ChatGPT 
Perplexity 
Claude 
Google AI Mode 
Grok Agentic Payments in 2026: Anthropic, Visa and Mastercard
Summary
Key takeaways
- Agentic payments are payment flows initiated or prepared by AI agents acting under delegated user or business authority.
- The defining requirement is verifiable authorization: merchants and payment providers need proof of what the user allowed the agent to buy, how much it may spend, and under which conditions.
- AI agents should not handle raw card credentials; tokenized payment credentials and scoped payment instruments are becoming the preferred model.
- Google’s AP2 focuses on the trust and authorization layer by using signed mandates and receipts that document user intent and the final transaction.
- Visa Trusted Agent and Mastercard Agent Pay address agent identity, tokenization, authentication, and transaction controls on existing payment networks.
- Agentic commerce protocols and agentic payment protocols solve different problems: commerce protocols coordinate discovery and checkout, while payment protocols prove authorization and move money securely.
- Human-not-present payments introduce additional risk because an agent may transact asynchronously after the user has left the session.
- Spending limits, merchant restrictions, product constraints, validity periods, approval thresholds, and revocation should be technically enforceable rather than stored only in prompts.
- Merchants still need deterministic payment validation, fraud controls, dispute evidence, receipts, and audit trails even when AI handles much of the buying journey.
- Agentic payments should be treated as an extension of established payment infrastructure with stronger identity and authorization controls, not as permission for an LLM to make unrestricted financial decisions.
When this applies
This applies when an ecommerce business wants AI agents to move beyond product discovery and recommendations into checkout or payment execution. It is especially relevant for merchants preparing for agentic commerce through ChatGPT, Gemini, Copilot, merchant agents, procurement agents, or similar systems. The model is also useful for B2B scenarios where an agent may reorder approved products, pay invoices, or transact within predefined budgets. It becomes particularly important whenever the buyer will not necessarily be present at the exact moment a transaction is executed.
When this does not apply
This does not apply when AI is only assisting with research, product comparison, support, or cart preparation and a human still completes the conventional checkout personally. It is also inappropriate when the organization cannot reliably authenticate users and agents, enforce payment limits, maintain audit records, or revoke delegated authority. Businesses should not introduce autonomous payment execution simply because an LLM can technically call a payment API. If the underlying authorization, fraud, credential, and dispute controls are not ready, human-controlled checkout remains the safer model.
Checklist
- Define exactly which payment actions an AI agent is allowed to perform.
- Separate product-selection authority from payment authority.
- Require explicit user or organizational consent before delegating spend.
- Define maximum transaction and cumulative spending limits.
- Restrict delegated authority by merchant, product, category, or supplier where appropriate.
- Add expiration dates or validity windows to agent payment permissions.
- Use tokenized credentials instead of exposing raw card information to agents.
- Verify agent identity before accepting agent-initiated transactions.
- Evaluate support for AP2, Visa Trusted Agent, Mastercard Agent Pay, and relevant commerce protocols.
- Enforce approval thresholds outside the language model.
- Maintain cryptographically verifiable or otherwise reliable evidence of user authorization.
- Log the original instruction, approved cart, payment authorization, and executed transaction.
- Define immediate revocation and exception-handling procedures.
- Test fraud, duplicate-payment, changed-price, unavailable-item, and dispute scenarios before launch.
- Expand toward human-not-present transactions only after supervised agent payments operate reliably.
Common pitfalls
- Treating a conversational instruction as sufficient payment authorization.
- Allowing an AI agent to access or retain raw card credentials.
- Storing spending restrictions only inside prompts instead of enforcing them technically.
- Confusing commerce protocols such as ACP or UCP with the authorization layer required to secure payments.
- Allowing an agent to change the final cart after the user has approved different purchase conditions.
- Failing to distinguish human-present and human-not-present transaction risk.
- Connecting agents directly to payment APIs without identity, fraud, and permission controls.
- Neglecting audit trails that prove what the user authorized and what the agent actually purchased.
- Designing around a single payment or agent ecosystem before the standards landscape has stabilized.
- Increasing payment autonomy before the organization has reliable revocation, exception handling, and dispute procedures.
Agentic payments are payments that an AI agent starts on behalf of a person, inside limits that the person sets, on rails that verify the agent and replace the card number with a scoped token. On September 2, 2026, Anthropic released Claude Commerce Agents, an open-source blueprint for shopping and merchant agents, and named Visa and Mastercard as ecosystem partners. The blueprint does not process payments. It leaves checkout to the merchant and payment credentials to the card networks. For merchants, this splits agentic commerce into two separate jobs: run the agent, and accept agent-initiated payments. Elogic Commerce, a member of the Claude Partner Network, explains both jobs below and what they mean for the next 24 months.
Key facts
- Date: Anthropic published “Building commerce agents with Claude” on September 2, 2026.
