KALYANI KOPPISETTI: Good afternoon, everybody. Looks like there’s some excitement over there. So thank you for all the clapping. We are excited to be here. Before we all get started, I actually wanted to share some personal experience that I experienced a week back. A week back, I was in India. And my mother-in-law said, “Hey, I need a security camera.” So I said, “Let me search for it.” I know how to use the browser. I looked at the browser, looked at it for hours and hours.

After three hours and four hours of browsing, there were so many different types available. And I was like, okay, you know what? I give up. I’m going to ask my husband to take a look at it and buy it because I couldn’t figure it out. How many of you had the same experience where there’s so much data in front of you that you couldn’t decide I want to buy this, and just walked away from the cart?

I can see a lot of hands. And that was me. When I looked at that, then I realized it. I had a friend a couple of years back. I mean, I still have that friend. We’re going to call her Sarah. [laughter] We’re going to call her Sarah. My friend Sarah was amazing at recommending anything. So if I needed something for tutoring, something for cookies or restaurant recommendations, hiking, gifts, chefs, anything I wanted, I would go to her. And she would give it. 

What made her recommendations perfect is not because she’s my friend. It’s because she took who’s asking the question and what is their personal preference into consideration. I’m a vegetarian, so she would not recommend me a place where they don’t have vegetarian options. Right. So she would take that consideration, and that’s a lot of things that we are seeing in AI agents right now. 

Danny and I were chatting the other day. And I said, “Hey, Danny, this is what happened.” And he said, “You know what? I started using AI agents for my shopping.” Danny, do you want to share a bit about it? 

DANNY SMITH: Yeah. Absolutely. So, again, show of hands, who suffers from paralysis by analysis, especially when it comes to big purchases? Yeah, I’m that person. So I would spend, sometimes, entire weekends just shopping for a TV. I needed to know like what are the best buys that are out there? Which has the right features? And I decided to offload that experience to AI. 

Now, I use a tool called Perplexity, which is a great tool—and a great partner of ours. But we’re here to talk about how you can build that on any platform for any merchant. But my experience was basically rooted in four main things that I wanted to focus on. So the first one is I want to be this guy. Okay. [laughter] 

KALYANI KOPPISETTI: Who doesn’t want to be that guy? 

DANNY SMITH: Yeah. He looks relaxed. All the shopping is happening in the background. So all the research time that I was putting in, I realized this is time that I could be spending with friends, family—way better spent not being on a computer going through endless filters, endless different websites, trying to do all the comparisons. The thing is I’m wasting too much time doing this. Secondly, I really needed to see objective comparisons. I didn’t want to hear marketing hype. 

You get fed a lot of information, like Kalyani just said. I needed a side-by-side comparison of exactly what I’m shopping for without any hype. Don’t try and sell me something. I’m shopping. I want to find the right experience, the right product to serve my needs. I don’t need to be sold. I want to basically have an objective experience. Third, no one has endless money. Right? Well, maybe some people do, but I don’t. So I need to have—

KALYANI KOPPISETTI: Neither do I. 

DANNY SMITH: I’m on a budget. And I need to balance price, but I really want to focus on the features that I’m looking for as far as the overall experience. The fourth thing I really wanted to land on is, especially when you’re shopping for things like technology—and in my example, I bought both a TV and a car recently. I bought my first EV. Now I did not do the payment of the car through Perplexity, but I did do the TV. The entire experience was bought end-to-end, never left the AI agent window. 

And the main thing was I wanted to future-proof, make sure that it had all the features— especially when it comes to a car—the safety features that I need. So these are the four things that I was focused on in landing at the overall recommendation. So to Kalyani’s point, we’ve gone from human-based recommendations to now using the power of AI to save time and get the right objective information we need. But what we’re here to do is talk to you about how we can use the partnership of AWS and Stripe to bring this functionality to any business, any platform. 

