCI for enablers
  • Carries merchant catalogue discoverability onto AI platforms
  • Pilots: Aldar, AWS, Diddo, Highnote, Mesh, Payabli, Sumvin
Expanded 29 Apr 2026

Agentic Ready

The issuer programme — the real moat

Banks test agent-initiated payments on live cards: card enrolment, tokenisation, authentication, authorisation, security mechanisms and operational gaps, before anything scales.

  • 20+ partners live across the UK and Europe
  • 85+ rolling out across Asia Pacific and Latin America
  • 30+ across CEMEA
  • Canada: BMO, CIBC, RBC, Scotiabank, TD
  • Singapore: 13 banks and fintech partners · Vietnam: ACB, MB, Sacombank, Techcombank, VPBank
Proven

Live transactions

Not slideware

Agent-initiated payments are clearing on live cards today, and the agent-enabling layer is already doing both consumer and business purchases.

  • 2 July 2026 — Worldline, ING and Visa execute a live agentic payment in Germany
  • Stablecoin settlement at a $7B annualised run-rate across nine blockchains by April 2026
  • Pilots with Skyfire, Nekuda, PayOS and Ramp running end-to-end consumer and B2B purchases in closed beta
  • InFlow issues network-grade Visa credentials plus a policy engine governing what an agent may spend them on

Why Connect matters more than it sounds

One integration. Every protocol. Including the ones Visa didn't write.

The protocol war has no near-term winner, and merchants cannot integrate nine times. Connect turns that fragmentation from a merchant problem into a Visa product: a neutral hub that speaks Trusted Agent Protocol, MPP, ACP and UCP through one endpoint, works with non-Visa cards, and is agnostic about whose token vault holds the credential.

It is the same strategic move as accepting every card scheme through one terminal — applied to agents. Neutrality is the position with the most leverage when you cannot be sure which protocol wins.

"Intelligent Commerce Connect brings that same, trusted payment acceptance infrastructure into the emerging world of AI-driven commerce, so businesses can let AI agents buy on behalf of consumers, securely and at scale."

Andrew Torre, President of Value-Added Services, Visa

Head to head

Visa, Mastercard and American Express are solving the same problem three different ways

NetworkUmbrellaAgent identityToken model Issuer routeProtocol stanceLive proof
Visa Intelligent Commerce · 30 Apr 2025 Trusted Agent Protocol — open spec, built with Cloudflare; signatures bound to a merchant domain and operation Tokens scoped to the agent; revocable without reissuing the card Agentic Ready — 20+ live UK/Europe, 85+ APAC & LatAm, 30+ CEMEA, Canada's big five Deliberately agnostic — TAP, MPP, ACP and UCP through one integration, non-Visa cards included Live agentic payment in Germany, 2 Jul 2026 (Worldline, ING)
Mastercard Agent Pay · 29 Apr 2025 Verifiable Intent — verification of the reason for the purchase; formal pilots from Feb 2026 Agentic Tokens, built on the tokenisation behind contactless, card-on-file and Payment Passkeys Through issuing banks — Citi and US Bank first, expanding globally through 2026; broadly available via certified processors, free to most merchants Joined Google's AP2 in Sep 2025 First live tokenised agentic transaction, 29 Sep 2025 · Agent Pay for Machines live Jun 2026 with 30+ partners
American Express Agentic Commerce Experiences (ACE) · 14 Apr 2026 Agent verification services inside ACE, plus intent intelligence and cart context Amex-issued credentials handed to registered agents Direct — Amex only has to convince itself Selective Industry-first purchase protection: Amex covers eligible customers for charges arising from AI agent error on registered agents

Where Visa is ahead

Nobody else has standardised the issuer layer

Agent Pay depends on issuing banks switching it on one at a time. Amex only has to persuade itself. Agentic Ready is the only programme testing live-card agent flows simultaneously across the UK, Europe, Asia Pacific, Latin America, CEMEA and Canada — and it is testing the unglamorous parts: enrolment, tokenisation, authentication, authorisation and the operational gaps in between.

Pair that with Connect's protocol neutrality and Visa is the only network a merchant or an agent platform can integrate once and reach globally, whichever protocol and whichever scheme ends up on top.

Where the gap is

The dispute layer does not exist yet

  • As of 2026 no jurisdiction has enacted regulation covering autonomous AI purchasing. Consumer law and card rules assume a human buyer.
  • The chargeback chain breaks in several places when the shopper is an agent — merchants operating under ACP own the liability without the evidence infrastructure to defend a dispute.
  • Amex moved first on the consumer-facing answer: a public commitment to cover agent-error charges on registered agents. That is a trust claim Visa has not matched publicly, and trust is the binding constraint on adoption.
  • Agent identity via TAP is the prerequisite for fixing this — signed intent and bound signatures are exactly the evidence a dispute needs. The rules that turn that evidence into liability allocation are still unwritten.
  • And the largest working agentic loop on earth does not involve a card network at all. See China, next.