- What shipped: an Apache-2.0 reference implementation of a shopping agent and a merchant agent, four vertical demos (retail, travel, telecom, ticketing), a Claude Code plugin, live demos and an engineering guide. Elogic Commerce covers the blueprint itself in the Claude Commerce Agents blueprint article.
- Runtimes: the Messages API, the Claude Agent SDK and Claude Managed Agents (beta). The code deploys on the Claude API, Amazon Bedrock, Microsoft Foundry and Google Cloud Vertex AI.
- Named partners: Visa, Mastercard and Accenture as solutions and ecosystem partners. Shopify, Priceline, Intuit, Klaviyo, Wix, Zomato, Fetch and Square gave customer statements.
- Vendor results: Anthropic reports carts up to 35% larger and shoppers 60% more likely to complete a purchase on agents built with Claude. These are Anthropic’s figures, not an independent benchmark.
- What did not ship: no payment protocol, no checkout, no wallet, no advertising layer. Payment stays with the merchant’s existing checkout or with an agentic payments provider.
- The rails: Visa Intelligent Commerce (Trusted Agent Protocol, scoped tokens, Intelligent Commerce Connect) and Mastercard Agent Pay (Agentic Tokens, Verifiable Intent) are the two card-network programs for agentic AI payments.
- Demand: Adobe Analytics reports that AI-referred traffic to US retail sites grew 393% year on year in Q1 2026 and converted 60% better than non-AI traffic by July 2026.
What Anthropic released on September 2, 2026
Anthropic released a blueprint for commerce agents that run on Claude. The blueprint is a public repository with working code, not a consumer product. It contains two agents.
The shopping agent lives inside the merchant’s app or website. It searches the catalog, assembles multi-item orders, remembers customer preferences, shows products and comparisons inside the conversation, builds the cart and hands the cart to checkout. It also answers service questions such as order status, returns and refund policy. Its guardrails constrain prices and products to real catalog data and block manipulative upsell patterns.
The merchant agent supports the people who run the store. It answers questions about sales performance, tracks inventory and flags problems, recommends pricing and promotions from the store’s own history, and drafts marketing campaigns. When the agent proposes a change, a person approves it before anything goes live.
Anthropic released the blueprint with reference implementations for retail, travel, telecom and ticketing, a Claude Code plugin that scaffolds an agent against a merchant’s own backend, and an engineering guide on architecture, latency, cost, memory, safety and evaluation. Shopify said it is building a reference storefront on the blueprint that connects to a merchant’s store through Catalog, UCP and Shop Sign-in. Priceline rebuilt its assistant Penny on Claude. Wix and Zomato said their engineers had the agents running in minutes.
Anthropic named Visa, Mastercard and Accenture as partners that will help clients and merchant communities use the blueprints. That is the sentence that Finextra, PYMNTS and Reuters led with. It is also the sentence most readers misread. Visa and Mastercard are not inside the blueprint. They sit below it.
What the blueprint does not do
The engineering guide states the rule plainly: no model tool call moves money or changes the business. On the shopping side, the checkout tool renders the cart with a button to place the order, and the backend interface the agent calls has no charge method at all. On the merchant side, every write produces a staged change with a server-generated ID, and the change applies only after a person approves it through a real surface.
So the blueprint contains:
- No payment protocol. It does not implement the Agentic Commerce Protocol (ACP), the Universal Commerce Protocol (UCP) or the Agent Payments Protocol (AP2).
- No checkout. The agent hands the cart to the merchant’s own checkout, or to an agentic payments provider the merchant chooses.
- No wallet and no stored credentials. The agent never holds a card number.
- No advertising or ranking layer. The agent’s search tool returns the merchant’s own ranking. The model decides what to show, not what to sell.
Anthropic’s own FAQ on the commerce page answers the question “Is Anthropic moving into commerce?” with “No.” The company supplies the reasoning layer. Everyone else keeps their job. The merchant keeps checkout. The processor keeps acceptance. The card network keeps the token. This design choice is the reason the announcement matters for payments. It makes the card networks the default answer to “who pays?”
What are agentic payments?
Agentic payments definition: an agentic payment is a payment that software starts on behalf of a person, with that person’s prior consent, within limits the person set in advance, and with proof of both the agent’s identity and the person’s intent attached to the transaction. The card networks call the AI program an agent. The person is the principal. The merchant is the counterparty. The network and the issuer are the verifiers.
An agentic payment differs from a stored card on file in three ways. The card credential is scoped to one agent and one policy, not to a device or a merchant. The request carries a cryptographic signature that names the agent. And the limit lives at the network, so it holds even if the agent is compromised.
How agentic payments work, in four steps
- Consent and policy. The person enrolls a card with their bank or wallet and grants an agent a policy: a spend ceiling, allowed merchants or categories, an expiry date. The policy is the contract.
