AWS re:Invent 2025 - Exploring dual pathways to agentic commerce with AWS and Stripe (AIM114)

🦄 Making great presentations more accessible.
This project aims to enhances multilingual accessibility and discoverability while maintaining the integrity of original content. Detailed transcriptions and keyframes preserve the nuances and technical insights that make each session compelling.

Overview

📖 AWS re:Invent 2025 - Exploring dual pathways to agentic commerce with AWS and Stripe (AIM114)

In this video, Stripe presents how AWS and Stripe enable secure agentic commerce through two modes: business-to-agent (B2A) and agent-to-consumer (A2C). The speaker introduces Stripe's Agentic Commerce Protocol (ACP), an open standard co-authored with OpenAI that makes human-readable checkouts machine-readable for AI agents. Key innovation includes shared payment tokens with built-in guardrails to prevent fraud and budget overruns. Live on Etsy with Walmart and Shopify implementations underway, customers can now purchase directly within ChatGPT. The presentation also covers Stripe's payments foundation model trained on billions of transactions, achieving 95% card testing detection and driving $1.4 trillion in annual payment volume. For on-site experiences, Stripe offers a reference architecture built with AWS Bedrock, enabling brands like Amazon Rufus and Walmart Sparky to deploy AI shopping advisors while maintaining control over customer discovery.

; This article is entirely auto-generated while preserving the original presentation content as much as possible. Please note that there may be typos or inaccuracies.

Main Part

Stripe's Financial Infrastructure and AI-Powered Payments Foundation Model

All right, thank you so much. I'm going to begin with a statement: agents are the new storefront. Discovery is shifting to generative AI-assisted agents. In 20 minutes, I'll show you two modes of agentic commerce and how AWS plus Stripe are helping to make payments instant, secure, and trusted in both of those modes. Let's jump right in with a lightning talk about Stripe.

Stripe is a financial platform. Think of us as the financial infrastructure of the internet. We do much more than just payments. Consider us to be programmatic money movement. We like to think of ourselves as the AWS of payments, but we do significantly more than just payments. As you can see here, we accept money globally in many different countries. We can automate revenue with recurring payments through our billing product. We can automate tax calculations, reporting, and data. We power many of the world's marketplaces, from DoorDash to Instacart to Lyft and many others. We also have extensive banking as a service and money management offerings such as Stripe Issuing and Capital. We're also market-leading in stablecoins. Think of us again as programmatic money movement and financial infrastructure to solve use cases across the board. We're very horizontal in how we approach things, and we're very proud of our partnership with AWS.

We have competencies in everything you see here: retail, media and entertainment, travel, hospitality, insurance, as well as a focused vertical in ISVs. This partnership really comes to fruition when you think about the amount of uptime and availability that we drive through both AWS's fantastic cloud infrastructure and our talented engineering team at Stripe. We drive five nines of availability consistently. We just completed Black Friday Cyber Monday, moved close to 40 billion dollars, and had six to seven nines of availability. We're very proud that the combination of AWS plus Stripe serves 4 million plus customers in a very scalable and reliable way.

These types of solutions drive real revenue for customers. By adding Stripe's optimized checkout suite, we can generate on average close to 12 percent of top-line revenue uplift. That's a huge number when you really think about it. We can modernize and optimize and drive efficiency through 30 percent faster launches through Stripe's innovation in API layers and developer SDKs and toolkits. We have some of the best documentation in the industry and we have a unified API, so we're very friendly to driving efficiency through innovation within our developer community. Finally, we reduce risk through many fraud data models that I'll discuss, powered by the combination of Stripe engineering and AWS infrastructure.

Earlier this year, we launched the world's first payments foundation model. It's trained on literally tens of billions of key signals. Think of it as using the breadth of the global economy to train an AI model. If you know anything about vector databases or vector spacing, it uses high-dimensional vector space to quantify things like different clusters and what these clusters mean to each other. Payments that share similarities, for example, that come from the same country or from the same card or that are positioned closer geographically from the same bank, we can cluster this information and use it to be very predictive in how we prevent fraud and bring other solutions to market.

Payments from the same email address, for example, are very important data points to build into vector spaces and to cluster. By using these solutions, we've driven 1.4 trillion dollars of annual payment volume last year in 2024. Ninety-two percent of cards using this payment foundation model have been seen before. Think of it as an AI model that's trained on billions of data points, so we can become very predictive and actually look at a transaction from start to finish. Instead of being reactive to it, we can actually be proactive in eliminating fraud. We've had millions of companies benefit from these capabilities.

The payments foundation model drives the entire end-to-end experience from checkout to authentication to fraud prevention with Radar, card testing, refunds, and disputes. The payments foundation model is key to delivering value end-to-end throughout the entire payments lifecycle.