New · 22 September 2026 · the bank view

Six banks just wrote the rules they want. Most of it already runs on Visa. The rest is where the fight will be.

Building Trust in Agentic Commerce sets out five principles and sixteen commitments for how AI agents should shop and pay. It is widely reported as a Bank of America and Commonwealth Bank paper; it is actually co-authored by six banks. Below: the paper, every commitment scored against Visa Intelligent Commerce and Trusted Agent Protocol, the gaps, and what the paper is really arguing about.

The paper

Building Trust in Agentic Commerce

Customer-centric principles for a trusted and thriving ecosystem · 13 pages · published 22 September 2026

Bank of AmericaCommonwealth BankCapital One NatWest GroupING GroupASB Bank
Read the paper (PDF) Announcement

The paper’s own autonomy spectrum

Standard online purchase
Product subscriptions
Agent searches, human executes
Human prompts, then reviews final purchase
Human pre-approves; agent executes later
Agent executes after one prompt, no review
Agent identifies the need and buys
TraditionalHuman in the loopAutonomous
  • StatusVoluntary and non-binding. “Does not prescribe conduct, any commercial position or any particular method or timetable.”
  • ScopeAny payment rail: cards, account-to-account and tokenised money. Businesses buying from merchants count as consumers.
  • Not in itNo fees, interchange or routing, and no implementation timetable (KeyBanc).
  • Their diagnosis“The focus to date has been on protocols, mostly led by payment networks and technology platforms”. Some transactions run outside any protocol at all.
  • Named risksOverspending and wrong orders, scams, fraud, agent impersonation, disputes that exclude the agent, and merchants losing the customer relationship.
  • What’s nextA second paper on implementation through “protocols, industry standards and policies”, plus consultation with policymakers. That is where fees and liability will be argued.

Scored

Sixteen commitments against eight Visa capabilities

9
Aligned
3
Partial
2
Tension
2
Gap
No coverage Indirect Partial Direct delivery

Overall 74% of the paper’s commitments are met by Visa capabilities in market today (aligned = 1, partial = ½, tension = 0.4, gap = 0). Safety scores lowest not because Visa’s security is weak, but because the paper defines safety to include liability, delegation management and cross-chain data sharing: the parts nobody has built yet.

This is an analyst’s assessment against published Visa capabilities, not a Visa position. Hover any cell for its score; the reasoning for every commitment is in the stepper above.

Against what Visa has said

Visa’s stated principles cover three of the five

Visa’s three published pillars for trusted agentic commerce (Charles Lobo, SVP & Regional Risk Officer, November 2025) map cleanly onto the paper:

  • Safeguarding consumer control → 2.2, 2.3
  • Enhanced payment security (agent-bound tokens, Payment Passkey) → 2.1, 2.3, 3.1, 3.3
  • Agent verification (“Know Your Agent”, Trusted Agent Protocol) → 1.1, 4.3

VIC’s product principles add trust, transparency and consent. What Visa has not stated a position on are the two principles about market structure: Choice and Interoperability. Visa practises interoperability (Connect, an open TAP); on choice, it is silent, and choice is the principle that names payment networks.

Read between the lines

Positioning
  • The paper credits “payment schemes” with laying building blocks, then states that the work so far has been led by networks and platforms, and that banks bring “decades” of trust.
  • Its Choice principle lists payment networks alongside devices, agents and wallets as platforms that could preference payment methods or charge “gatekeeping fees”.
  • Its scope is rail-agnostic. Account-to-account is named in the definition on page four.
  • The day after publication, consultant Richard Crone told American Banker that banks are disadvantaged because “payment networks — Visa and Mastercard — are not representing their interests.”
  • Four of the six authors (BofA, Capital One, CBA, ASB) sit in exactly the three markets Visa surveyed in its own trust research (US, Australia, New Zealand), where 85% of consumers said control over agent data is important.

The gaps

Where the paper asks for something Visa does not deliver today

Gap · 2.5

Liability that follows the error

The paper asks
Disputes that involve every party, with liability landing where the risk or error entered.
Visa today
Strong evidence (Commerce Signals, CE 3.0) inside an issuer–acquirer process that the agent provider is not party to. No published agent liability framework; Amex moved first on agent-error cover.
What would close it
Agent-specific reason codes and a liability model that brings KYA-registered agent providers into the rules.
Gap · 1.2

How agents rank and sponsor

The paper asks
Transparency over how agents prioritise options, including sponsored ones, and how data is used.
Visa today
Out of scope for VIC: ranking lives inside the agent. Meta now plans to take “a small fee from transactions”.
What would close it
Make ranking and sponsorship disclosure a condition of Know Your Agent onboarding.
Partial · 2.4