- Credential tokenization. The card network mints a token bound to that agent and that policy. This is agentic tokenization for secure AI commerce: the agent and the merchant receive the token, never the card number. Visa issues these tokens through Visa Intelligent Commerce. Mastercard issues Agentic Tokens through its Digital Enablement Service.
- Agent identity and intent. The agent signs each request so the merchant can tell a verified agent from a scraper. Visa’s Trusted Agent Protocol uses HTTP message signatures for this. Mastercard adds Verifiable Intent, a signed record of what the person asked for, that each later transaction references. AI agent verification in agentic commerce is this step.
- Authorization and settlement. The merchant’s checkout sends the token to its processor. The network checks the token’s scope and the policy limits before it approves. Where regulation requires strong customer authentication, a payment passkey confirms the person. Settlement runs on the same card rails the merchant already uses.
Example. A shopper tells a merchant’s agent: “Get me a tent, a sleeping bag and a stove for a weekend with two kids, under 400 euros.” The agent searches the catalog, builds a three-item cart and shows it. The shopper taps “place order”. The checkout receives a Visa token scoped to this agent, checks the agent’s signature, sends the token to the processor, and the issuer approves within the 400-euro limit. The agent never saw a card number. The merchant’s fraud system saw a verified agent, not a bot.
| Step | Human checkout today | Agentic payment |
|---|---|---|
| Who starts the payment | The person, on the merchant’s page | The agent, with the person’s prior consent |
| Who holds the credential | The merchant or its PSP (card on file, network token) | The network, as a token scoped to one agent and one policy |
| How identity is proven | Session, device, 3-D Secure | Agent signature plus passkey or SCA when required |
| Where the limit lives | Issuer credit limit only | Policy at the network, plus issuer limit, plus the merchant’s own caps |
| What the merchant’s fraud system sees | A human session or an unknown bot | A verified agent with a directory entry and a signed request |
| Who approves the order | The person, by clicking | The person, by tapping approve in the agent’s UI, or in advance through the policy |
Where Visa and Mastercard sit
Agentic commerce payments do not need a new network. That is the point both networks make in every release. Visa and Mastercard do not build the agent and do not own the surface where the shopper talks. They own the leg between the issuer and the merchant’s processor, and they are extending that leg with agent-aware tokens, agent identity and intent records. The agentic payments infrastructure is the same infrastructure, with three new fields.
Visa Intelligent Commerce
Visa Intelligent Commerce is Visa’s portfolio for AI-driven payments. It has three parts that matter to a merchant.
- Trusted Agent Protocol (TAP). Introduced in October 2025 with more than ten partners, TAP is an open framework built on existing web infrastructure. Agents sign their requests with HTTP message signatures, and merchants use the signature to separate legitimate purchasing agents from malicious bots. Cloudflare and Akamai integrate TAP at the edge, so the same systems that blocked bots now admit verified agents. Visa’s Agent Directory gives merchants a consistent record of who the agent is.
- Scoped tokens and spend controls. Visa issues tokenized credentials to agents with consumer-set limits. A travel agent’s token can be valid only for one airline and one trip window.
- Intelligent Commerce Connect. Launched April 8, 2026 through the Visa Acceptance Platform, this is the merchant on-ramp. One integration gives payment initiation, tokenization, spend controls and authentication. Visa states it accepts agent-initiated payments over TAP, the Machine Payments Protocol (MPP), ACP and UCP, works with major token vaults, and can make merchant catalogs discoverable on AI platforms.
Visa completed live agentic transactions with about 30 European issuers in July 2026, at merchants including lastminute.com and Frasers, with every transaction authenticated by a Visa Payment Passkey to meet Europe’s strong customer authentication rules. On June 10, 2026, Visa and OpenAI announced tokenized Visa credentials for agent-initiated checkout inside ChatGPT. Visa agentic payments are live, not planned.
Mastercard Agent Pay
Mastercard Agent Pay is Mastercard’s program for agent-initiated transactions. It launched on April 29, 2025, piloted with Citi and U.S. Bank in September 2025 and reached all US cardholders in November 2025. It has three parts.
- Agentic Tokens. An extension of the Mastercard Digital Enablement Service. Each token is bound to one agent, one merchant scope and one consent policy. The merchant receives the token, the acquirer routes it, and Mastercard de-tokenizes it. The card number never reaches the agent or the merchant.
- Verifiable Intent. Formal pilots began in February 2026. The agent records the person’s request as a signed Intent Artifact on Mastercard infrastructure. Each later transaction carries a reference to that artifact, so the network can check what the agent bought against what the person asked for. This is a different trust model from Visa’s. Visa verifies the agent. Mastercard verifies the intent as well.
