Google’s Agent Payments Protocol provides a framework for AI to automate consumer purchases
Google LLC is giving artificial intelligence agents even more autonomy with the launch of a new protocol that makes it possible for them to make payments on behalf of their users.
The Agent Payments Protocol, known as AP2, is backed by more than 60 merchants and financial institutions, and is aimed at AI agents that can shop and make buying decisions at the request of humans. It’s designed to be interoperable across AI platforms, payment systems and vendors and fully traceable so every transaction can be audited.
In a blog post, Google Vice Presidents Stavan Parikh and Rao Surapaneni stressed that the company is committed to full openness and transparency, and sees the AP2 protocol evolving collaboratively over time. It hopes to attract the involvement of the “entire payments and technology community”, as well as standards bodies to ensure agentic payments are regulatory compliant. Google has open-sourced the protocol and made the full specification available on GitHub in line with this stance.
Parikh and Surapaneni said AP2 is designed for a future where humans routinely rely on AI agents to shop for products and services on their behalf, and envisions a world where retailers also rely on agents.
AI agents have already come a long way. With the rise of browser-based agents that can surf the web and navigate through menus and other elements like people do, it’s already possible for them to browse the internet and find cheap flights or hotels, and in the case of businesses, it’s possible for AI agents to compare different suppliers, for example. However, today’s agents still stop short of completing purchases or transactions, and instead ask users to take this final step themselves. It’s this final step that Google wants to automate with AP2.
In one example, Parikh and Surapaneni said this can provide a lot of convenience and allow shoppers to take advantage of the best deals available. For instance, they gave the example of a chatbot user who asks it to shop for a bicycle, which triggers a spontaneous, time-sensitive offer from a bicycle shop’s own AI agent. Alternatively, someone might ask their agent to book travel and lodgings, providing it with only the dates, location and their budget. In this case, “the agent can then interact with both airline and hotel agents, as well as online travel agencies and booking platforms, and once it finds a combination that fits the budget, it can execute both cryptographically-signed bookings simultaneously.”
Automating such transactions are very complex, not only on the technology side, but also socially. This is why AP2 always requires AI agents to obtain two separate approvals from users before any purchase can be made. First, it needs to obtain an “intent mandate”, which is where the user tells them they are looking for a new bicycle and provides an appropriate budget and specifications, enabling it to go search for a specific product and negotiate a price. Then, it also requires a “cart mandate”, which is when the user gives the final approval for it to buy a specific item once it has been identified.
There is a provision in the AP2 protocol for entirely automated purchases too, where the AI agent can be permitted to automatically generate the cart mandate once the specific item has been found. In such scenarios, the agent must obtain a more detailed intent mandate, which specifies price limits, the timing of the purchase and other rules. The goal is to ensure that the user retains control of their wallet and their finances and that the agents don’t start making purchases willy-nilly, and also to maintain a fully auditable trail of each transaction, so it can be examined in case it’s suspicious.
Google has obtained a lot of industry support, with multiple payment processors getting onboard with the idea of AP2 and agentic payments. Mastercard Inc. Chief Digital Officer Pablo Fourez said the protocol can help to shape the future of agentic commerce, but stressed there’s still a lot of work to be done. “These efforts include critical work with standards bodies such as the FIDO Alliance, where we are advancing verifiable credentials to capture and secure consumers’ intent in this dynamic new context,” he explained. “we’re playing an essential role in securing the payments ecosystem – ensuring that trust and safety remain at the core of every transaction.”
Google is also collaborating with cryptocurrency firms such as Coinbase Global Inc., the Ethereum Foundation and the crypto wallet provider MetaMask on an extension called x402 that will enable agents to utilize digital stablecoins for purchases.
“Blockchains are the natural payment layer for agents, and Ethereum will be the backbone of this,” said MetaMask’s AI lead Marco de Rossi. “With AP2 and x402, MetaMask will deliver maximum interoperability for developers and will enable users to pay agents with full composability and choice while retaining the security and control of true self-custody”.
Google isn’t alone in trying to enable AI agents to make payments. Some of its rivals in the AI and payments industries are working on similar efforts. For instance, the AI search startup Perplexity Inc. offers a “Buy With Pro” service in its agentic AI browser, and Stripe Inc. has developed its own software to enable AI agents to make purchases using its platform.