Real-time data for both sides

The paper asks
Agent identity, merchant of record, intent and purchase details, to issuers and acquirers, in real time.
Visa today
Instructions are validated in VisaNet and signals are captured, but the per-party data dictionary is not published.
What would close it
Publish which agent fields ride in authorisation, and to whom. The cheapest gap on the page to close.
Partial · 2.1

From safe path to safe rule

The paper asks
Unsafe practices, such as agents entering real card details into websites, discouraged and “where possible prohibited”.
Visa today
VIC provides the tokenised alternative; Grok Bot and Comet still check out with saved cards.
What would close it
A rules position on agent-keyed PANs, phased in behind Connect’s one-integration on-ramp.
Tension · 3.4

Who may use commerce signals

The paper asks
Consent to govern any use of agentic data beyond what each party needs.
Visa today
A rich new dataset of instructions and outcomes, with no published use policy beyond disputes. Banks say they lack equivalent data rights.
What would close it
A published data charter for Commerce Signals, with issuer access to their own cardholders’ records.
Tension · 4.1

Rails, and the fees nobody mentioned

The paper asks
No unreasonable restrictions or requirements on providers, across any rail.
Visa today
Protocol-neutral and card-inclusive via Connect; card-centric by design. Fees are the unwritten half of this principle.
What would close it
Engage on the second paper before it is drafted, not after it is published.

The opportunity

Every gap is one Visa could close with the issuers who wrote this, rather than be positioned against them

The paper’s strongest asks (delegation management, auditable records, real-time data, dispute participation) are all things issuers want to offer their customers, and all things VIC already half-builds. An issuer API for viewing and revoking every agent delegation (2.2), a published agent data dictionary (2.4) and a Commerce Signals data charter (3.4) would turn Visa’s stack into the implementation path for the banks’ second paper. Liability (2.5) is the hard one, and the one where Amex already has a head start.

Stress test

The five principles, applied to this month’s launches

A first-party merchant agent passes almost by construction. General-purpose agents that act as you are where the principles strain, and they strain differently. The network can make the payment safe; it cannot make the agent trustworthy.

PrincipleMeta MuseGrok BotInstinctGopuff Go (Grok-powered)
Transparency Partly

Totals confirmed in chat. But Meta now plans a fee on each transaction, and ranking isn’t disclosed.

 Strains

Signs in with your own sessions; a merchant cannot tell the agent from the person.

 Partly

Acts through your accounts; purchases surface as explicit Link approvals.

 Meets

First-party: the agent is the merchant, so whose interests it serves is obvious.

Safety Meets

Link one-time cards, PayPal and Shop Pay; every total confirmed before money moves.

 Strains

One shared cloud computer: cookies, sessions and credentials shared across every bot. No scoped credential.

 Partly

Link’s scoped one-time cards, yet testers saw an unauthorised email and a successful prompt-injection.

 Meets

Checks out inside Gopuff’s own app and payment flow.

Privacy & Data Partly

Meta says Muse chats don’t feed its ads systems; Zuckerberg was pressed on trust at Connect.

 Partly

Signed-in sessions persist on a cloud machine you don’t control.

 Strains

Perpetual, irrevocable training licence; plain-text email; kept reading Gmail after access was revoked.

 Meets

Gopuff says Grok cannot be trained on its customer data.

Choice Meets

Three wallets (Link, PayPal, Shop Pay) and a named retail set.

 Partly

Any site you can log into, but only with whatever card is saved there.

 Partly

Link where available; otherwise it stops at the Pay button.

 N/A

One merchant by design; shoppers remain free to use others.

Interoperability Meets

Shopify catalog via Agentic Storefronts; PayPal merchants worldwide.

 Strains

No commerce protocol: it drives websites directly.

 Partly

Link’s merchant network; a browser everywhere else.

 — N/A

A closed loop inside one app.
principles-live.mp4
4:13 segment, cut from a 37-minute interview · 23 September 2026 YouTube ↗

The principles, tested live

Zuckerberg on whether you should trust Muse

The day after the banks published, the Wall Street Journal’s Joanna Stern pressed Mark Zuckerberg on Muse privacy and trusting Meta with an agent that holds your accounts. It is the paper’s Transparency and Privacy principles, argued in public by the company with the most to prove on both.

Plays the “Muse privacy and trusting Meta” chapter only.

The players

Thirteen players, twenty-nine recordings of the flow.

Nine consumer agents, the three companies quietly making them possible, and the card rails underneath. Each panel carries the mechanics of the flow, the spec, and two recordings of the shopping experience itself — mostly 30 seconds to two minutes, chosen to show the interface doing the work rather than someone talking about it. Toggle between them above the player; marks a segment cut from a longer official video so it opens on the moment that matters.