- Agent Pay Acceptance Framework. Mastercard’s merchant and processor program. In Latin America it launched with Checkout.com, Getnet, Yuno and others. Mastercard joined Google’s AP2 and UCP coalitions, and PayPal tokens work inside the Agent Pay framework.
For agentic AI payments fraud prevention, both networks make the same claim: the token, the signature and the policy give the issuer more signal on an agent transaction than on a normal card-not-present sale. Merchants should expect issuers to price that signal into authorization rates over time.
| Dimension | Visa Intelligent Commerce | Mastercard Agent Pay |
|---|---|---|
| Announced | April 2025; TAP in October 2025; Intelligent Commerce Connect April 8, 2026 | April 29, 2025; all US cardholders November 2025 |
| Credential model | Network tokens scoped to a specific agent, with consumer-set limits | Agentic Tokens from MDES, bound to agent, merchant scope and consent policy |
| Agent verification | Trusted Agent Protocol: HTTP message signatures, Agent Directory, edge partners Cloudflare and Akamai | Agent registration on the network; agent identity bound into the token |
| Intent model | Consumer authorization and policy checked at authorization | Verifiable Intent: signed Intent Artifact referenced by each transaction (pilots from February 2026) |
| Merchant on-ramp | Intelligent Commerce Connect via the Visa Acceptance Platform; supports TAP, MPP, ACP, UCP | Agent Pay Acceptance Framework via acquirers and processors |
| Live status (September 2026) | Live transactions with about 30 European issuers; ChatGPT checkout with OpenAI | Live for US cardholders; regional rollouts in Latin America and elsewhere |
| What this means for your checkout | Accept scoped Visa tokens; read TAP signatures at the edge | Accept Agentic Tokens through your processor; log the intent reference |
What this changes in a merchant’s checkout
Three things change. First, the processor must accept network tokens that carry an agent flag, and the merchant’s risk rules must read that flag. Second, the edge must read agent signatures, so the web application firewall becomes part of the sales channel. Third, spending limits become network-enforced, but the merchant keeps its own caps. The blueprint enforces caps on the resulting cart state, not on the request, because an agent will retry and rephrase in ways a human never did. Merchants should do the same.
The agentic commerce stack, and where each piece sits
The agentic payments infrastructure is a stack, not a race. Each protocol or product sits on one layer. Most confusion about the Anthropic announcement comes from reading the blueprint as a competitor to ACP or UCP. It is not. It is the layer above them.
| Layer | Standard or product | Owner | What it does | What a merchant does |
|---|---|---|---|---|
| Agent runtime | Claude Commerce Agents | Anthropic (open source) | Runs the shopping and merchant agents on the merchant’s own surface; calls the merchant’s systems as tools | Fork, connect to catalog, cart, orders and policies; choose a checkout |
| Communication | MCP (Anthropic), A2A (Google) | Open standards | Connect models to tools and agents to agents | Expose commerce tools over MCP if third-party agents will call them |
| Commerce journey | UCP (Google, Shopify and 20+ partners, January 2026); ACP (OpenAI and Stripe, September 2025) | Coalitions | Standardize how an outside agent discovers, carts and checks out; UCP covers the full journey, ACP covers checkout | Enable on the platform to reach Google AI Mode, Gemini and ChatGPT shoppers |
| Agent trust | Visa Trusted Agent Protocol and Agent Directory | Visa, with Cloudflare and Akamai | Prove which agent is calling and that it has commerce intent | Turn on at the edge; whitelist verified agents |
| Payment authorization | Visa Intelligent Commerce tokens; Mastercard Agentic Tokens and Verifiable Intent; AP2 (Google) | Card networks; Google | Bind a credential and a policy to an agent; prove the person’s mandate | Ask the processor to enable agent tokens; log intent references |
| Settlement | Card rails; MPP (Stripe and Tempo); x402 (Coinbase); stablecoins | Networks, processors, chains | Move the money, including machine-to-machine micropayments | No change for card volume; watch MPP and x402 for API and content sales |
Claude Commerce Agents sits at the top of this stack, on the merchant’s own surface. It calls the merchant’s search, cart and order systems as tools. It hands the cart to the merchant’s checkout, and the checkout uses whatever rail the merchant already has. That is why Visa and Mastercard appear in the announcement as partners and not as components. Anthropic built the brain. The networks own the wallet. The merchant owns the store.
