Google has announced a new open protocol, the Agent Payments Protocol (AP2), for AI agents that can make purchases and decisions on behalf of users. It is supported by 60 merchants and financial institutions.
Google has announced a new open protocol, the Agent Payments Protocol (AP2), for AI agents that can make purchases and decisions on behalf of users. It is supported by 60 merchants and financial institutions.
The system is designed to interface between AI platforms, payment systems, and vendors, providing the ability to track every transaction, TechCrunch reports. The full specification for AP2 has been published on GitHub .
«We are committed to developing this protocol through an open collaborative process, including through standards bodies, and invite the entire payments and technology community to build this future with us,» said Stavan Parikh and Rao Surapaneni, vice presidents of Google and Google Cloud, respectively.
The protocol is designed for a future in which AI agents routinely purchase goods on behalf of customers and engage in complex interactions with AI agents at retailers. One example in Google’s post involves a chatbot user asking their agent to buy a bike for a ride, which results in a spontaneous promotional offer from a bike shop agent.
In another example, a user requests a weekend getaway and accommodation, specifying only dates, location, and budget.
«The agent can interact with both airline and hotel agents, as well as online travel agencies and booking platforms, and when he finds a combination that fits the budget, he can simultaneously complete both cryptographically signed reservations,» Google explains the idea.
Implementing these kinds of transactions is both technically and socially complex. AP2 requires agents to register two separate assertions before a purchase can be made: first, an «intent mandate» (essentially an AI message saying, «I’m looking for a polka dot tie»), which allows the agent to search for a specific item and negotiate with sellers; then, a «cart mandate,» which provides final confirmation of the purchase once the specific item has been found.
The protocol also includes provisions for fully automated purchases, where an agent can automatically generate a cart authorization after finding an item. Such circumstances require a more granular authorization specifying price limits, terms, and other interaction rules. The goal is to create a transparent chain of events that can be reviewed in the event of fraud.
Previously, cybersecurity experts at startup Guardio prepared a study called «Scamlexity» that looked at the vulnerability of browsers with AI agents, such as Opera Neon and Perplexity Comet. They found that AI is bad at distinguishing between phishing and other types of scams that are filled with the Internet.