Side by side

The whole field on one row each

AgentStatusLaunched Where checkout happensMarkets Payment railCost to the shopper
Meta Muse Live 8 Sep 2026 Agent's own browser, on the merchant's real site US app · PayPal merchants worldwide Link single-use cards · PayPal · Shop Pay Free · $20 · $100 / mo; Meta plans a transaction fee
Alexa for Shopping Live 13 May 2026 In-assistant, incl. third-party sites via Buy for Me US Stored Amazon card & address Free, no Prime needed
Google AI Mode & Gemini Live Jan 2026 (UCP) In-surface button; merchant stays merchant of record US, CA, AU, UK UCP + AP2 mandates; any PSP Free
Alibaba Qwen × Taobao Live 15 Jan 2026 Fully in-conversation, end to end China Alipay Free
OpenAI ChatGPT Pivoted 29 Sep 2025 → 6 Mar 2026 Merchant's own checkout, handed off from chat US-led; discovery global ACP; merchant PSP; PayPal Free tier upward
Microsoft Copilot Live 8 Jan 2026 Inside Copilot chat, no redirect US PayPal, Shopify, Stripe Free
Perplexity Comet In litigation Jul 2025 · free Oct 2025 Agentic browser drives the merchant's real checkout Worldwide — injunction vacated PayPal; merchant checkout Free · Enterprise tier
Grok · SpaceXAI Beta Go 3 Jun · Grok Bot 11 Aug 2026 Go: inside Gopuff's app · Grok Bot: your own logged-in accounts US (Go) · unrestricted (Grok Bot) None native — saved cards in the accounts it signs into SuperGrok Heavy · $120–$200 / mo
Instinct Invite-only Raised $350M, 26 Aug 2026 Merchant sites via Link; otherwise stops at the Pay button US Stripe Link one-time cards, per-purchase approval Not disclosed
Shopify Live 24 Mar 2026 Merchant's own store, surfaced into every AI channel Wherever the channel is live Shop Pay and merchant PSP; speaks ACP and UCP Included; Catalog needs only an API key
Stripe Live 11 Dec 2025 Wherever the agent is — Stripe supplies the token, not the surface Global Shared Payment Tokens, Link agent wallet, Issuing for agents Not disclosed
Anthropic Live MCP Nov 2024 · agents 2 Sep 2026 Nowhere by design — the merchant's own checkout, untouched Global None. No payment protocol, no checkout, no ad layer. Apache-2.0, free
Card & wallet rails Scaling Apr 2025 onward Not a surface — the authorisation layer beneath all of it UK, LatAm, US pilots Agent Pay, Intelligent Commerce, AP2, x402 n/a

The turn nobody predicted

The biggest story of 2026 is not that an agent can buy for you.
It is that in-chat checkout stalled — and the industry regrouped around discover in AI, buy on site.

3×
Walmart measured checkout inside ChatGPT converting roughly three times worse than a click-through to Walmart's own site. OpenAI mothballed Instant Checkout in March 2026 and gave merchants their flow back.
+393%
Traffic from AI sources to US retail sites, year over year in Q1 2026, per Adobe — and it converted 42% better than other traffic in March. The referral is working even where the checkout is not.
1.5%
Agentic commerce's share of US ecommerce in 2026 on eMarketer's forecast. Enormous announcements; a rounding error of actual spend. Both things are true at once.

The protocol war

Five standards are competing to be the language a purchase speaks.

The surfaces get the headlines. The protocols decide who owns the order, who eats the fraud, and whether a merchant integrates once or eight times. The safe posture for a retailer is to support UCP and ACP, with MCP underneath for catalog depth — betting the whole strategy on one protocol is the most common and most expensive readiness mistake being made right now.

Live

MCP

Anthropic · open-sourced Nov 2024

The oldest and most widely adopted of the five, and the only one not built for commerce. It is how an agent asks any compliant server what tools and data it has — which turns out to be exactly how a merchant makes its catalog legible.

  • Donated to the Linux Foundation, Dec 2025 — backed by Block, OpenAI, Google, Microsoft, AWS
  • SDK downloads ~97 million a month
  • Stripe, Shopify and WooCommerce all run MCP servers
  • UCP is designed to speak it; it is the catalog-depth layer beneath the commerce protocols
Live

UCP

Google · open standard · Jan 2026

A universal cart that moves between agent, merchant and payment provider across the whole journey — discovery, buying, and post-purchase support.

  • Built with Shopify, Etsy, Wayfair, Target, Walmart
  • 20+ endorsers: Adyen, Amex, Mastercard, Stripe, Visa, Home Depot, Macy's, Flipkart, Zalando
  • Amazon, Meta, Microsoft and Salesforce also listed as backers
  • Speaks MCP, A2A, AP2 and plain REST
Live

ACP

OpenAI + Stripe · Sep 2025

The protocol outlived the product it was built for. Merchants can adopt it with their existing payment provider, and it is designed to work across agents, not just ChatGPT.