Two paths to the AI shopper
AI shopping agents reach a merchant on two paths. On the first path, the merchant runs its own agent on its own storefront. The Claude blueprint is built for this path, and platform features such as Shopify’s agentic storefronts serve it too. On the second path, the shopper talks to someone else’s agent, such as ChatGPT, Google AI Mode, Gemini, Copilot or Perplexity, and that agent reaches the merchant through a commerce protocol. The two paths have different economics. Most merchants need both. The Elogic Commerce AI in ecommerce guide covers the second path in detail.
| Question | Your own agent (Claude blueprint, platform-native agent) | Other people’s agents (ChatGPT via ACP, Google AI Mode and Gemini via UCP) |
|---|---|---|
| Who owns the conversation | You | The assistant platform |
| Who owns the customer data and memory | You, under your consent and retention rules | The platform; you receive order data |
| Who owns checkout | You | ACP: the platform-mediated checkout with Stripe; UCP: your checkout or Google Pay, merchant-hosted |
| Fee on the transaction | Your normal processing cost, plus model cost per completed task | ACP: a platform fee, reported at 4% per transaction; UCP: no platform fee reported |
| Discovery | Your existing traffic and logged-in customers | The platform’s ranking; ACP is curated, UCP is open to any compliant merchant |
| Payment credential | Your PSP, with Visa and Mastercard agent tokens as they roll out | Stripe Shared Payment Token (ACP), Google Pay, Shop Pay, network agent tokens |
| Where it wins | High-consideration and logged-in journeys, B2B, service and returns, merchant operations | Top-of-funnel discovery and first purchases |
| Lock-in | Low: the architecture adopts a new model as a config change | Medium: catalog feed formats and approval processes per platform |
The first path is where the margin stays. The second path is where the new traffic comes from. Visa and Mastercard serve both paths. Intelligent Commerce Connect explicitly accepts payments over TAP, MPP, ACP and UCP, which means a merchant that enables agent tokens once can accept an order from its own agent and from ChatGPT with the same credential model.
What the 2026 data says
The demand is measured, not forecast. Adobe Analytics tracks more than one trillion visits to US retail sites. Its 2026 figures, as reported by Adobe and Digital Commerce 360:
- AI-referred traffic to US retail sites grew 693% year on year in the 2025 holiday season, 393% in Q1 2026, 138% in May 2026 and 62% in July 2026. Since October 2024 it has grown 1,219%.
- In March 2025, AI-referred traffic converted 38% worse than non-AI traffic. In March 2026 it converted 42% better. By May 2026 the gap was 54%, and by July 2026 it was 60%.
- In July 2026, AI-referred visitors generated 53% more revenue per visit, spent 59% more time on site, bounced 33% less and added items to cart 28% more often than other visitors.
- In Adobe’s survey, 39% of consumers had used AI for online shopping and 85% of them said it improved the experience.
- Adobe also reports that about 34% of retailer homepage content is not readable by AI models, and product pages average 66% visibility.
Shopify reported AI-referred traffic up about 8 times and AI-referred orders up nearly 13 times year on year in Q1 2026. Accenture’s research, quoted in the Anthropic announcement, found 85% of consumers open to collaborating with an AI agent, and nearly three in four would trust a personal AI agent more than their best friend to make a purchase for them.
Elogic Commerce reads these numbers in one way. The traffic is real and it converts, but almost all of it is still referral traffic: the assistant sends a person to the site and the person pays. The next step is the one the networks and Anthropic just built for: the agent pays. When that happens, the conversion figures above stop describing visitors and start describing orders that arrive without a session.
The future for merchants: five shifts
The future of ecommerce is not a chat window on the storefront. It is a set of changes to what a store must be able to do when the customer is software. Elogic Commerce sees five shifts, each with a second-order effect that arrives later and costs more.
Shift 1. Agentic checkout turns checkout into an API with a policy
Today checkout is a page. For an agent it is a contract: what an agent may buy, in what quantity, at what price, with which fields protected. The blueprint’s rules are the template. Writes accept only server-issued IDs. Caps apply to the resulting state, not the request. Fee and disclosure text comes from approved copy, byte for byte. Second-order effect: legal, finance and engineering must own one machine-readable policy together. Most merchants have this policy spread across three teams and no document.
Shift 2. The storefront becomes a data product
An agent buys from the catalog, not from the page. If the attribute is missing, the agent cannot compare, and the product is invisible. Adobe’s finding that a third of homepage content is unreadable by AI is the same problem seen from the crawler side. Second-order effect: the PIM and the product feed become revenue infrastructure, and product data quality becomes a board metric. Elogic Commerce’s ecommerce GEO services exist because this shift arrived before the agents did.
Shift 3. Bot blocking becomes agent admission
For a decade the security team blocked automated traffic. Now a signed agent request from Visa’s Agent Directory is a customer. Cloudflare and Akamai already carry TAP. Second-order effect: the WAF rule set is now a sales-channel decision. Blocking unverified agents is still correct. Blocking all agents is a revenue decision made by the wrong team.
Shift 4. Memory and consent become regulated assets
An agent that remembers a shopper’s size, allergy or default store is more useful. It is also holding personal data. The engineering guide treats memory as a data-handling problem: a validator on the write path, a way for users to see and delete facts, a retention period, and a per-deployment switch so regions that cannot carry the obligation run without memory. Agentic commerce compliance starts here, before payments. Second-order effect: EU merchants will run agents with less memory than US merchants, and the experience gap will be visible.