Looking at some real-world statistics, 95% of card testing and attacks have been detected in real time using this payments foundation model, which represents a 21% improvement over prior methods. By using this foundation model, we're being proactive rather than reactive in preventing fraud. Smart disputes is another powerful capability where, if you need to dispute a charge, you would normally have to submit 10 to 15 or 20 pages of documentation. Smart disputes uses generative AI to automatically generate this paperwork and submit it end-to-end. DoorDash implemented it and achieved a 10% reduction in chargeback costs, which for a company of that size represents a very significant amount of revenue. Vimeo and Squarespace saw a 13% increase in chargebacks recovered. Automation and AI can be applied to every aspect of the end-to-end payments lifecycle, including fighting disputes.

To wrap up, the Optimized Checkout Suite boosts revenue by creating frictionless checkout experiences for your customers. We also have a one-click digital wallet called Stripe Link. You can visit link.com and sign up for an account. Think of this as a phone-agnostic digital wallet. It's not Apple Pay or Google Pay, but it competes with those from the standpoint of what it delivers, and it's available to you no matter what device you use or what browser you use. Stripe Link is a one-click wallet that powers frictionless checkout experiences for us.

Business-to-Agent Commerce: Agentic Commerce Protocol and Shared Payment Tokens

Let's jump into the topic at hand: agentic commerce. I promised we'd talk about two modes of agentic commerce, but there are actually three models that exist today. There's business-to-agent, otherwise known as B2A, agent-to-consumer, A2C, and then A2A. I'll briefly touch on A2A because it's more of a future model. Think of it as a world where an autonomous agent from the buyer's perspective negotiates with an autonomous agent from the seller's perspective, and there's no human involved at all in the checkout. You're basically having an agent shop on your behalf while the seller has an agent representing them. You're using policy engines and logic to come to a conclusion. That's a future use case we're working on, but we have already launched real-life solutions today in the first two use cases.

I'll start with business-to-agent, which involves participating in agentic channels. Quick show of hands: who here has used any AI surface to try and buy something or at least get a recommendation to shop for something? So our stats show one out of four. That was more than half, just roughly scanning the audience. We have a very sophisticated audience here that's used AI for shopping. The key thing we wanted to talk about is that AI is the new search, as I mentioned earlier, and AI is being the new storefront. We also have AI-powered advisors, which is the agent-to-consumer use case we'll discuss as well.

With business-to-agent, what we're seeing is an emerging ecosystem between buyers, AI agents or orchestrators, and merchants and businesses. Businesses want their products to be discoverable in these AI chat surfaces, and buyers want a frictionless, very safe, secure checkout experience. Payments is the one layer of trust that cannot be compromised in agentic commerce. Think about giving an agent your bank account information or your credit card and saying go out and shop for something. That may not be the best idea. We want to build something that's fast, secure, reliable, and fraud-proof. This is the ecosystem we're working with. We've partnered with all the AI companies you see there, and we're laser-focused on making sure that this new ecosystem and this new way of purchasing is safe, reliable, and secure.

So what have we launched? We launched this in late September at our US New York tour. It's called Agentic Commerce Protocol, ACP for short. It's an open standard, Apache 2 licensed, community-driven, and can be used by any AI surface. We co-authored it with OpenAI, but it is available to be used on Perplexity, Gemini, Microsoft Copilot, X, Meta, and all the others that you saw on the previous slide.

The problem we were trying to solve by giving an agent a virtualized debit card to complete a purchase is very difficult. In the past, we tried using browser automation and realized that agents are not very good at figuring out human-readable checkouts. Agents are very probabilistic, so you get different outcomes every time. The same agent could go to the same checkout and figure it out in five seconds or take ten minutes to figure it out.

We launched an experience similar to that with Perplexity, learned a lot, evolved, and came back to the drawing board. We realized what we needed to do was make human-readable checkouts machine-readable. We needed to create an abstraction layer and bring that human checkout to the agent in a way that the agent can be successful. Agents are really good at deterministic outcomes. Agents can hit an API and make a consistent API call every time.

That's the thought model around what we were trying to solve. What we brought to market is a protocol that makes human-readable checkouts machine-readable. If you want to be discoverable as a business in these new machine-readable checkouts, you have to make your product catalog machine-readable and feed your product catalog straight to the agents. We had to publish protocols, product feed specs, and delegated payment specs. There's a lot of documentation that I'll share with you at the end of the presentation.

I get a lot of questions about the difference between AP2, which Google launched two weeks prior to ACP. Basically, AP2 is a principles framework. There's a lot of really good thought leadership from Google and many partners that went into developing AP2. Stripe participated somewhat at the beginning of those conversations, but Stripe takes a different approach. We like to build things and get them launched out in the market.