  • Target, Sephora, Nordstrom, Lowe's, Best Buy, Home Depot, Wayfair integrated
  • Survived the March 2026 Instant Checkout retreat
  • Now framed as discovery-first, merchant-owned checkout
Live

AP2

Google + 60 partners · Sep 2025

Not a cart — a signature. The shopper's authorisation becomes a cryptographically signed mandate, so a merchant can prove exactly what a human approved and an agent cannot exceed it.

  • Visa and Mastercard both joined in Sep 2025
  • Coinbase among the 60+ partners
  • Pairs with UCP rather than competing with it
Scaling

Agent tokens

Mastercard, Visa, PayPal, Coinbase

The credential layer. A verified agent gets a token scoped to one purchase and one amount, so the agent never touches a real card number.

  • Mastercard Agent Pay — Agentic Tokens, first live transaction 29 Sep 2025
  • Visa Intelligent Commerce, Trusted Agent Protocol, Tokenized Agent Payments
  • PayPal Agent Ready — live in ChatGPT and Perplexity
  • Coinbase x402 revives HTTP 402 for stablecoin settlement

What a handshake actually looks like

One cart, one signed mandate, one scoped token.

Strip away the branding and every live implementation does the same four things: hydrate a cart the agent did not invent, get a human to sign the total, mint a credential that cannot be reused, and leave the merchant as merchant of record.

The differences are all in step four. Amazon and Alibaba keep it. Google and OpenAI hand it back. Meta routes around it with a virtual card and a browser.

China vs the rest of the world

Two ecosystems, three orders of magnitude apart.

The Western map has nine competing protocols and one market with in-agent checkout at scale. China has one integrated stack and production volume. In the week of 5–11 February 2026, Alipay AI Pay processed over 120 million agentic transactions. The busiest Western protocol was doing roughly half a million a week.

China

Vertical · integrated · closed
  • One company owns the whole chain. Alibaba has Qwen, Taobao, Alipay and Cainiao. ByteDance has Doubao, Douyin Mall and Douyin Pay. No protocol translation, no multi-party coordination, no handoffs.
  • Alipay AI Pay passed 100 million users on 23 February 2026 — the first AI-native payment product to get there.
  • On 17 August 2026 Alipay unveiled a full-stack agentic commerce foundation and the AHA (Agent Hub Access) multi-agent interconnection protocol suite at its first AI ecosystem conference in Hangzhou.
  • UnionPay's APOP (Agentic Payment Open Protocol) launched in Hong Kong on 2 April 2026: native China access to 9.6 billion cards across 183 countries, regulator-first design aligned to PBOC, intent management as a protocol primitive, cross-participant single sign-on. Caveat — no developer tooling and no major Chinese tech partners announced; only Evonet and Hoppa demonstrated.
  • Doubao has 226M monthly actives and completes multi-platform shopping in under 30 seconds. Qwen took nearly 200 million "one-sentence orders" over the 2026 Spring Festival.
  • Platforms spent $647M subsidising AI shopping adoption — Alibaba $431M, Tencent $144M, Baidu $72M. Demand was bought, then it stuck.

Rest of world

Horizontal · fragmented · portable
  • Nine protocols with divergent incentives — TAP, UCP, ACP, AP2, MPP, x402 and Agent Pay in the West, APOP and AHA on the other side of the wall.
  • An agent must navigate separate identity, discovery and settlement layers, with a different answer on every surface.
  • One market with in-agent checkout at genuine scale (the US), three more for Google's checkout button, and Europe switched off on data-protection and payments-regulation grounds.
  • Ownership is split by design: the model, the marketplace, the wallet and the logistics are four different companies with four different P&Ls — which is exactly why a protocol was needed at all.
  • But fragmentation is not only a cost. It is why the Western stack is protocol-agnostic, portable across wallets and issuers, and legally survivable in 200+ jurisdictions. China's loop is faster because it is closed. That is a trade, not a verdict.

What this means if you work at a network

In China's agentic loop, there is no card network.

Payment resolves inside Alipay or Tenpay. The model, the marketplace and the wallet sit on the same balance sheet, so there is no interchange, no scheme, and no acceptance integration to sell. China's payments market is roughly $43.65 trillion and Asia Pacific is projected to reach $33 trillion by 2031 — so a stack that speaks only TAP, UCP and ACP is, on this reading, built for part of the addressable market.