Shift 5. Margin moves to whoever owns the agent
An order placed through a third-party assistant carries that platform’s fee on top of card fees. An order placed through the merchant’s own agent carries card fees and model cost per completed task, and the blueprint’s prompt caching brings the model cost down to a small fraction at 90 to 99% cache hit rates. Second-order effect: merchants with their own agents gain negotiating leverage with the platforms, the same way merchants with their own apps did with marketplaces. Machine customers, agents that buy on a schedule for a business, will be the first to prefer the merchant’s own agent because they need the policy contract, not the discovery.
The future of agentic commerce for a merchant, then, is not a choice between paths. It is a sequence: fix the data, expose the checkout as a contract, admit verified agents, run your own agent where the margin is, and use the protocols where the traffic is.
What to do in the next 90 days
| Window | Actions | Output |
|---|---|---|
| Days 1 to 30: audit | Measure how much of the catalog an AI model can read: attributes, availability, price, policies.Ask the PSP whether it accepts Visa and Mastercard agent tokens today and what the agent flag looks like in the authorization message.Pull edge logs and count signed agent requests versus unknown bots.Write the checkout contract on one page: what an agent may buy, caps, protected fields. | A readiness score, a PSP answer, a one-page policy |
| Days 31 to 60: pilot | Pick one narrow flow: product discovery for a category, reorder for logged-in accounts, or the merchant agent on inventory.Fork the blueprint or enable UCP or ACP on the platform, depending on the path.Keep the stage-and-approve pattern. No agent action moves money.Write 50 to 100 evaluation cases per flow with the people who see the failures: care, pricing, legal. | One agent flow in production behind a switch, with an eval suite |
| Days 61 to 90: harden | Turn on TAP at the edge and whitelist directory agents.Enable agent tokens with the PSP and map the agent flag into risk rules.Enforce caps on cart state and serialize cart writes per session.Measure cost per completed task and cache hit rate, not cost per call. | Agent-initiated payments accepted; unit economics known |
The future for ecommerce platforms
AI agents in ecommerce need seven things from a platform: a catalog API with full attributes, a cart API, a checkout handoff, order status, machine-readable policies, agent identity at the edge and agent token acceptance in payments. An agentic commerce platform is a platform that supplies all seven without custom work. In September 2026, none supplies all seven natively, and the gaps decide where an integration partner is needed. The table uses public information; Elogic Commerce scores the platforms in more depth in the Agentic Commerce Readiness Index.
| Platform | Native agent surface | Protocol and payment status | What to watch |
|---|---|---|---|
| Shopify Plus | Sidekick, agentic storefronts; a reference storefront on the Claude blueprint connecting through Catalog, UCP and Shop Sign-in | UCP native, co-developed with Google; Shop Pay in UCP; scored 86 of 100 in the Elogic Commerce Agentic Commerce Readiness Index | How fast the Claude reference storefront reaches general availability; ACP for ChatGPT alongside UCP |
| Adobe Commerce | Adobe’s AI features for catalog and merchandising; agents are custom-built against GraphQL and REST | No native ACP or UCP announced; strongest for the own-agent path on complex B2B catalogs; needs an integration partner such as Elogic Commerce’s Adobe Commerce developers | Adobe’s agentic roadmap; partner-built UCP adapters |
| Salesforce Commerce Cloud | Agentforce commerce agents | Own-agent path is native; payments through existing SFCC cartridges | Agentforce support for network agent tokens |
| commercetools | None native; API-first design fits the tools-over-systems pattern | Composable checkout; protocol adapters are straightforward to build | Partner adapters for UCP and ACP |
| BigCommerce | AI features in the admin | Native ACP support reported, which reaches ChatGPT shoppers | UCP support; token acceptance through its payment partners |
| Shopware | Agentic Commerce Lab reference build on Claude (MIT licence) | Reference build proves the own-agent path; protocol support through the community | Movement from lab to product |
| Medusa.js | None native; developer-led | Tools are written directly against the modules; no protocol layer | Community modules for UCP and TAP |
The second-order effect for platforms is larger than for merchants. Platform value moves from storefront themes to agent-grade APIs and policy engines. The platform that ships stage-and-approve primitives, protected fields and per-session write serialization as product features will win B2B, because B2B buyers already work that way. Platforms will also want a share of agent-originated orders, and merchants with their own agents will be the ones able to say no.
What is different in B2B
The card networks built agentic payments for consumers first. Most B2B volume does not run on cards. It runs on net terms, purchase orders, credit lines and invoices, with approval inside the buyer’s procurement system. So B2B agentic payments look different in three ways.
