Explainer: How Google’s AP2 Protocol Might Transform the Web 

Across 2025, we have heard a lot about AI agents. In fact, the idea of agents has been touted as the next big thing in technology. An AI agent is effectively a tool that can work independently across the internet, completing tasks based on your instructions. For instance, you might ask an AI agent to search for a French restaurant in New York City, and book a table for two people on a specific date. Technically, the agent can complete that task, meaning you don’t have to do it yourself. 

However, it’s worth noting that roll-out has been relatively muted. There are various reasons for this, including limitations to the agents themselves, as well as other barriers. Yet, Google recently made a significant announcement, which will remove one of those barriers, potentially transforming the web as we know it. 

AP2 Could Change the Web

On September 15th, Google announced Agent Payments Protocol (AP2), creating a new payments standard for the internet, which effectively allows AI agents to make financial transactions. To add weight to the announcement, Google has partnered with Mastercard, PayPal, Coinbase (meaning crypto and stablecoins will be part of the standard), Revolut, Etsy, WorldPay, and many more. While it’s a Google announcement on paper, it’s open-source and about the broader internet. 

So, why is this a big deal? Well, no such infrastructure existed before, at least across the web. Previously, you would be able to use, for example, Expedia’s AI bot Romie and ask it to, say, book a hotel room based on your instructions. But Romie can only work under strict orders on Expedia’s platform. Now with AP2, it is opened up to the wider internet.  

AP2 works as follows: You ask the AI agent something like, “Find me some Nike sneakers under $100. The sneakers must be mostly white.” This creates an Intent Mandate, and the AI bot will presumably return with some options. If you tell the AI agent you are happy to proceed, it creates a Cart Mandate, and the bot will go off and purchase those sneakers.

You can be more direct to set up the Cart Mandate at once. For example, you might say, “Buy two Taylor Swift tickets as soon as they go on sale. Don’t spend over $250 per ticket.” That direct command would set up a Cart Mandate instantly. These Mandates are contracts, recorded on a blockchain-type ledger with cryptography. Thus, there is an evidence trail in case you argue that the AI agent is making mistakes. 

How Widespread Will Adoption Be?

It’s easy to see how this is important although there are several questions. First, how widespread will it be? Could you, for example, have an AI agent trained on card games visiting a casino to play online blackjack, placing bets without your input? Could AI agents spam platforms like Vinted or eBay, making it impossible for ‘real’ users to get a deal? There will, of course, be guardrails in place, but it will be intriguing to see how quickly adoption rolls out. As mentioned, the partners are a who’s who of fintech and eCommerce, so this could permeate the internet pretty fast. 

Overall, though, the adoption will be correlated to the performance and reliability of the AI agents. ChatGPT, Claude, Gemini, Llama, and so on are hugely impressive, but AI is certainly not perfect yet. You’ve all probably seen the “how many Rs in strawberry” style test that LLM models tend to fail. Would you really trust AI to book a flight for you as it stands today? Still, as with everything in technology, the rollout might have a few bumps on the road before it is perfected. But Google and its partners are laying out the framework for an agentic web economy. It could transform how the internet works and how we work with it.