Three things worth watching:

  • Hong Kong and Singapore — the bridge markets where both ecosystems coexist, and where Visa's Agentic Ready is already live with 13 Singapore partners. If interoperability gets proven anywhere, it gets proven here first.
  • Ant International — Alipay's global arm participates in Google's AP2 while building China's domestic rails, which makes it the closest thing to a working bridge between the two standards.
  • EMVCo — jointly owned by the Western networks and UnionPay, it has stood up a Digital Identity and Payments Task Force. It may be the only institution structurally capable of reconciling incompatible agentic standards.

Merchant readiness

The bottleneck is not the agent. It is the catalogue.

Agents are ready to buy. Most merchants are not ready to be bought from. Early-2026 research found 40% of ecommerce businesses still standardising product pages for agentic AI and 33% that had not started at all — leaving roughly a quarter past the starting line, while agentic AI traffic grew 7,851% year on year. The tooling is no longer the blocker; Shopify, Stripe and Anthropic have each removed a different piece of the work. The backlog is.

Derived from early-2026 survey data: 40% still standardising product pages, 33% not started. The remainder are past initial preparation — not finished.

Illustrative model, not survey data. Only the entry stage and the ~27% "past initial prep" figure are sourced; the intermediate stages are an analytical frame for where merchants actually fall out. Each stage is a hard technical gate, not a maturity score.

The fix list

Six things a merchant has to get right, in this order

1 · Structured product data

ACP is push: you submit structured feeds to OpenAI, which surfaces them in ChatGPT. UCP is pull: you publish a machine-readable manifest and agents call your commerce APIs directly. MCP goes deeper still — your catalog, cart and policies exposed as tools an agent can call rather than text it has to parse. Either way: a product without machine-readable attributes will not be recommended, however well the description reads.

2 · Stop blanket-blocking

User-agent blocking misses most AI agents anyway; behavioural classifiers catch them. The working 2026 playbook is purpose-based control — identify the agent, then decide what it is allowed to do. Visa's Trusted Agent Protocol exists precisely to make that decision possible.

3 · An agent-callable checkout

A human-shaped form is not an interface. An agent needs a cart it can hydrate, a price it can trust, a total it can present for approval, and an authorisation it can carry. Everything upstream is wasted if this step requires a mouse.

4 · A delivery promise that is true

Most retailers are optimising to appear in AI discovery. Far fewer have asked whether fulfilment can honour a date an agent just quoted on their behalf. That distance between digital presence and physical capability is the real readiness gap — and an agent that quotes wrong generates a dispute, not a complaint.

5 · Post-purchase in the same channel

Order status, returns, after-sales. Alibaba's Qwen handles all three inside the original conversation via a skills library; that is the bar being set. A checkout that completes in-agent and then dumps the customer into a call queue is not a finished experience.

6 · Dispute evidence

Capture agent identity, the signed intent and the cart context at the moment of authorisation. Merchants under ACP carry the liability today without the evidence to defend it, and no regulator has written the rules yet. The merchants who log this now will be the ones who win disputes later.

Who is actually working on this

Three companies have made merchant readiness their product rather than their excuse

Most of this page is about companies competing for the shopper. The readiness problem is being solved by a different set, and largely by removing work rather than adding standards:

  • Shopify removed the decision entirely. Agentic Storefronts went on by default for ~5.6 million stores on 24 March 2026 and the merchant never touches ACP or UCP — they toggle channels in Admin. Shopify also supplies the single best piece of evidence that any of this matters: AI searches powered by its structured Catalog convert at 2× the rate of those using scraped data. Same products, same agents, different data quality.
  • Stripe removed the integration. The Agentic Commerce Suite gives a merchant a hosted ACP endpoint, checkout with tax and shipping, and agent-payment processing with fraud controls — and lets platforms like Wix, WooCommerce, BigCommerce, Squarespace and commercetools pass it through to their own merchants.
  • Anthropic removed the ceiling. MCP, open-sourced in November 2024 and donated to the Linux Foundation, is the layer that makes a catalog callable rather than merely readable — Stripe, Shopify and WooCommerce all run MCP servers, and UCP is built to speak it. Then in September 2026 Anthropic published an Apache-2.0 blueprint for commerce agents a merchant runs on its own property, with Visa, Mastercard and Accenture distributing it to their merchant networks. It ships no payment protocol, no checkout and no ad layer, so nothing routes through Anthropic.

The pattern is worth noting: nobody fixed readiness by publishing another specification. They fixed it by making the specification somebody else's problem — a platform default, a hosted endpoint, or an open blueprint. Any retailer still treating readiness as a standards decision is solving the wrong problem.

Two postures emerged, and the courts just moved one

The agent-access question
  • A few large players closed the door. Amazon blocked third-party agents — plausibly to protect a $56B advertising business that depends on shoppers browsing Amazon's own site.
  • Most others opened partner integrations instead, on ACP or UCP, trading control for reach.
  • On 4 August 2026 the Ninth Circuit vacated Amazon's injunction against Perplexity, finding that when a user directs an assistant to complete a task, it is the user — not Perplexity — accessing Amazon's computers. Comet is operating on Amazon again.
  • The lawsuit continues in San Francisco and Amazon can seek rehearing, so whether terms of service can lawfully exclude agents is still open. But the default just shifted toward access.

