When Machines Pay: AI Agents Are Set to Spend Trillions

Artificial intelligence assistants will soon be able to make purchases and pay for services autonomously. Making this possible will require two different forms of money and clearly defined limits.

Imagine telling your AI assistant: “I need new running shoes. I’m running a marathon in April, and my budget is €180.”

A few minutes later, an order confirmation arrives in your inbox. No comparison websites, no shopping carts, no passwords. The agent has researched the options, compared products, placed the order, and completed the payment, all on your behalf.

The Agent Goes Shopping

This is no longer a distant vision of the future. Major AI providers are transforming their language models into what are known as AI agents: software systems that do more than answer questions. They can independently carry out tasks.

OpenAI and Stripe have introduced a joint protocol enabling purchases directly within a chat interface. Visa and Mastercard have launched their own agent-based payment initiatives, called “Intelligent Commerce” and “Agent Pay,” respectively. Google has proposed a standard for authorizing payments made by AI agents.

All of these initiatives are addressing the same fundamental question: How can we allow a machine to spend money on someone’s behalf without things going wrong?

The economic stakes are enormous. Consulting firm McKinsey estimates that, by 2030, AI agents could facilitate up to $1 trillion in retail sales in the United States alone, and between $3 trillion to $5 trillion globally.

Buying running shoes is still a relatively straightforward use case. Things become more interesting when an agent starts spending money independently as part of the purchasing process.

Before selecting a particular model, for example, the agent might purchase the latest running shoe review from a specialist magazine. Rather than paying for an annual subscription, it buys access to a single article for just a few cents. It analyzes the results, matches them against the user’s foot shape, and only then places the order.

The magazine has gained a paying reader who would never have become a subscriber.

This model extends far beyond publishing. An AI agent might pay for a weather forecast, a dataset, a translation, computing resources, or a credit report. It might even delegate part of its work to another AI agent.

In this emerging environment, software systems will increasingly transact directly with one another - at enormous scale and often for extremely small amounts.

Two Forms of Money

A closer look reveals three fundamentally different types of agentic payments.

In the first scenario, an AI agent purchases goods or services from a merchant, much as a human would: running shoes, a flight, or the weekly groceries. These transactions typically involve amounts of €50, €200, or €2,000. They also involve delivery obligations and the possibility that something might go wrong.

In the second scenario, the agent pays for a service that is delivered immediately, such as access to a magazine article, a data request, or a minute of computing time. These payments may amount to just a few cents and occur thousands of times a day.

In the third scenario, one AI agent pays another. A research agent, for example, might commission a translation agent to complete part of its assignment. There is no human on the receiving end, and no physical product changes hands. What is being exchanged is information and work.

These three scenarios call for different payment instruments.

For purchases from merchants, card payments remain the appropriate solution. If the running shoes never arrive, the wrong size is delivered, or the merchant goes bankrupt, customers need a way to recover their money.

Over decades, card payment networks have developed the infrastructure for chargebacks, buyer protection, and clear liability rules.

For micropayments between machines, however, cards are poorly suited.

Card payments typically involve a fixed transaction fee plus a percentage-based charge. For an article costing only a few cents, this makes the transaction economically unviable. Moreover, while a card payment may be authorized within seconds, the merchant often receives the funds only after a delay.

This is where stablecoins offer a distinct advantage.

Stablecoins are digital forms of money issued on a blockchain and pegged one-to-one to a currency such as the euro or the US dollar.

They can be divided into fractions of a cent and transferred across the world in seconds, without relying on correspondent banks. On low-cost blockchains, a transfer can cost just a fraction of a cent, with the recipient receiving the funds almost immediately.

In Europe, the Markets in Crypto-Assets Regulation (MiCAR) provides a regulatory framework for stablecoins, treating relevant fiat-referenced tokens as electronic money and subjecting their issuers to regulatory supervision.

This regulatory foundation is what makes stablecoins particularly attractive to publishers, data providers, and software companies.

The objective is not cryptocurrency speculation. It is to make regulated euro-denominated money available in programmable form.

The future of payments, therefore, is not a choice between cards and stablecoins. It is about using both, depending on the use case.

Cards will remain essential wherever consumers need protection. Stablecoins will come into their own wherever machines need to settle transactions directly with one another.

AI Agents Need Guardrails

One uncomfortable question remains: Do we really want to give machines the ability to spend money?

Language models make mistakes and can be manipulated. Ask an agent to find “the best running shoes,” and it may end up choosing the most expensive pair.

An agent purchasing an article for a few cents could also become trapped in a loop and repeat the transaction tens of thousands of times.

Yet AI agents are unlikely to disappear. The answer is not to prevent their use, but to establish clear boundaries.

First, spending limits must be enforced independently of the agent itself.

An AI agent tasked with purchasing specialist literature, for example, could be authorized to spend no more than €5 per month - without the ability to exceed that limit by even a single cent.

The proposed European Payment Services Regulation (PSR) is expected to require payment service providers to offer such spending limits.

Second, agents need clearly defined mandates.

These mandates take the form of cryptographically signed authorizations specifying what the user intends to purchase and how much they are willing to pay.

Google’s Agent Payments Protocol (AP2) is designed around precisely this concept.

It allows merchants to verify that a legitimate customer has authorized the agent’s actions. At the same time, customers can demonstrate exactly what they authorized if a dispute arises.

Third, common rules and standards are essential.

How does an AI agent prove its identity? Who is liable when it makes a mistake?

With MiCAR, the ongoing reform of European payments legislation, and the European Digital Identity Wallet, Europe is well positioned to help shape these rules.

The money that AI agents use must be divisible into fractions of a cent. But it must also be subject to a defined purpose, a spending limit, and an expiration date.

Anyone selling goods, content, or data should prepare for a future in which their next customer may not be human.

It may be a machine.

Originally published in German in Frankfurter Allgemeine Zeitung (F.A.Z.) on October 28, 2026.

English translation prepared from the original German article. All rights to the original article remain with Frankfurter Allgemeine Zeitung GmbH.

About the Authors

Dr. Alexander Bechtel

Alexander Bechtel is Global Head of Digital Product Strategy at DWS and a member of AllUnity’s Supervisory Board. Previously, he held various positions in money markets and digital currencies at the European Central Bank and Deutsche Bank. He earned his doctorate from the University of St. Gallen and was a visiting researcher at Stanford University.

Peter Grosskopf

Peter Grosskopf is Chief Technology Officer at AllUnity, a BaFin-regulated electronic money institution issuing MiCAR-compliant stablecoins denominated in currencies including the euro, Swiss franc, and US dollar.

He also leads the development of an agentic paywall that enables AI agents to pay for content and data on a per-request basis.

Previously, he co-founded Solarisbank.