- The token is a purchase order, not a card. A B2B agent stages an order against a contract price and a credit limit. The approval routes to the buyer’s approver, not to an issuer. Virtual commercial cards can carry the network tokens where the buyer wants card rails, but the policy contract lives in the ERP.
- The merchant agent is the first production win. Inventory alerts, contract-price checks, reorder proposals and campaign drafts sit inside the blueprint’s merchant agent today, with human approval built in. For a distributor, that is worth more in the first year than a consumer shopping agent.
- The integration is the product. The blueprint’s rule is to build tools on top of the systems you already run. In B2B those systems are SAP, Microsoft Dynamics 365, Oracle NetSuite and Odoo. The agent is only as good as the ERP connection under it.
Elogic Commerce builds on this pattern for manufacturers, distributors and B2B2C brands. The agent proposes. The ERP approves. The person signs off. That is the same shape as the blueprint’s maker-checker rule, applied to a business that already had one.
Agentic Payments Readiness Checklist
Elogic Commerce uses this checklist to score a merchant before an agent pilot. It complements the platform-level Agentic Commerce Readiness Index. Twelve checks, four areas. A merchant that passes eight is ready to pilot. A merchant that passes eleven is ready to accept agent-initiated payments.
| Area | Check | Pass condition |
|---|---|---|
| Data | 1. Catalog completeness | Every sellable item has price, availability, key attributes and policy text readable through an API |
| Data | 2. Policy documents | Returns, shipping, fees and disclosures exist as structured text with an owner and a version |
| Data | 3. Product feed | A feed or UCP manifest validates without errors and updates within the hour |
| Checkout | 4. Checkout API | A cart can be created, priced and handed to checkout without a browser session |
| Checkout | 5. Checkout contract | A one-page policy states what an agent may buy, caps, and protected fields |
| Checkout | 6. Cap enforcement | Limits apply to the resulting cart state and cart writes are serialized per session |
| Payments | 7. Agent token acceptance | The PSP accepts Visa and Mastercard agent tokens and passes the agent flag to risk rules |
| Payments | 8. Strong authentication | Passkeys or SCA are available for agent-initiated orders in regulated markets |
| Payments | 9. Intent logging | The order record stores the agent identity and, where present, the intent reference |
| Trust and control | 10. Agent admission at the edge | TAP signatures are verified and directory agents are admitted; unverified automation is still blocked |
| Trust and control | 11. Stage-and-approve | No agent action moves money or changes prices without a person or a policy applying it |
| Trust and control | 12. Evaluation suite | 50 to 100 cases per flow, including injection, cross-user data and regulated-language checks, run before every release |
Frequently asked questions
What are agentic payments?
Agentic payments are payments that an AI agent starts on behalf of a person, with the person’s prior consent, inside limits the person set, using a token that replaces the card number and a signature that identifies the agent. Visa and Mastercard both run programs for them: Visa Intelligent Commerce and Mastercard Agent Pay.
What does agentic payment mean?
An agentic payment means the payer is software acting for a person, not the person clicking pay. The word agentic refers to an AI agent that can take actions. The payment still runs on card rails, but it carries three extra things: a scoped credential, proof of the agent’s identity and a record of the person’s intent or policy.
Can you give me an example of an agentic payment?
A shopper asks a retailer’s AI agent for a tent, a sleeping bag and a stove under 400 euros. The agent builds the cart. The shopper taps approve. The checkout receives a Visa token scoped to that agent, verifies the agent’s signature, and the issuer approves within the 400-euro limit. The agent never saw the card number.
How do agentic payments work?
In four steps. The person sets a policy and enrolls a card. The network mints a token bound to the agent and the policy. The agent signs its requests so the merchant can verify it. The merchant’s processor sends the token to the network, which checks the policy before the issuer approves. Settlement uses the merchant’s existing card rails.
Does Claude Commerce Agents process payments?
No. Anthropic’s blueprint builds the cart and hands it to the merchant’s checkout. No tool call in the blueprint can charge a card. Payment runs through the merchant’s existing checkout or an agentic payments provider. Visa and Mastercard are named as partners that help merchants accept agent-initiated payments, not as components of the blueprint.
Is Anthropic moving into commerce or payments?
Anthropic’s own FAQ answers no. The company supplies the model and an open-source reference implementation. It does not run a marketplace, a checkout, a wallet or an advertising layer. Merchants keep the customer relationship and checkout, and the card networks keep the credential.
What is the difference between Visa Intelligent Commerce and Mastercard Agent Pay?
Both issue card tokens scoped to an AI agent. Visa’s Trusted Agent Protocol verifies the agent with HTTP message signatures and an Agent Directory, and Intelligent Commerce Connect is the merchant on-ramp that accepts TAP, MPP, ACP and UCP payments. Mastercard’s Agentic Tokens bind the credential to the agent and a consent policy, and Verifiable Intent adds a signed record of what the person asked for. Visa verifies the agent. Mastercard verifies the agent and the intent.