KALYANI KOPPISETTI: Yeah. That’s right. You know, whether my friend who recommended based on personal preferences or the AI agent that knew Danny’s preferences, this is happening now. We are not talking about the future. It’s something that’s happening now, and we all need to somehow start interacting with those agents. Sometimes we have interacted with those agents without knowing it. Sometimes we know that it’s an agent. 

So there is a slight difference between knowing it’s an agent and not knowing it’s an agent. But before we get started in any of the demos and everything else, let me first properly introduce myself. I’m Kalyani Koppisetti. I’m a partner solutions architect at AWS. Based on my 25 years of experience in the financial services industry, I can personally understand and sympathize with you when you have to balance innovation with security. 

In my current role at AWS, I work with partners like Stripe, build solutions for businesses, and help businesses build cutting-edge technologies but still maintain the trust that they put in the business by the customers. So, Danny? 

DANNY SMITH: Yeah. So my name is Danny Smith. I’m a partner solution architect at Stripe. My background is roughly 20 years in cloud and automation. But my role at Stripe is to partner with our amazing Stripe partner ecosystem—there’s a lot of partners in the audience, especially AWS—and going to market and helping to optimize payment experiences overall. 

KALYANI KOPPISETTI: All right. So let’s look at generative AI trends through the years. Typically, when you see slides like these trends, you see at least 5 years, 10 years. And the thing is we are only showing 3 years because of the tremendous changes that’s been happening. In 2023, businesses were like, “What is generative AI? What do you do? What is this chatbot doing? Is it secure? Can I say this?”

And then they were entertained by it, but they didn’t know how to use it in their businesses. So they started looking at POCs and saying, “Okay, how about this POC? How about that POC?” And then in the year of 2024, what we started seeing is our customers started implementing generative AI in a real way, which is in front of your customers. You see some of the AI assistants there. So we are already seeing some of that happen. 

But now it’s the year 2025. What we are talking about is no longer just generative AI. We are talking about AI agents. We are talking about agentic systems. And just briefly, an AI agent—how we define it at AWS— is something that can take an autonomous decision based on certain rules that it was pretrained on. Right. So to actually make this real, one of the things that we did was to create a shopping experience. I know a lot of you love shopping. Just kidding. A lot of you don’t love shopping, either. 

But whether you love shopping or not, there are things that you always have to purchase whether it is for your house, whether it is your gifts. Whatever it is, you still have to make purchases. So what we created is called Shopping Advisor [AI Shopping Assistant]. It’s a personal shopping agent. These are the AWS services we use. We use Stripe Payments for payment because we have a lot of assistants. We wanted to show how assistants become agents in this particular demo. 

So when we started building this demo, we looked at it from three different angles. One is customers. We always need to start with customers, see what is good for them. Right. Right now, the way the product discovery is, is through umpteen number of filters, and finally you get to it, and you purchase one item, and then go back to the filters and find another item. So we wanted that discovery to be conversational in our experience. We wanted that checkout experience also to be AI-driven and not just show every possible payment option in front of you. 

And then we also looked at our products. You have your products. How do you make your products AI-enabled? How do you make it AI-consumable—creating that knowledge base, creating the dynamic knowledge base? If you want to introduce a new product into this thing, you don’t have to wait six months or you don’t have to prepare the technology to surface that product to your customer. So we looked at that. 

And then finally, the most important thing is employees. Your customers are as happy as your unhappiest employee because they are interacting with your customers. So what we did is we looked at the services that AWS has and said, “How can we amplify the developer experience for employees?” So we used an AWS service called Amazon Q Developer to develop the entire ecosystem. 

And all of this we built on the cloud because that gives us the flexibility to scale up or down based on what we need. Sometimes you may have one customer interacting. Sometimes you may have 1,000 customers interacting. So we wanted to make sure that that balance is created. Again, this is a sample example of a demo that we created. 

