Human-Not-Present: Agentic Commerce, Network Tokens and Payments

Have you ever wished that someone else could handle your shopping for you? In early September, Visa and Mastercard took a step toward making that a reality.

Both companies announced new sets of tools aimed at improving integration with AI-powered payments systems and, specifically, AI agents. This comes after both companies announced their new agentic ecommerce platforms earlier in the spring: Visa Intelligent Commerce and Mastercard Agent Pay. Both platforms are designed to support AI systems that can make secure, autonomous purchases on behalf of users.

This new era of agentic commerce will require new technologies and a new approach to online payments. But some key pieces of the puzzle, like network tokenization, are already in place. 

In this article, we’ll look at what AI agents are, what they could mean for the future of online shopping, and how payments infrastructure will need to evolve to support fast, secure and truly autonomous transactions.

Today, the term “AI agent” is used loosely, which has caused no shortage of confusion. So, let’s lay down some foundational ideas so that we’re all on the same page.

A true AI agent is a system that can analyze information, make decisions and then act autonomously. That means it can complete complex, multi-step tasks without the need for human intervention. It still requires initial instructions from a user, who can always go in and change the rules or shut it down. But once an agent has been programmed, it has the capacity to act on its own. That’s the key differentiator. A true AI agent is not just a language learning model (LLM) acting as a step in an automation loop triggered by something else. It runs its own loop.

Today, even the most advanced AI agents fail 70% of simple tasks. The flagship agents from the biggest players, like OpenAI, are demonstration-only systems that require constant human intervention to work. But Visa’s and Mastercard’s initiatives are laying the foundation for a future in which true AI agents can operate with high-enough accuracy and reliability that they can autonomously make purchases on behalf of users. That future may still be years away, but when it arrives, the payments space will need to be ready for it.

The eventual goal of agentic ecommerce is to allow shoppers to give an AI system plain-language shopping instructions, in the same way we currently interact conversationally with ChatGPT. From there, the system will act as an autonomous personal shopper. The agent will be able to go out into the (digital) world and make purchases, with the full authority to buy within rules and limits, just like a human personal shopper would.

A common example cited is searching for flights. It’s already easy to get alerts when flight prices drop. But a truly agentic AI system could regularly perform new searches, compare price trends and, most importantly, actually book and pay for a flight, all on its own, without constant prompting. All the user would have to do is set the rules and then wait for an email with their ticket confirmation.

You can jerry-rig AI systems to do something that sort of resembles this today, but the system can break easily and can’t rerun the task on its own. It either succeeds on the first try or not at all. A true agent would be able to run as many searches as necessary, for days, weeks or even months until it finds the right deal. A single error wouldn’t cascade and break the task because the agent could simply restart the loop correctly.

Imagine your weekly grocery shopping. With agentic ecommerce, you could give an agent your grocery list and rely on it to set up deliveries on its own. It could autonomously check the offers each week and source all your standard items from different stores at the best possible prices, factoring in delivery costs, rewards and more. In an even more advanced potential future application, it could be set to dynamically adjust your order based on a computer vision system in your refrigerator or even your monthly spending, through an application program interface (API) connection with your bank.

To date, progress towards agentic shopping has been slow. Perplexity Buy With Pro launched in November 2024, enabling users to do product research through the LLM’s chat interface and then click a buy button embedded within the results to complete a purchase. But while that was a big step for AI-enabled shopping, it still requires significant user intervention to complete the task, so it’s not really an agent. OpenAI also now offers shopping through ChatGPT, but, again, it isn’t agentic.

Systems like Mastercard Agent Pay and Visa Intelligent Commerce will probably enable Perplexity-like shopping flows on other platforms in the very near future, but, ultimately, their goal is to help power the full-agentic experience that AI companies are working toward.

Making Agentic Payments a Reality

A big part of the gap that needs to be closed for highly autonomous ecommerce to work is the payment itself. Current AI shopping tools, like the ones from Perplexity and OpenAI, enable conversational shopping and make product research faster and easier, but they still require the user to click to buy. Fully agentic commerce would eventually eliminate that step. But how?

The key is the API connections that agentic AI systems will use to process payments. Visa’s Intelligent Commerce Developer Center identifies five critical APIs that will all work together to balance security, authentication, alignment, and personalization.

API #1: Network Tokenization

When a user adds their credit card to an AI agent and gives it permission to pay, a unique network token is issued to the agent to enable secure, eventually autonomous transactions. This network token works essentially like any other, except the request is sent to the network by an autonomous AI system instead of the merchant.

API #2: Authentication

Authentication is required at multiple steps to ensure safe, authorized use. First, step-up authentication has to happen when issuing the network token. Step-up or passkey authentication is also required anytime a user makes changes to payment instructions. Finally, until agents are truly reliable, authentication will be required before executing transactions.

API #3: Payment Instructions

Users will provide agents with instructions on what they can buy, along with key limitations, like hard spending limits or feature requirements, like a certain clothing size. The payment-instructions API enables the system to validate that the purchase is aligned with the user’s instructions before allowing the agentic payment to go through.

API #4: Signals

A collection of commerce signals sends data back to the card networks, including the outcome of the transaction and the user’s initial instruction. This not only helps the networks refine the system, but it also helps automate and speed up dispute resolution in cases where a user claims their agent acted incorrectly.

API #5: Personalization

Personalization is an opt-in connection that allows a cardholder’s agents to tap into the wealth of data and insights available through the card network. This would allow agents to not only make purchases but also make highly personalized recommendations for new products or services, improving the experience for users.

Think about the way subscription payments work today. With the customer’s permission, a trusted vendor uses a saved payment token to securely rebill and process the payment automatically each period, with no customer reapproval or intervention required. Now imagine if that permission was extended; instead of a vendor reprocessing a single, pre-approved payment, it’s an AI agent using a tokenized credential to autonomously make payments for anything the user has pre-approved it to shop for.

That is the eventual promise of fully agentic ecommerce — saved time, saved money and better product discovery. 

There are obvious hurdles to overcome, like:

• Determining who would be liable when hallucinations or agent misalignment occur
• Convincing shoppers to trust imperfect autonomous systems with their credit cards

But, if there’s one thing we should be used to by now, it’s AI taking tasks that seem fringe at first and shocking us with their performance a year later. So, who knows how quickly true or near-agentic shopping could arrive.

If agentic ecommerce takes off, the entire industry will change. Online merchants will need to optimize their sites for agent traffic that may bypass the normal user interface altogether, and checkout and payment flows will need to adapt to a world where it isn’t just the card that’s not present, but the human, too.