Meanwhile the demand side is not waiting

Why readiness is urgent, not optional
  • 45% of consumers already use AI for at least part of the buying journey (IBM Institute for Business Value, 2026).
  • AI-driven orders across Shopify stores grew 15× between January 2025 and January 2026.
  • Agentic AI traffic grew 7,851% year on year, and over 95% of it landed in three sectors: retail and ecommerce, streaming and media, travel and hospitality.
  • AI crawlers now account for 20.3% of verified bot traffic, AI-search bots another 6.5% — and bots overall now pass half of all web requests.
  • Gartner expects 40% of enterprise applications to embed AI agents by 2026.

B2B agentic commerce

The bigger prize is procurement — and it is further along than retail.

Consumer agentic checkout is a 1.5%-of-ecommerce story with a trust problem. B2B procurement is a multi-trillion-dollar flow where the buyer actively wants the purchase automated, the approval chain already lives in software, and nobody needs persuading that a machine should place the order.

Already happening

This is not a 2028 story

Nearly two in five B2B buyers are already using agentic AI somewhere in purchasing — evaluating products, configuring orders, reviewing contracts, benchmarking prices. Forrester expects about one in five B2B sellers to face agent-led quote negotiations by the end of 2026, and a third of B2B payment workflows to involve AI agents by the same point.

Consumer agentic commerce is where the brand gets won. B2B agentic commerce is where the volume is.

How agentic procurement actually runs

Policy at the top, humans on the approval, the machine in the middle

  1. The agent reads the policyApproved supplier list, budget constraints, vendor preferences, historical order data. It operates inside those boundaries rather than inventing its own.
  2. It finds and comparesStructured supplier data, quotes requested through APIs, prices benchmarked, contracts checked against terms already agreed.
  3. It prepares the purchase orderHumans set the policy and approve the result. Approval chains and contracts stay fully in force — the mechanics of buying are automated, not the authority to buy.
  4. Next: agent-to-agent negotiationBuyer and seller systems interacting directly, settling within predefined rules, optimising discount and transaction speed with minimal human involvement. This is where the efficiency story turns into an autonomy story.

Small businesses

A capability jump, not an efficiency gain
  • No category manager, no negotiated rates, no PO system, no procurement team. An agent that sources, compares and buys inside a spend cap is a department they never had.
  • They have no software to govern the agent, which means the credential has to do the governing: a scoped card with a hard limit is the policy engine.
  • This is the segment where agent-scoped tokens and single-use virtual cards are not a convenience but the entire control framework.
  • It is also the segment most exposed if the dispute rules stay unwritten — a small business has no leverage in a chargeback fight.

Corporates

Throughput inside an existing chain
  • Nothing here is greenfield. The agent slots into an approval chain that already exists in procurement software, so its value is throughput, not autonomy.
  • What they need is auditability: which agent, acting under whose authority, against which policy version, with what evidence retained.
  • That maps almost exactly onto what Trusted Agent Protocol already emits — agent identity, bound signatures, session ids — which makes the commercial credential the natural place to enforce it.
  • Level III data and virtual-card controls give the finance team the reconciliation trail that a consumer flow never needs.

Where Visa sits in B2B

Connect was built for procurement too, not just consumer shopping

  • Intelligent Commerce Connect extends agentic commerce into B2B procurement and bill pay — one integration, four protocols, non-Visa cards included. The same neutral-hub position, applied to a larger flow.
  • Visa has applied Intelligent Commerce to its own B2B payments automation platform, streamlining corporate bill pay while letting customers earn cashback on card payments — a reference implementation, not just a spec.
  • Early Intelligent Commerce pilots with Skyfire, Nekuda, PayOS and Ramp are already executing end-to-end consumer and B2B purchases in closed beta.
  • The virtual card is the natural agent credential: scoped, single-use, spend-capped, with Level III data attached. Javelin identifies five forces converging on virtual-card B2B adoption in 2026 — Level III data enforcement, embedded payments, customised pricing, agentic AI and macroeconomic uncertainty.
  • The marketplace side is moving independently: Alibaba has launched an AI sourcing agent aimed at global trade, which means agent-led sourcing arrives whether or not Western procurement suites are ready for it.
  • And the first general-purpose agent vendor has shipped a procurement agent: SpaceXAI’s Haggle Bot (4 Sep 2026) reads vendor spend, contracts and Ramp data, benchmarks prices and drafts negotiation counters — but will not spend, sign or send without a human. SpaceXAI reports ~$100K identified in a week (company-reported). The agent does the finding; the commercial card still does the paying — which is precisely where Visa sits.

