EMILY GLASSBERG SANDS: Well, we are all living it. Today’s models can reason, and understand, and act with remarkable sophistication. At Stripe, we combine these advancements with our own unique data to put artificial intelligence (AI) directly to work for your business. At its core, Stripe’s AI is powered by our massive scale. Last year, we processed $1.4 trillion—and that’s across every industry, from startups to global enterprises and from buyers in nearly every country. So our AI is truly trained on the breadth of the global economy, but as important, it’s also a powerfully dense network. For example, 92% of your card transactions are on cards we’ve already seen elsewhere, and growth of that Stripe network is only accelerating.

Payments volumes on Stripe grew 38% last year. And now, even businesses processing transactions off Stripe are using Stripe products, like Radar and Checkout, creating a virtuous cycle that strengthens our AI and in turn boosts your performance. So today, I’m going to walk you through how we’re helping you harness AI across the payments lifecycle. I’ll give you a deeper overview of our new foundation model, and I’ll show how you can reap the opportunities in agentic commerce.

Stripe’s been using AI for over a decade across our payment stack and that AI is a force multiplier. It’s a practical, productionized engine that learns from transactions across our network to make the best decisions for you. The payments lifecycle is critical to your bottom line, and Stripe’s AI powers every step shown here, from personalized checkout and fraud prevention to maximizing transactions approvals and managing disputes. That all happens in real time, optimizing jointly across conversion, cost, and fraud in service of your profit.

So let’s take a quick look at how AI advances over the last year are boosting performance in each of these stages, starting at the beginning with checkout. Stripe’s AI now customizes your checkout further, from selecting the right payment method for each cart and customer to removing unnecessary fields so you convert more business. Next, when that customer tries to pay, we help with compliance, automatically triggering authentication as needed and we do so, importantly, while minimizing friction.

Since Sessions last year, our AI advances cut down unnecessary authentication challenges by 20% and simultaneously reduced fraud by 8%. Then comes fighting fraud directly. So every day, new attacks emerge and our fraud detection product, Radar, helps block bad actors in real time, stopping that fraud before it hits your business.

We’ve made a ton of progress here. Industry-wide, e-commerce fraud is up 15% year on year. But Radar users are actually seeing dispute rates continue to fall, down 17%. And we’re expanding Radar’s reach across both payment providers and payment methods. Radar now works for transactions processed on other providers. DoorDash, for example, saw a 10% reduction in chargeback costs from adopting Radar for their off-Stripe transactions. And because of your requests, we’ve expanded Radar to cover more payment methods, including the much-asked-for Automated Clearing House (ACH) and Single Euro Payments Area (SEPA) with new specialized AI models. Early adopters, like Zoom, are seeing ACH fraud down 20% and SEPA fraud down 42%. 

Okay, great. So then we’ve confirmed that a payment is safe, but you still need the networks and issuers to actually approve it. False declines—so these are legitimate transactions that are mistakenly rejected due to suspected fraud—are a real problem for online businesses. Adaptive Acceptance is our AI-powered optimization tool. It helps you recover legitimate transactions wrongly declined by banks, and this includes by retrying initially declined payments in real time. Last year alone, Adaptive Acceptance recovered over $6 billion for businesses on Stripe, and this is thanks to our advanced transformer models. So I want to show you here the retry successes before last year. This is the share of retries that were successful. Then we updated to a deep neural net, which boosted retry success rate by 60%.

But this is my favorite part. As AI does, all on its own, it keeps getting better. Year to date, we’ve seen a further 27% increase in retry success, recovering way more revenue for our users in April than in any previous month. And then finally, the optimal amount of risk for a business isn’t zero. And even as fraud protections improve, refunds and disputes continue to be an important part of your business. So yesterday we launched Smart Disputes, an AI tool that handles disputes for you. It predicts which disputes to fight and then goes and creates tailored evidence submissions for the ones worth fighting, saving your business both time and money.