DANNY SMITH: Yeah. So Kalyani, let’s say I’m a customer, okay? Mother’s Day is this weekend. I already have paralysis by analysis, as I discussed earlier. My mom, let’s just say—and she would kill me for saying this, but it’s true—she’s a challenging person to shop for. [laughter] So I don’t keep track of the latest fashion trends. I don’t know her style. But she’s my mother, so I know a lot about her. So I can feed some of this information into a model and we can learn over time. But yeah. Walk me through an experience of using your AI Shopping Assistant to buy a Mother’s Day gift for my dear mother. 

KALYANI KOPPISETTI: Okay, Danny. I’m not only going to help you pick something for your mother; I’m also going to get something for myself. 

DANNY SMITH: Oh, okay. It’s on me. Go ahead. Have fun. 

KALYANI KOPPISETTI: So here’s a typical any-company website where you go. If you wanted women’s dresses, you have to go to women’s dresses. If you want handbags, you have to go to handbags. You have to go to jewelry. Umpteen number of choices. And sometimes those choices are important, and that’s what the customer wants. I’m not saying you should replace it. You should totally have it. 

But let’s say you have a customer like Danny, who doesn’t want to spend too much time but wants the crisp information to be in front of them. So we can create something like a shopping advisor. I’ve already logged in as Danny. So it’s asking, “What can I help you with today?” We have AWS services listed over here. I’m saying, “Need gift ideas for Mother’s Day.”

So now that natural language query is being taken. And the Bedrock agent—and I’ll go more into what Bedrock and everything else is at a later point of the conversation—but the AWS service is going back and thinking and coming back. I paused the video over here. It’s not that it’s taking that long, just so you know. I just wanted to finish my sentence there. 

DANNY SMITH: It runs lightning fast. 

KALYANI KOPPISETTI: Yeah. It is fast. So I said, “Need gift ideas for Mother’s Day.” I didn’t provide any other context other than I need a gift. So I’m not sure what it comes back with. Let’s see what it comes back with. And now it’s actually looking at the thing and coming back with, okay— 

DANNY SMITH: Whoa. [Laughter] 

KALYANI KOPPISETTI: Looks like there’s three options that you can choose from, Danny: studs, the sparkly earrings, and the shoulder bag. I’m assuming your mom would want the studs, so I’m going to go with the studs. So all I have to do is add studs to the cart. Again, I’m not clicking here and there. If you wanted to, there is a button there to say “add to cart.” So you can do that. Now, the studs have been added. As I said to Danny, I’m going to purchase something else. I have jeans. 

I want to put together an outfit. I’m going to ask, “Hey, I need an outfit suggestion for my options for my jeans.” So let’s see what that comes back with. Okay. It came back with the top and earrings. As much as I like those earrings, I don’t want them, but I’m going to go buy the top. I add it to cart. Now you can see on the right side—you can see the earrings and the top. 

Now I’m going to ask it to create—you know, show me some handbags because I want to show you the experience of going to different products within, and then finding that product and being able to shop for it. So I’m going to ask for handbags. 

DANNY SMITH: This is going to make me look like an amazing son, by the way. My mom’s going to know this wasn’t me, but it was like an assistant helping me. This is amazing. 

KALYANI KOPPISETTI: Yeah. By the way, all the images that you see, everything that you see is being generated by Amazon Nova. We’ve used the LLM model to generate the images. So the handbags and everything basically used the description and did it. So I like that shoulder bag. So I’m going to go with that red shoulder bag with my jeans and my top. 

So now I’m ready to check out. All I have to do is check out, and it’s preparing the checkout. Okay. So that pops up your Stripe Checkout experience. Here you have multiple options. So let’s say I put in Danny’s email. So Danny has registered himself in a Stripe product called Link. So as soon as I entered his email, the checkout experience recognizes that Danny is in Link and it asks to confirm that number. So if I want to, I can enter it. 

And because it’s a test account—again, you can’t use Danny’s email account to buy things; it’s a test account. You can use that credit card information. Or I can come back and say, “You know what? I don’t want to use Link. I want to use the credit card information. Or it’s too much. I want to use buy now, pay later.” I can do that. 