The supplier-side trap

72% of suppliers say their sales processes are mostly or highly automated. Only 47% of buyers agree. When a buyer's agent evaluates suppliers it favours clean, structured, machine-readable data — so suppliers still quoting from PDFs and manual processes do not get outcompeted. They become invisible. There is no bad review, no lost bid, no feedback signal at all: they simply never enter the comparison set. It is the same failure mode as the consumer catalogue problem, with a longer sales cycle and a bigger invoice attached.

Chronology

Seventeen months from the first agent token to Meta putting a browser in your pocket.

Apr – Sep 2025

The rails go first

Mastercard Agent Pay (29 Apr) and Visa Intelligent Commerce (30 Apr) land a day apart. OpenAI and Stripe ship Instant Checkout and ACP on 29 Sep; Visa's Trusted Agent Protocol follows on 14 Oct. Mastercard runs the first live tokenised agentic transaction.

Jan – Mar 2026

Standards, then the retreat

Sundar Pichai announces UCP at NRF in January. Microsoft ships Copilot Checkout on the 8th; Alibaba wires Qwen to all of Taobao on the 15th. Klarna backs UCP in February. In March, OpenAI mothballs in-chat checkout.

May – Sep 2026

Scale and litigation

Amazon merges Rufus and Alexa+ into Alexa for Shopping on 13 May, then sells the engine to rival retailers on the 27th. UCP checkout reaches Canada, Australia and the UK. Amazon wins an injunction against Comet in March, then loses it on appeal on 4 August. Grok Bot enters beta (11 Aug) and Instinct raises $350M (26 Aug). Meta launches Muse on 8 September and adds PayPal and Shop Pay within a fortnight — and on 22 September six banks publish their own principles.

Geography

Almost all of it is American. The largest working loop is Chinese. Europe is switched off.

Autonomous checkout is not generally available in Europe: data-protection, payments and AI regulation make an agent holding a credential hard to clear. Meanwhile Deloitte expects Asia Pacific to drive roughly two thirds of the world's new retail sales over the next five years, with nearly three quarters of APAC consumers already using AI to discover and compare products.

MarketMuseAlexaGoogle QwenChatGPTCopilotCometGrokInstinct
United States ●●● ○◐●◐●●
Canada ○○● ○◐○◐◐○
United Kingdom ○○● ○◐○◐◐○
Australia ○○● ○◐○◐◐○
European Union ○○○ ○◐○◐◐○
China ○○○ ●○○○○○
Rest of world ○○○ ○◐○◐◐○
●Checkout completes in-agent ◐Discovery only, or a browser agent using your own saved card ○Not available

Card-network programmes run on a different map: Visa's Agentic Ready launched with UK banks and issuers before expanding to Latin America — the rails are ahead of the surfaces outside the US.

Reality check

The referral is real. The in-chat checkout, mostly, is not.

Two charts hold the whole argument. AI is now a serious source of retail traffic that converts better than average. And the closer the checkout sits to the chat window, the worse it performs.

AI-referred traffic to US retail sites

Indexed, Q1 2025 = 12. Adobe Digital Insights reported +393% year over year in Q1 2026, and +1,324% since it began tracking the channel in October 2024.

Relative conversion by where the checkout lives

Indexed to ordinary web traffic = 100. Composite of Walmart's ~3× finding on ChatGPT handoffs, Adobe's +42% on AI-referred traffic, and Microsoft's +53% on Copilot-included journeys. Directional, not a single controlled study.

The trust gap is the real constraint

Every live implementation stops and asks a human to confirm a total. Muse does it in chat. UCP does it with a signed AP2 mandate. Nobody has shipped a genuinely unattended purchase of any size, and the ones that come closest — Alexa's price-drop rules — work because the human set the ceiling in advance.

Attribution breaks before checkout does

When an agent browses, fills and buys, the merchant sees a session that looks like nothing in their analytics. That is why UCP and ACP both carry agent identity, and why Visa built a Trusted Agent Protocol at all: the ecosystem needs to tell a good agent from a scraper.

Amazon v. Perplexity just moved the boundary

A district court blocked Comet from Amazon in March 2026. On 4 August the Ninth Circuit vacated the injunction, holding that when a user directs an assistant to complete a task, it is the user — not Perplexity — accessing Amazon's computers, which undercut Amazon's Computer Fraud and Abuse Act claims. The suit continues and Amazon may seek rehearing, so whether terms of service can exclude agents is unresolved. But the browser-driven model just got a lot more viable.

Further viewing

The long-form versions, for when two minutes isn't enough.

Full sessions, architecture walk-throughs and network-side material — run times shown. These are the ones that did not belong in the console.