Google on Tuesday (Sept. 16) introduced the Agent Payments Protocol, a bid to standardize how artificial intelligence (AI) agents authorize and complete purchases, while giving banks and merchants a clean audit trail. The company framed AP2 as a foundation for AI-driven commerce, meant to ensure that automated transactions carry the same accountability as human clicks.
Why a Payments Protocol for Agents
AI agents are shifting from chat to checkout, challenging a core assumption of eCommerce: that a human presses “buy.” AP2 proposes a consistent way to prove that an agent’s request reflects a user’s instruction and to identify who approved the transaction. Google said more than 60 firms are supporting the launch, including American Express, Mastercard, PayPal, Etsy, Revolut and Coinbase.
The mechanism rests on tamper-evident mandates, cryptographically signed records that travel with each transaction. In a human-present flow, the agent captures an intent mandate when the user requests an item, followed by a cart mandate that approves the final basket. A third payment mandate signals to networks that an AI agent executed the order. Together, these credentials create a non-repudiable trail that issuers and merchants can use in fraud investigations and dispute resolution (AP2 GitHub Spec).
How AP2 Fits Into Real Transactions
Google illustrated scenarios that highlight AP2’s role. A shopper might direct an agent to monitor a jacket in a specific color and buy if the price falls below a threshold. A traveler can set a budget and let the agent coordinate flights and hotels, then approve both bookings at once. For unattended purchases, such as buying concert tickets at release, a detailed intent mandate sets rules upfront, allowing the agent to generate a cart mandate later without a human present.
While the initial version focuses on card rails, AP2 was built to be payment-agnostic. Google released an extension linking the protocol to the x402 crypto standard, developed with Coinbase, MetaMask and the Ethereum Foundation, so agents can transact directly from crypto wallets. Coinbase published an explainer on the integration, positioning it as a way to unify AI commerce with decentralized payments.
That roadmap signals where Google expects AI-driven payments to go. Future iterations of AP2 are slated to include real-time bank transfers and stablecoin support, bridging the gap between traditional card networks and instant push-payment systems.
Advertisement: Scroll to Continue
The Broader Ecosystem: Mastercard, Visa, PayPal and Perplexity
Google is not alone in preparing for AI commerce. Mastercard is rolling out its Agent Pay service globally, part of a toolkit that helps merchants and developers create agent-ready checkout flows. Visa, meanwhile, has opened its MCP server and released a no-code Agent Acceptance Toolkit so developers can plug agents into Visa APIs and test real transactions.
Platforms are experimenting as well. In May, PayPal partnered with Perplexity to power its “Buy With Pro” service, embedding agent-led purchases inside its browser. The move signaled that agentic checkout is no longer theoretical; it is routing real orders. These examples show why a standard like AP2 matters: it gives merchants and issuers a uniform way to handle agent-driven requests rather than negotiating bespoke integrations.
Risk, Conversion and the Adoption Test
For risk teams, AP2’s biggest contribution is accountability. Each mandate documents what the user allowed, what the merchant promised, and what the network processed. That evidence can cut down on false disputes and help issuers make better approve-or-decline decisions. Agentic commerce will still need strong identity checks, but mandates could reduce fraud without adding friction.
Adoption will hinge on two tests. First, developer uptake. The AP2 GitHub repository contains full specifications, sample flows and reference code, lowering the barrier for pilots. Second, merchant economics. If AP2 reduces fraud losses and false declines, or lifts conversion by enabling richer agent-to-agent offers, merchants will adopt. If it adds steps without visible upside, they may wait.
The final question is regulatory and network acceptance. If mandates are recognized as sufficient proof of authority, AI-driven transactions can scale through existing payment pipes. If not, each platform may need custom agreements. For now, AP2 provides a foundation: it shifts proof of purchase from inference to verifiable, signed intent, giving the ecosystem a common starting point to build on.
