One customer asked the AI assistant to find running shoes under $150 and buy them based on reviews. Another customer programmed the agent to reorder office supplies whenever the inventory dropped to a certain level.
The AI is no longer just recommending products for customers to purchase. Instead, the AI is participating in the purchase itself.
This concept is known as agentic commerce—transactions between consumers and artificial intelligence. In 2026, agentic commerce is moving from concept to infrastructure for commerce and payments. According to Visa, nearly half of American shoppers used AI during their shopping experience last year.
For merchants, the challenge is to accept payments from AI agents without becoming vulnerable to fraudulent transactions.
How Agentic Commerce Changes Ecommerce Checkout
Traditional ecommerce assumes a person is interacting directly with a website or app. Agentic commerce introduces software between the customer and merchant.
That agent may:
- search the merchant’s catalog
- compare products
- check inventory
- evaluate shipping costs
- apply customer preferences
- build a cart
- choose an approved payment method
- complete checkout
- monitor fulfillment
- initiate a return
Emerging standards are being built to make those interactions more structured.
Google’s Universal Commerce Protocol (UCP), introduced in 2026, provides a common way for agents, businesses and payment providers to exchange commerce information across discovery, purchasing and post-purchase interactions. Its Agent Payments Protocol (AP2) adds authorization records that can show what the customer permitted an agent to do.
AP2 can establish guardrails such as an approved merchant, spending limit or refundability requirement. A payment authorization can then be tied to the specific cart the agent intends to purchase.
Visa and Mastercard are addressing a related problem: how does the merchant know the automated visitor is actually a trusted shopping agent?
Visa’s Trusted Agent Protocol uses signed information designed to help merchants recognize approved agents, identify their commerce intent and receive payment or customer information through existing web and API infrastructure. Mastercard’s Agent Pay framework similarly uses registered agents and agent-specific network tokens so merchants and payment participants can recognize agent-initiated transactions.
The result is an important distinction:
| Ecommerce Traffic | Merchant Question |
|---|---|
| Human shopper | Is this customer and payment legitimate? |
| Ordinary web crawler | Should this automated traffic access the site? |
| Malicious bot | Is this fraud, scraping or card testing? |
| Trusted shopping agent | Is this recognized software acting with customer authorization? |
| Unknown AI agent | What evidence proves who it represents and what it is authorized to do? |
Merchants should not solve agentic commerce by simply relaxing bot controls. They need a way to recognize trustworthy automation without opening checkout to malicious automated traffic.
How AI Agent Payments Work
An AI agent does not necessarily need access to a customer’s reusable card number to complete payments.
Instead, new models for agentic payments use tokenized credentials and customer-authenticated instructions to enable agents to make purchases on behalf of their customers.
Visa’s Intelligent Commerce initiative, for example, is creating payment tokens and controls for AI agents that associate each payment with the customer’s original instructions for the agent. The system can also provide merchants with signals about commerce transactions that link the purchase to the customer’s intent.
Mastercard’s Agent Pay system uses its registered agents to tokenize customer purchases and verify customer intent behind those purchases.
Google’s AP2 system uses digital mandates to allow customers to authorize agents to purchase specific items from a customer’s cart. The purchase record will note which cart was purchased. In April 2026, Google released version 0.2 of AP2, which includes support for “Human Not Present” transactions based on the customer’s pre-authorized instructions to the agent.
For merchants that receive these payments from AI agents, the systems aim to answer the following questions:
- Which agent initiated the purchase?
- Which customer authorized the agent to make the purchase?
- Which items was the agent permitted to purchase for the customer?
- What were the spending limits for the agent making the purchase?
- Which cart was submitted for payment?
- Which payment credential was used to complete the transaction?
- Did the purchase match the customer’s instructions for the agent?
This trail of purchase information will be especially important if the customer later claims that they did not authorize their AI agent to make that purchase.
Get Checkout and Fraud Systems Ready for AI Agents
Merchants don’t have to replace their entire payment system to accommodate AI agents shopping on their sites.
