Agentic Pay Systems: Google’s Agent Payments Protocol
Google’s Agents Payments Protocol (AP2) sets out to introduce an agentic payment tool compatible with most platforms.
Merchants, payment providers and consumers can safely enable transactions across all types of payment methods.
The collaboration has named partners such as Adyen, American Express, Ant International, Coinbase, Etsy, Forter, Intuit, JCB, Mastercard, Mysten Labs, Paypal, Revolut, Salesforce, ServiceNow, UnionPay International and Worldpay. The list is non-exhaustive.
Agentic Pay introductions are on the rise
In recent months, the introduction of agentic AI in the payments space has risen considerably.
In a report from PYMENTS, agentic AI was predicted to handle up to 20% of e-commerce tasks in 2025.
Payhawk recently introduced four agents to assist in purchasing travel tickets, procurement, financial controller and, of course, payments.
The Agent claims to deflect 40% of helpdesk inquiries by advising on failed transactions, blocked cards and funding issues.
Similarly, Mastercard also introduced agentic payment options with its Agent Pay.
Agent Pay allows AI agents to complete transactions on behalf of the user. The programme uses tokenisation to facilitate transactions.
At the time, Mastercard stated it would partner with IBM for more B2B cases, and Microsoft for scaling agentic commerce.
What about Google’s Agentic Protocol is different?
In what Tom Mason, Principle Specialised Engineer, Applied AI at the office for the CTO at Google Cloud calls a “foundational step”, the introduction of AP2 aims to provide a chance for open communications between agents and merchants.
Tom continues in a statement on LinkedIn: “This initiative introduces a standardised set of API specifications designed to create a secure and interoperable communication layer between LLM-based agents and payment processors.”
“The core objective is to enable AI agents to execute transactional flows on behalf of users, moving beyond informational queries to direct commercial actions.”
The protocol also enables various different types of payment methods, such as real-time bank transfers, card transactions and stablecoin.
Introducing an open, shared protocol allows merchants to facilitate secure and scalable experiences with the transparency needed to properly assess and mitigate risk.
Tom adds: “By establishing a common protocol, AP2 aims to solve the integration challenge and provide a scalable framework for building robust agentic commerce applications. “
Integration is a key part of the service, and Google state that the AP2 can be used as an extension of its Model Context Protocol and Agent2Agent protocol.
“This is a foundational step in architecting an ecosystem where autonomous agents can securely interact with the existing financial infrastructure.” says Tom.
Google claims that the introduction of AP2 can unlock new commercial experiences, such as smarter shopping, personalised offers and coordinated tasks.
How Google’s AP2 works
In order to drive home the concept of trust in AI, Google’s AP2 will use Mandates.
Mandates are verifiable proof of a user’s instructions. They act as cryptographically-signed contracts, designed to be tamper-proof.
Split into two sections, the mandates act in accordance with human presence.
When humans are present, and request an item, the instruction is captured by the Intent Mandate. Google claim this is the auditable part of the entire interaction process.
The Cart Mandate creates an unchangeable, secure record of the exact items and price, locking in the exact amount to ensure that the amount seen is the amount that is paid for.
If the user is pleased with the accuracy of the item in the cart and approves the action, this is classed as a Cart Mandate.
The absence of a human means a delegated task. If a user mentions a time-specific task, such as the acquisition of concert tickets as soon as they are on sale, they are prompted with signing an Intent Mandate up front.
The Intent Mandate lists requirements for the purchase, eg budget, timing and other conditions.
With the Intent Mandate, the agent now has undisputable permission to carry out the research.
Final permission is still needed for the automatically-generated Cart Mandate, which is presented to the user once the conditions laid out in the Intent Mandate have been met.
