AI agents spend money for you, but the payment system is still stuck in the button era.

Written by: David Marcus, former President of PayPal and Head of Payments and Crypto at Meta

Translated by: AididiaoJP, Foresight News

AI agents can buy and sell for us, but today's payment systems still assume that there is a person at the screen to press "confirm."

In early 2015, while I was at Meta, the company acquired a small natural language company called Wit.ai, and we began collaborating with the Facebook AI Research team (FAIR). That summer, we tested the mobile assistant M in Messenger. M is an assistant that can message like a friend. You direct it to make restaurant reservations, send flowers to loved ones, or arrange travel, and it takes care of everything without needing to touch any Apps. Users quickly loved it.

At launch, I described M as "artificial intelligence trained and supervised by humans." This was not wrong, but we never managed to automate more than 30% of the conversations. At that time, large models had not yet emerged, and the modeling capabilities were insufficient, resulting in demand that could not be met, leading to the product eventually being shut down.

In 2019, we launched Libra, a digital asset native to the internet that could flow globally like sending messages. Libra died due to political reasons, a story I have mentioned elsewhere. I am still grateful for Mark Zuckerberg’s enthusiastic support for these two ventures.

In both cases, I misjudged the timing by a few years. M needed models that could work independently, and Libra required government acknowledgment of digital assets. Now, both of these things have happened. The models are good enough; stablecoins, essentially the prototype of the original Libra, are already written into laws in the United States and Europe.

The two bets I made too early are now happening simultaneously, which deserves serious attention. In my view, the combination of the two means that Apps and screen buttons will disappear, while a whole new way of flowing funds will emerge.

The interfaces we are familiar with are disappearing

The disappearance of interfaces I am talking about refers to the familiar UIs, such as Apps, websites, and buttons, all fading away. You will still see photos and maps; I'm not that extreme. The way we interact with AI and computers will become like interacting with people: just talking or typing, and things get done.

For more than 40 years, we have been learning the language of computers: programming, menus, forms, drop-down boxes, hamburger icons, and "Are you sure?" pop-ups. We became so accustomed to it that we forgot we were always translating. Now, computers are starting to speak our language.

This is also changing the definition of "operating system." Today's operating systems are mainly App launchers; whoever controls the launcher sets all downstream rules. When people no longer launch Apps, the operating system becomes the agent you converse with, and it currently has no master.

The focal point of attention will also change. For the past 20 years, people's attention has been centered around information flows and Apps. Next, it will shift to agents: personal agents working for you and corporate agents working for companies, which will increasingly make decisions on our behalf.

Whoever can control the winning agents will become the most powerful company of this era. I believe this is no longer debatable. Meta's Muse is doing exceptionally well and is just getting started. Google will likely join in, and cutting-edge modeling companies will not be absent; any company that can create a strong economic network effect among stakeholders will compete for this. Whether these agents succeed or whether these companies can make money all hinges on business, because reasoning, virtual machines, and underlying model training all cost money. Moreover, an agent that cannot make purchases cannot be considered very capable.

The payment system is built on buttons

At PayPal, the product is the button. During my years at PayPal, we spent many years optimizing button experiences, cutting steps and saving seconds to help more people complete checkout. Now, I find myself writing its obituary.

Today, almost all financial products are built around a screen: bank Apps, checkout pages, card number forms, and verification codes sent to your phone. The entire risk framework of payments — authentication, fraud, disputes, liability — assumes that there are people involved at many stages. Remove that person, and the assumptions collapse entirely.

When your agent buys something from an enterprise's agent, there is no checkout page and no one to send the verification code. The question becomes whether you authorize this agent to make such payments. When this system was designed, no one needed to ask that question. Card networks and some new protocols are already looking for answers, and everything underneath must keep up.

The interface for digital funds will undergo profound changes because money will no longer flow through UIs.

Stablecoins are no longer the answer

This statement coming from the initiator of Libra may sound strange: I believe stablecoins are no longer a standalone entity.