Many of the agentic payment systems being developed are compatible with existing e-commerce systems and payment methods. Visa’s Trusted Agent Protocol, for example, works with both existing e-commerce websites and payment APIs.
However, effort must be made to make the data and controls surrounding the e-commerce site agent-ready.
Make Product and Policy Data Machine-Readable
For AI agents to shop on merchant sites, there must be accurate product data available about the following:
- Price
- Inventory
- Variants
- Shipping
- Taxes
- Returns
- Warranties
- Subscriptions
- Geographic restrictions
- Cancellation policies
Visa’s 2026 research into consumer trust highlights product data as a concern for merchants. If not implemented correctly, it could cause issues during checkout for AI agents. If the data on return policies is not clearly defined for human customers, the AI agent will simply parse the information it finds and proceed with the purchase.
Preserve Customer Intent
Merchants should retain information about each transaction to reconstruct it in its entirety.
Information that may be useful in reconstructing the transaction may include:
- identity of the agent
- customer and account information
- customer authorization
- spending and purchase constraints
- contents of cart
- amount to spend
- payment token information
- authorization result
- shipping information
- timestamp of transaction
- fulfillment records
- cancellation records
- return and refund records
While it is important to retain this information, it is only for the purpose of demonstrating which transactions were authorized and completed by the merchant—not for retaining any sensitive payment token information.
Confirm Gateway and Processor Compatibility
A merchant account, gateway, and AI checkout layer all have different jobs.
Payment Nerds works with NMI and Authorize.net for e-commerce and other merchants that are eligible for these merchant accounts. However, each merchant will have a different gateway and acquiring bank, so compatibility with these components is essential before integrating an agentic checkout system.
Before implementing an agentic checkout for merchants, these questions should be answered:
- Does your payment gateway accept this tokenized payment credential?
- Does your gateway support agentic transaction metadata?
- Do your fraud detection tools understand agentic transactions?
- Are your checkout system’s APIs compatible with agentic checkout?
- Will your acquiring bank accept these transactions?
- Can your reporting package identify transactions that originated from agentic checkouts?
- How will your checkout gateway handle refunds?
These questions should be answered before developing any agentic checkout software. The specifications and integrations for agentic checkout systems are still emerging and evolving rapidly.
Control VAMP Risk and Avoid Agentic Commerce Mistakes
Agentic commerce will introduce a unique type of fraud in which some legitimate customers may appear as bots.
Visa has recognized this issue in its Trusted Agent Protocol. Merchants have historically blocked bots because many of them look like credit card data-scraping tools or even malicious bots. The Trusted Agent Protocol was established to help merchants distinguish between commercial agents and bots.
This distinction is specifically important for understanding how the Visa Acquirer Monitoring Program (VAMP) monitors Visa fraud reports and disputes. VAMP also monitors for enumeration activities. Enumeration involves bots attempting to enter credit card details into shopping carts to determine which details are valid for online transactions.
Merchants should not whitelist all AI traffic to avoid blocking shopping agents. This will allow automated activity that is flagged by anti-fraud systems to pass through undetected.
Instead, when using AI shopping agents, merchants can rely on signals of identity and intent to verify the transactions. These signals can include information like:
- cryptographically verified agent identity
- authenticated customer intent
- tokenized payment credentials
- authorization velocity
- device and network signals
- transaction amount
- cart consistency
- account history
- unusual decline patterns
- enumeration detection
The fact that an agentic system initiated a transaction does not exempt it from fraud, dispute, or acquiring program controls placed upon all transactions that traverse the Visa payment network.
Common Agentic Commerce Mistakes
Many problems can be avoided by treating the AI checkout as a new channel for transactions rather than simply adding another feature to an existing website.