If I build my own AI shopping agent, do I still need UCP or ACP?
Usually yes, for reach. Your own agent serves the shoppers already on your site and your logged-in customers. UCP reaches Google AI Mode and Gemini shoppers, and ACP reaches ChatGPT shoppers. Both paths can use the same agent tokens from Visa and Mastercard, so enabling agent-initiated payments once serves both.
Who are the best providers for agentic payments?
For card volume, the providers are the networks, Visa Intelligent Commerce and Mastercard Agent Pay, delivered through your processor. Stripe, Adyen and Checkout.com have announced or piloted support. American Express, Discover and PayPal have their own programs. For machine-to-machine and micropayments, Stripe’s Machine Payments Protocol and Coinbase’s x402 are the current options. The right provider is the one your processor already supports.
How can marketplaces and B2B sellers prepare for agentic commerce?
Start with data and the checkout contract, not the agent. Make the catalog fully readable through an API, write a one-page policy for what an agent may buy, and confirm agent token support with the processor. For B2B, connect the agent to the ERP and route approvals to the buyer’s approver. Elogic Commerce recommends the merchant agent as the first pilot for distributors, because it produces value with human approval built in.
About Elogic Commerce
Elogic Commerce is a B2B and B2B2C commerce engineering agency founded in 2009, headquartered in Tallinn with offices in New York, London, Stockholm, Dresden and Prague, and a team of more than 200 specialists. It builds on Adobe Commerce, Shopify Plus, BigCommerce, Salesforce Commerce Cloud, commercetools, Shopware and Medusa.js, and connects storefronts to SAP, Microsoft Dynamics 365, Oracle NetSuite, Odoo and Visma. Elogic Commerce is a member of Anthropic’s Claude Partner Network, an Adobe Commerce Silver Solution Partner and a Hyvä Bronze Partner, and holds ISO 27001, SOC 2 Type II and ISO 9001 certifications. Its AI work includes AI solutions for ecommerce, the Agentic Commerce Readiness Index and AI Chat Core.
Ready to accept agent-initiated payments? Elogic Commerce runs a two-week agentic payments readiness assessment: catalog readability, checkout contract, processor token support and edge admission, scored against the checklist above. Talk to an Elogic Commerce architect.
Sources
- Anthropic, “Building commerce agents with Claude”, September 2, 2026. https://claude.com/blog/claude-for-commerce-agents
- Anthropic, “A guide to the anatomy of effective commerce agents”, September 2, 2026. https://claude.com/blog/the-anatomy-of-effective-commerce-agents
- Anthropic, commerce-agents repository. https://github.com/anthropics/commerce-agents
- Finextra, “Anthropic launches AI commerce agents with Visa and Mastercard”. https://www.finextra.com/newsarticle/48346/anthropic-launches-ai-commerce-agents-with-visa-and-mastercard
- PYMNTS, “Anthropic built the shopping brain and skipped the wallet”, September 3, 2026. https://www.pymnts.com/news/artificial-intelligence/2026/anthropic-built-the-shopping-brain-and-skipped-the-wallet/
- Visa, “Visa Opens the Door to AI-Driven Shopping for Businesses Worldwide”, April 8, 2026. https://corporate.visa.com/en/sites/visa-perspectives/newsroom/visa-intelligent-commerce-connect-ai-shopping-for-businesses.html
- Visa Developer, Trusted Agent Protocol. https://developer.visa.com/capabilities/trusted-agent-protocol
- Visa, “Visa and Partners Complete Secure AI Transactions”, December 2025. https://investor.visa.com/news/news-details/2025/Visa-and-Partners-Complete-Secure-AI-Transactions-Setting-the-Stage-for-Mainstream-Adoption-in-2026/default.aspx
- Mastercard, Mastercard Agent Pay. https://www.mastercard.com/us/en/business/artificial-intelligence/mastercard-agent-pay.html
- Mastercard, “Mastercard unveils Agent Pay in Latin America and the Caribbean”, December 2025. https://www.mastercard.com/news/latin-america/en/newsroom/press-releases/pr-en/2025/december/mastercard-unveils-agent-pay-in-latin-america-and-the-caribbean/
- Adobe, “AI traffic grows but retail sites lag in AI search visibility”, April 2026. https://business.adobe.com/blog/ai-traffic-surge-retail-sites-not-machine-readable
- Digital Commerce 360, Adobe AI-referral data, June and August 2026. https://www.digitalcommerce360.com/2026/08/19/adobe-ai-referral-traffic-data-july-2026/
- Universal Commerce Protocol and Agentic Commerce Protocol comparisons, 2026. https://universalcommerceprotocol.blog/en/agentic-commerce-protocols-compared/
Get in Touch
Looking for a partner to grow your business? We are the right company to bring your webstore to success.