This assertion is useful. It allows legislators to grasp something to regulate, and as a result, stablecoins have been recognized as legal digital assets. The European MiCA has already been fully implemented, and the U.S. GENIUS Act has also become law and will take effect in January. This battle has been fought for many years, and it has essentially been won. Stablecoins equal acceptable digital assets.

However, stablecoins themselves do not solve most real-world problems. Holding digital dollars is indeed useful. But using them to do things is often challenging. To be widely useful, stablecoins must combine with these components:

• A fast settlement network that enables value to arrive in seconds, anytime, at near-zero costs;
• An account system, also known as embedded wallets, which allows money to reside within the products or agents used;
• A fiat payment system, both domestic and cross-border, allowing you to move between all existing systems, essentially absorbing the liquidity of the old track into the new one;
• A card network that enables balances to be spent anywhere cards are accepted, and cards are almost ubiquitous.

Agent-native payment layer

Combining all these elements — stablecoin settlement, accounts, fiat connectivity, and cards — but without a frontend, only a layer of simple APIs, which agents can use to transfer value whenever needed, whether buying or selling for individuals or companies. This is the "agent-native payment layer."

The APIs may be simple, but what lies behind them is not. You need licenses, banking partners, global issuing and acquiring, real-time connections to local payment systems in over 100 countries, liquidity, compliance, and everything operating simultaneously. This is the toughest and most complicated work in payments, and it is why very few companies can succeed at it.

You tell the agent to fix the fence before the weekend. It finds a contractor, most likely an agent of the contractor, agrees on a price, pays a deposit, and waits for the photos of the completed fence to come in before paying the final amount. You never touched an App or went through the payment process. You simply stated what you wanted.

For example, consider a company whose agent manages a sum of money and a set of rules: paying 400 contractors in 30 countries every Friday, increasing advertising spending when data is appropriate, paying cloud bills, and refunding customers for damaged orders. Each transaction lands at different endpoints: a bank account in this country, a mobile wallet in that country, a card, a stablecoin address. The agent should not be concerned with these details. It should just say, "pay this person this amount," and then it arrives in seconds, anywhere, around the clock.

The agent has its own inbox and some have their own phone numbers, but the current AI economy lacks banks, and agents do not have accounts. To enable agents to have accounts, the underlying money must:

• Be digital;
• Allow real-time settlement, 24/7, regardless of bank hours;
• Reach every kind of endpoint on the planet;
• Be affordable, where agents pay mere cents for an API call, and fees should not be higher than the payment itself;
• Allow people to safely grant agents real spending power, with limits that can also be revoked instantly;
• Clearly state who bears the loss if someone tricks the agent into making an incorrect payment.

Who will build the agent-native payment layer

The payment infrastructure will spawn a number of giants, with some reaching valuations in the hundreds of billions. At this momentum, in the eyes of those winning the agent race, this is just a small piece.

The difference comes from how both sides make money. At the payment layer, the earnings are from 1 to 300 basis points on the volume, which is 0.01% to 3%. However, an agent that decides what the buyer buys and from whom is far more valuable to the seller than merely transferring funds. Companies operating these agents will receive compensation, much like today's markets and advertising platforms.

This does not mean that payments are unimportant within the agent platforms. On the contrary, their payment teams will have to tackle one of the most challenging tasks in the company: putting together the fragments of licenses, banks, card networks, local payments, and compliance to create a secure, user-friendly experience that can operate in every country and is suitable for both buyers and sellers. The underlying funds layer will become as reliable and critical as the protocols that run the internet.

For those who can successfully build the agent-native payment layer, this is a tremendous opportunity.

This time, it may not be too early

The first wave of agent shopping started slowly; I can already hear people saying I was too early for the third time. I believe money is a significant reason for this. Agents have been borrowing payment flows built for humans, with each purchase ultimately involving card forms and confirmation buttons.

This time, the models are good enough, and the laws are written. What is missing is money that agents can use without a screen. Whoever builds it will create something that all agents depend on. If I have indeed gotten ahead of the curve again, I will accept that as well.