Common mistakes include:
- allowing any bot to identify itself as a shopping agent
- disabling fraud rules for shopping agents
- allowing shopping agents to access card credentials
- failing to preserve proof of authorization from the customer
- attempting to process a cart for a value that exceeds the customer’s limit
- hiding shipping and subscription terms from the product data
- failing to distinguish between shopping agent traffic and other traffic in reporting
- assuming that using a shopping agent eliminates the potential for chargebacks
- ignoring returns for products purchased with a shopping agent
- building the AI shopping functionality around only one protocol
- failing to test for fraudulent shopping agents
- allowing shopping agents to authorize without checking for card testing
- ignoring VAMP rules because the transaction originated from a shopping agent
- assuming that the current payment gateway will support all new agentic commerce payment standards
Finally, merchants should leave room for human intervention within the shopping agent. For example, it’s likely the agent will be used to purchase ordinary office supplies. However, the customer may want to authorize purchases of expensive products, subscriptions, or non-refundable items.
Agentic Commerce and AI Payment Questions
Q: What is agentic commerce?
A: Agentic commerce is a concept where AI agents are able to complete tasks for the customer, such as finding products, comparing prices, adding items to a cart, and even completing the payment process for the customer. The customer establishes the AI agent’s authority.
Q: What are AI agent payments?
A: AI agent payments refer to instances in which an AI agent takes action to complete a payment for a consumer or business. Technologies like tokenization and authenticated AI agent instructions can associate those payments with the consumer who authorized the agent.
Q: Can an AI agent make a purchase without the customer approving every transaction?
A: Depending upon the AI and virtual assistant software being used, the customer may authorize the agent to make purchases from specific merchants or to a specified spending limit without having to approve each individual transaction.
Q: What is the difference between UCP and AP2?
A: UCP ensures that the communication between the agent and the commerce platforms is standardized. AP2 will focus on authorizing the payment and providing proof of the transaction, as per the client’s instructions.
Q: How can a merchant differentiate between an AI shopper and a malicious bot?
A: There are various signals that trusted agent frameworks emit when they are acting on merchants’ behalf that can be used to differentiate between bots and agents. These signals can be used in addition to the merchant’s current fraud control systems.
Q: Do merchants need a new merchant account to facilitate agentic commerce?
A: Merchants will not necessarily require a new merchant account. Various agentic payment systems are being developed to work with merchants’ existing payment systems. However, merchants will have to ensure that their current merchant account can take the required payment credentials.
Q: Will AI agent purchases have chargebacks?
A: Card payments can lead to chargebacks. While the agentic-payment frameworks will keep better records of customer intent, merchants will still need to present proof of fulfilled purchases.
Q: How does VAMP apply to AI agents?
A: Visa’s protocols will still apply to all transactions, regardless of the purchasing agent. However, merchants must ensure that purchases made by AI agents do not weaken fraud controls designed to prevent bots from enumerating products and purchasing them.
Q: Should merchants implement agentic checkout now?
A: Merchants do not need to rebuild their checkout systems right away. However, e-commerce businesses may want to start reviewing their products and the APIs that facilitate their transactions.
Get Ready for AI-Native Checkout
AI agents are already purchasing products. Visa, Mastercard, and Google are all building the infrastructure around the same requirements: identify the agent, verify the customer’s intent to purchase the product, protect the payment credential, and have enough information to understand what happened during the purchase.
Merchants do not need to surrender control of the customer checkout experience to AI agents. Make commerce data easier for AI agents to understand, verify what the current gateway can handle, and leave fraud, VAMP, and dispute controls to the merchants themselves. The merchants of the future will not simply accept AI purchasing agents but will understand which automated customers to trust.
Sources
- Visa. “Visa Intelligent Commerce.” Accessed August 2026.
- Visa. “Trusted Agent Protocol.” Accessed August 2026.
- Visa. “Trusted Agent Protocol Specifications.” Accessed August 2026.
- Visa. “5 Trends Shaping the Future of Agentic Commerce.” Accessed August 2026.
- Google Developers. “Under the Hood: Universal Commerce Protocol.” Accessed August 2026.
- Google Developers. “Developer’s Guide to AI Agent Protocols.” Accessed August 2026.
- Google. “Agent Payments Protocol and the FIDO Alliance.” Accessed August 2026.
- Mastercard. “Mastercard Agent Pay.” Accessed August 2026.