We like to innovate with design partners and launch something live, learn, iterate, and evolve from that, and then let the protocol mature as it's being delivered out in the wild. Two weeks after AP2 launched, we launched ACP in co-authorship with OpenAI. Part of the technology we had to solve for was shared payment tokens, which I'll talk about coming up. We are live today on Etsy, and I'll show you an example. We're implementing with Walmart, Shopify, and many other large retailers. This is a live solution that we're talking about. It's not a principles framework. It is live and something that we launched.

What is a shared payment token? Think of an example of giving an AI agent your credit card and saying you want to buy a model airplane, but it hallucinates and tries to buy a real airplane and drains your bank account. This is a nightmare scenario that we do not want to embrace in agentic commerce. You have to have something called guardrails, which is basically a budgeted amount that the agent cannot exceed.

Think of a shared payment token as a Russian doll of layers of logic around a payment method that's been tokenized. You've got your payment method on file, Stripe Link or your Visa card, and then you've clicked on OpenAI ChatGPT saying you want to buy this model airplane for five dollars. The shared payment token is instantly minted at that point and gives you a five-dollar budgeted amount that cannot be exceeded. Inside this logic layer is fraud signals. If this card has been used previously in card testing attacks, we'll pass that signal to the merchant. This is a very important part of the protocol. We have to make sure that these extra layers of logic are included as part of the shared payment token. This is live today.

You can take out your phones right now if you want to. You can make history and be one of the world's first agentic shoppers. I've bought many things for my dog as I do these demos on stage, but you can make history. This is something you can basically go into ChatGPT right now and say, buy me some dog treats on Etsy. You'll find some that say "visit," which means they have not adopted ACP, but interestingly, you'll find some that say "buy." So you can click on buy and actually complete that purchase inside the chat session while never leaving that chat experience.

The goal is OpenAI wants you to stay in OpenAI, get these hyper-personalized recommendations. It's going to really get to know your dog over time, like what treats it likes, what size, what costumes it likes for Halloween, all of the above. Getting those hyper-personalized recommendations are what the agentic agents are all about, but we also want to make this a safe, fast, secure payment experience that meets you at the point of inspiration. OpenAI was very clear with us where they said, we do not want the user leaving the site to complete the purchase.

Agent-to-Consumer Mode: On-Site AI Shopping Advisors and Implementation Architecture

We're going to talk about Mode 2. Mode 2 is AI stylist or what we call agent to consumer. Some retail brands do not want to give up all modes of discovery to these external AI agents. Not everybody wants to sell things in ChatGPT, right? Luxury brands, hotels, there's lots of use cases for agent to consumer. Some of these have already been launched and are available out in the wild today. You may have seen Amazon Rufus, which is an on-site AI-powered assistant. So again, you're using generative AI to optimize the experience. You're still getting these hyper-personalized recommendations, but you're just not using ChatGPT or Perplexity as your front-end portal. You're actually going to an on-site branded experience.

Walmart has Sparky, Ralph Lauren just launched Ask Ralph. I've got some nice quotes there as far as where the industry is heading for agent to consumer. The really important thing is Stripe launched this at Stripe Sessions earlier this year in May. We have an on-site shopping advisor. We have a reference architecture that we've built with AWS that allows any business to implement their own Ask Ralph or Sparky or Rufus. We've worked with some major brands using AWS generative AI front end, Bedrock, trained very well on knowledge bases, product catalogs, and guardrails.

A lot of these AI agents that have been released out in the wild are not really good. Rufus is really good. Sparky is good. There's a few good ones. Since we've built with Bedrock, we have built-in guardrails, so you cannot make these agents go off the rails and do something unpredictable. For the architects and developers in the crowd, this is my reference architecture that I built for the AI stylist, and you can see the Stripe interaction is a clear API call. So you don't have to use shared payment tokens if you're making a direct API call from an on-site shopping assistant.

Stripe also has an MCP server. So you can actually call our MCP server directly and generate what we call a payment link to complete the transaction. You can use lambda functions that we can talk about offline. Meet me in the corner after the session if you want to talk about any of this. Basically, you can interact directly with Stripe's API or our MCP server or our Agentic Commerce SDK to build out on-site payments experiences that are powered directly at that level.

With the last minute, I'll open it up for any questions. You can also scan the QR code here. The main thing is you can go directly to agenticcommerce.gov and see the protocol that we published with OpenAI. We have a GitHub repo and we're very encouraged to make this an open community developer-led experience. We've realized that at Stripe we use this talking point a lot. We haven't won yet. We released this protocol. It's been very effective so far, but there's so many more use cases that we want to talk to our merchants about, that we want to work with our developer community to build this out and make it more mature and evolve over time.

With that, I am 20 seconds away from being done. Any 5-second questions? Well, thank you so much. I'll be hanging out after the show if you have any questions.

---

; This article is entirely auto-generated using Amazon Bedrock.