DANNY SMITH: Yeah, buy now, pay later. One of the key things we want to point out, as part of Optimized Checkout Suite, is dynamic payment methods. Because of the amount of the transaction, it does surface Klarna, some of the buy now, pay later options. If the amount was smaller, it would remove those options automatically. So you can design those custom rules. There’s AI and also custom logic that you can add to only surface the appropriate payment methods for the right checkout experience. 

KALYANI KOPPISETTI: So we just wanted to spend some time on the architecture that we used. So we talked about the experience for the customer. We talked about security. So some of the supporting services that we used are basically AWS IAM service, Amazon Cognito, to make sure that we authorized the user before they entered the shopping agent and go in with the payment options. 

For most of our services, we’ve used serverless architecture. We have all of our back-end processing in Lambda. The front end is done by Amplify and hosted in front of CloudFront. And from a product preparation perspective, we had the product catalog. As I said, we used LLM to generate the product catalog. I basically told the LLM, “Hey, I’m trying to create a demo. I need different types of products. What product category should I create?”

It gave me product categories. And then I said, “Okay, for all these product categories, give me a prompt. How do I generate the information?” Then it generated the information. And I said, “Okay, take this information and create the images for me.” So it created that image. So all of that data was in the S3 bucket. And then we used Amazon Bedrock to convert that into a knowledge base that the agent can access and surface those products. 

Again, we restricted those products to three or four for this demo. That would be based on your own experience as to what you need to do. So from there onwards, we used Lambda also to call Stripe API for the checkout. So Danny, do you want to walk through the— 

DANNY SMITH: Yeah. I want to spend a little bit of time to walk through the actual integration technique that we used. So as Kalyani said, we used AWS Lambda, which is a serverless function call, function as a service. But we actually introduced a Stripe layer. So we’re going to publish a blog on this and make it public on how exactly we built this. 

But long story short, a Lambda layer allows you to put Stripe function calls into any of your Lambda functions that you write across the board. So we chose the Checkout Session API and also Payment Links. We have two different variables that we put into the equation as far as how you can generate the checkout experience. 

As you saw this morning, Will Gaybrick debuted the Order Intents API. We’re going to be adding that very soon. Long story short, this is a very modular solution that allows you to make any Stripe object or API call directly into your agentic AWS workflow. So that was the integration technique. And again, we’re going to publish information on how you can build that as a reference architecture. And then all the Stripe goodness that comes after the payment comes into play here. 

So we’ve got Stripe Radar, obviously, that’s going to help you with fraud prevention. We’ve got things like adaptive acceptance and card account up there that’s going to help you with soft declines and making sure that we’ve optimized the conversion experience. And then with our out-of-the-box integration with AWS EventBridge, this is very key. So, basically, completing a shopping experience, you’ve got to trigger events that will basically start shipping these products, whether it’s digital products, physical products, or whatever. That’s definitely part of the experience. We have an out-of-the-box integration with EventBridge. 

And then finally, I did want to land on things like Stripe Data Pipeline. So, after you’ve completed the payment experience, how are you training these AI models to become more intelligent, to learn more about my mother’s style and liking? I always talk to customers, partners. You’ve got all this—payments data is customer data, and it’s experience data. So what are you doing to learn from that data to optimize the experience? 

And then, secondly, what are you doing with this data to help train the AI models to make them more intelligent and get better over time? That’s where Stripe Data Pipeline integration with S3, and then AWS Redshift, allows you to run analytics on that data and train these models to become more intelligent over time. So it’s more than just completing the checkout experience. It’s the integrations that we have with the back end on AWS that helps you learn and improve the overall experience from a shopping perspective.

KALYANI KOPPISETTI: Again, this is the reference architecture for this shopping advisor. Your particular implementation may look different. Your data may be in an RDS database. But things can be—you know, we just wanted to show how a reference architecture would look like for this demo. From this, there’s a lot of AWS services and Stripe products that we talked about in this thing, but I want to zero in on the generative AI stack specifically because that is a stack that we used a lot. 

