A new Wild West is opening up for the payment industry. In recent weeks, Visa, Mastercard, Coinbase, and PayPal have announced their proposals for so-called agentic payments. The stakes are high – creating a background for trade supported by artificial intelligence.
Let's imagine the near future. The holiday season is approaching and we want to plan a vacation, let's say on the Greek islands. Instead of struggling through hundreds of offers, we assign the task of booking flights and hotels to our AI assistant. Based on our preferences, the AI finds the right components of a “vacation package” and… that's it.
Today, our “intelligent agent” (operating autonomously on the network, carrying out the task imposed on it, e.g. ChatGPT Operator) would turn out to be just a slightly improved version of the Internet search engine. It would be able to find and sift through offers, but the obstacle would be the reservation and payment. In particular , the last stage, payment, requires human intervention . Somewhere in the background there must be a payer – a human who agrees to spend the funds belonging to him and who is responsible for the transaction. The problem is to be solved by so-called agentic payments , i.e. tools that give artificial intelligence the ability to use money .
4 models of agentic payments
The problem of AI agent “payment dependency” can be solved in several ways. The available options were neatly presented on his blog by one of the best fintech experts – Simon Taylor:
- The agent can only mediate in the payment process, but the necessary data is provided by a live person. This model is technically feasible today – after completing part of the task, the agent waits, for example, for the payment card data to be provided and the transaction to be approved by the user.
- The agent stores the data necessary to make a payment transaction (something like card on file, i.e. saved payment card data, as in services requiring a subscription) and only requires confirmation of the operation by a human. This slightly simplifies the process, but it is still far from the agent's autonomous operation.
- The agent has its own “virtual payment card”, which has parameters imposed by the user – e.g. spending limit, set of industries (MCC codes) in which it can be used to pay, etc. The card can be issued on demand, as a one-time instrument in the form of a token.
- The agent has its own wallet holding stablecoins. In this world, transactions triggered by algorithms are “normal,” and the payment can be almost invisible to the user, although they will probably want to have the final say in confirming the selected option.
The last two options show the two most likely directions – payment cards (and the large card organizations behind them) and competitors from the crypto world will probably fight for new territories of digital trade.
Stablecoins have obvious advantages at first glance. They are designed for the digital world, and programmable transactions (where a flow of funds triggers an algorithm based on a set of conditions) are built into their capabilities from the start. You could say that giving an AI agent a wallet and a key to it is like a parent sending a child to the store with cash . They have freedom of action, up to the amount we have funded in the wallet.
Payment cards, in turn, are more than just a piece of plastic. First, payment organizations have built a complex system of connections between merchants, financial institutions and intermediaries operating on a global scale. Second, they are backed by a “trust architecture” – rules for settlements, division of responsibility, risk management. Third, they are constantly modifying their main product, as evidenced by the vision of a card as an identifier ( credential ) giving access to various payment instruments (see the vision of Visa Flexible Credential, which we wrote about in Bankier.pl). It is therefore only a matter of time before card giants adapt to the world of agentic commerce.
A flood of announcements, agentic commerce is coming
That the competitive battle on the new market front is beginning in earnest is evidenced by a series of events in recent weeks. Almost every major player in the payments world (except, interestingly, Google and Apple) has presented its vision of agentic payments.
Visa has proposed a package called Visa Intelligent Commerce. The platform is to enable AI agents to perform transactions using tokenized payment cards. Payers will retain control over their money, receiving tools that allow them to precisely define the boundary conditions of transactions. Partners of the venture include OpenAI, Perplexity, Anthropic, Stripe and IBM.
Mastercard tried to “steal the show” from its main competitor and presented Mastercard's Agent Pay a few days before Visa's presentation. The vision is similar and is based on so-called agentic tokens. Microsoft appears among the partners.
PayPal has gone a bit further. PayPal Agent Toolkit is a set of tools for developers that lets them connect to PayPal account features. AI agents will be able to use PayPal APIs, including to track, for example, order fulfillment and delivery stages.
Interesting news also came from the stablecoin camp. Coinbase presented the x402 protocol, based on forgotten options of the foundation of internet services. Internet users are well familiar with the “404 error” – this is a message that appears when the server we are visiting does not find the requested page. The 402 response, on the other hand, means “payment is required”, e.g. to open a page with paid content.
Coinbase proposes an open standard for transactions based on the 402 code. When the AI agent receives a response with this code, it will “pull out the wallet” and pay, and then gain access to the resource (e.g., a paid API).
“Much like HTTPS secured the web, x402 has the potential to define the next era of the internet; one where value moves as freely and instantaneously as information. We are laying the foundation for an economy run not just by humans, but by software—autonomous, intelligent, and always available,” Coinbase’s Erik Reppel said in a statement.
Technology alone is not enough
Solving the challenges related to transactions carried out by autonomously operating agents will probably require more than just well-thought-out technological standards. There are many potential pitfalls lurking in the background, for example, of a legal nature. What will be the framework for an “unauthorized transaction” when payment is initiated by artificial intelligence? What should the procedures for complaints and handling errors look like? This is what pioneers in the new market will have to deal with.