As I said, the foundational infrastructure? AWS has the chips for it. We have the SageMaker AI where you can train the models. Again, we haven’t trained the model for this particular demo, but you should be thinking about how do you train your model to understand your customer needs? And what are those three products that you want to surface? How do you want to think about that? 

So for those things, you should think about how do you train that model, and AWS has a service for that. And then the middle layer is where the magic happens, in my opinion. We have a service called Amazon Bedrock. So what Amazon Bedrock does is basically it gives you a way to interact with all the foundation models that are out there from our partners, as well as Amazon. It is the easier way to integrate with one API and access those things. 

On top of it, Amazon Bedrock also has services like how do you create a knowledge base. It will guide you through the process of creating the knowledge base. You can use APIs for it. You can also create an agent. We’ve introduced multiagent orchestration. We’ll talk about some of these things in a later part. But when you’re having multiple agents come together and come up—in our example, let’s say I said, “Hey, I’m traveling to San Francisco for a conference. I’m going to do a wine tour, and I’m going to do this.” 

Then it would give the suggestions not only based on what it knows about me and what it knows about current trends. It also will incorporate into it what is the weather like in San Francisco around this time of the day and do I need a jacket. So all of those things are different. Not one agent can do all of the things. It can do it, but I would advise not to. But basically having that multiagent orchestration and using the tools it has is necessary to come up with what your customer wants. 

And then the final layer is the applications to boost productivity. And as I mentioned, I used to be a software engineer like 25 years back, and I knew how to code everything. Over the years, that has not been the case. That’s not been my friend. And UI coding is not something that I’ve ever done. But to build this, me and my team, who’s sitting in the front, used Amazon Q Developer extensively. 

We asked it to help us with the prompts. We asked it to create some code. We asked it to optimize the code. We said, “Hey, are we doing this right?” So we used it like basically what these days we call vibe coding. 

DANNY SMITH: Yeah. We vibe coded this thing. 

KALYANI KOPPISETTI: Yes. We had a really good vibe coding, and I was very happy with it because, at the end of the day, we were able to create something to touch and just to experience within a short period of time. So, with that, I just wanted to focus on the generative AI stack. But also we want to talk about cloud infrastructure. If you host all of this and are not able to scale, or you have a failure because way too many customers want to use your product, that’s not good as well. So what I want to do is I want Danny to—maybe we’ll switch here. 

DANNY SMITH: Yeah. Reliability. Let’s talk about reliability, a pretty intense topic these days based on current events. But we are very, very proud of our partnership with AWS from a reliability standpoint. Just me being in this role recently, I’ve gotten to know a lot of our friends at AWS, but also I’ve gotten to really know our Stripe engineering team very well. 

This is an amazing team of people who are very well prepared for any possible scenario. So the combination of our brilliant engineers and AWS’s amazing infrastructure gives us statistics like what you see on the screen. So I just—hands down, I think they’re the best in the business. That’s just my personal opinion. But, you know, reliability. Nothing matters if all this stuff is not up and running and if it’s not scalable and reliable. 

So that’s a key part of what we go to market with, with our friends at AWS. We’re going to go through this really quick, because we already talked a lot about Stripe Link. So you saw in the demo, when she put in my email address, I’m part of the Link network. It came up with a nice green button. I could have completed that checkout in about five seconds. So that’s the optimized experience. You can log out and use credit cards, Klarna, BNPL—as you saw, everything built into the suite. 

The overall message of Optimized Checkout Suite is we want to basically only surface the payment methods that make the most sense in an optimized fashion for the shopping experience and the checkout experience. The way we frame it is the payments experience is the customer experience. You’re not a customer until you pay for something. And the checkout experience is the payment experience. So if you get lost in a bad checkout experience and you abandon the cart, then all this technology that we’ve built—it’s landed nowhere. So that’s a huge part of Stripe’s go-to-market. 

We definitely want to talk about four models of agentic payments that exist today. So we demonstrated—so the first two have a lot of human intervention. And the last two are going to be what we’re doing as we move into the future. So the human input. The agent waits for human input in the checkout flow. That’s what we demoed today. So we did the AI shopping assistant. It built a nice cart. It opened a checkout window. And then we put in my information, and we completed the payment. 

So there’s still human input. It’s part of an agentic workflow, but there’s a lot of human input making sure that the experience is secure and safe. We moved into wallet authorization. So Stripe Link is a wallet; Amazon Pay. These are great examples. You could basically have an agent with the wallet credentials complete that checkout experience for you. So, again, if I were to use Stripe Link or Amazon Pay, we can basically authorize using a wallet. 

Third, and this is very interesting. So when I mentioned earlier Perplexity, Perplexity Pro is powered by—we use Stripe Issuing on-demand to create a virtual debit card that has a budget. So, basically, safety is the number one thing that we’re caring about. You don’t want an AI agent draining your bank account. So it’s like you want to make sure that you give it a budget and an expiration date. 

So Stripe Issuing is a powerful way of using a virtual credit card per transaction. You can generate an infinite number of these. We are using that to gate the experience. Some users are doing that today. That’s pretty cutting edge. We are going to move forward and start to make that available as we work together with our friends at AWS. 

And then, finally, this is true agentic payments where there’ll be no checkout, where you’re funding a stablecoin wallet. Stripe’s very big on AI and stablecoins, as you saw today and as you’re seeing all week. Stablecoin wallets are basically the true agentic payment flow where there is no checkout. You’re basically empowering a stablecoin wallet with the same gating that I mentioned earlier as far as the debit card. And then it’s communicating with another back-end API, and they’re just moving the money between merchant and buyer and completing the entire experience with no involvement. 

So we’re going to be working towards that as a future use case. But these are the four models that exist today and that—you know, some of them are live currently. And some of them we’re working on in the future. 

KALYANI KOPPISETTI: So we talked about a lot of agentic AI and everything else. This is all innovative technology that’s coming up. But at the end of the day, you have to balance your innovation with your security. So I would say, from an innovation perspective, make sure you’re building focused agents. Build focused agents and composable architecture so that you can learn from it and improve on it. 

Build smart knowledge bases. Your knowledge base should be easy to search. And product discovery. Again, this is one example that we took. It can be applied in all industries and all use cases. So from a lessons-learned perspective or best-practices perspective, make sure you do that. And you’re scaling with intelligence. You just don’t suddenly start putting all agents all in your front end. Start thinking small and doing it from there. 

And then, from a security perspective, make sure we have clear boundaries of what the agent can or cannot do. Make sure there’s safeguards. You’re monitoring your agent activity. Have your logs. Look at those logs, and make sure there’s no anomalies. If there’s anomaly detection, you have some parameters defined to take action on it. And it should not be something that you think of afterwards. So security should be at the forefront of your design from the get-go. Think about security as as important as your customer experience. 

And from a scaling perspective and from a use-cases perspective, as I said, a lot of experimentation is happening. Think about your use case and build on it gradually. In our particular experience, we just talked about product discovery, and we didn’t take into consideration the customer’s previous purchase history and all of those things, but you could technically add that over the things as you’re experimenting with your customers, seeing what your customers really want. 

And then governance is important. I’m going to click to the next slide just because I do like surprises. I’m here. Danny’s here. But I want Hasan, and Balu, and Mohan—I don’t know if you’re here—to stand up. So these are the three folks who helped us build this. Thank you so much. [Applause] 

Without you, this demo wouldn’t have happened. Chintan, unfortunately, was not able to join us here. But he was also critical in terms of helping us understand some of the concepts behind some of these things. So, thank you for joining us. 

DANNY SMITH: Thank you so much.