I mentioned earlier, sort of in passing, that our AI operates as a system tuned to drive your profits. Well, under the hood, our transformer-based AI is evaluating thousands of potential pathways for each transaction across all of the metrics that matter—authorization rate, and fraud risk, and refund likelihood—and then it’s choosing that one path that’ll boost your profits. So consider the seemingly minor decision of whether to request 3D Secure (3DS) authentication where it’s optional. Letting our AI decide whether to authenticate a given customer reduces fraud by 32% on Stripe Checkout, our prebuilt checkout form offering, with no negative impact to conversion. And so today we’re making this feature opt-in also for businesses not using Checkout, and you can turn it on right in your Radar dashboard. 

Given the criticality of AI to your business, we’re doubling down on putting our data to work for you. Last year in Sessions—any repeat offenders in this same AI breakout? A few of you? Thank you, I’ll see you again next year—I laid out our vision for foundation models in payments. For years, Stripe developed highly specialized machine learning models for checkout optimization, for fraud detection, and for authorization. Each was built for a single narrow task, and they’ve served our users well. But foundation models give us a chance to go much further, learning from more data more holistically to improve all those applications faster.

Yesterday, we announced the industry’s first-ever Payments Foundation Model. It’s trained on tens of billions of transactions and distills each charge’s key signals into a single, versatile embedding. You can think of the result as a vast distribution of payments in a high-dimensional vector space. Okay, everyone knows how many dimensions a cube is, but I’m talking hundreds of dimensions. And then the location of each embedding captures rich data, including how different elements relate to each other. Payments that share similarities naturally cluster together, so transactions from the same card issuer are positioned closer, those from the same bank even closer, those with the same email address closer still, and so on. And we can then use these embeddings to make all sorts of different kinds of predictions about each future transaction in real time as it comes through, based on where it is in that high-dimensional vector space. And this is producing real results across our business.

Our Payments Foundation Model is already delivering wins that weren’t possible with traditional methods, like in card testing. Thanks to our new foundation model, we now identify over 95% of card-testing attacks in real time, a 22% improvement over prior methods. And this is just the beginning. We’re weaving our payments foundation model across our stack to unlock the next set of performance gains for all of you.

Of course, better AI isn’t just revolutionizing payments performance. We’re all seeing AI transition from knowing to doing, right? From answering your questions to actually performing tasks. And this shift is bound to fundamentally transform commerce. But in order for it to all be real, agents have to be able to collect, and move, and spend money. Now, for years, Stripe has optimized payments, and checkouts, and billing flows for human buyers and businesses. Today, we’re doing the same for these AI agents. We launched our agent toolkit in November last year to make it easy and safe for agents to transact directly. And it’s being downloaded by developers, including—I hope—some in this room, thousands of times every week as startups and skilled businesses build new kinds of commerce agents. This includes agents that can shop for you, from booking your flights to sourcing and buying the best coffee beans on your behalf.

I wanted to show you one of those agents in action. So here I am chatting with Barista Agent. She’s a new coffee-buying agent that recently popped up. You should check her out. She’s pretty good. I tell the agent what I’m into, and she helps me discover good beans from across the Internet. So it’s conversational. Yes, I do like my coffee very chocolatey, okay? And then when the agent finds a bean that I like, I can buy it. But here’s the thing. The agent doesn’t actually have any inventory at all. It just goes out and purchases for me from across the Internet. On the back end, the agent’s using a secure, single-use card from Stripe Issuing to execute the transaction, which both keeps spending controlled and keeps all my data safe.

So that’s a commerce example, but these agents can also help operate your business, tackling all sorts of back-office financial tasks. I don’t know—issuing a refund, or updating a card on file, or changing a billing plan. You know, all the stuff that you and your team have to do but wish you didn’t have to do so you could get back to more impactful work. So I want to show you how ElevenLabs, which is an AI-powered voice generation startup, uses our toolkit to let a voice agent handle customer support. Let’s take a look.

[VIDEO]

Subscribe to the Eleven Labs Pro Plan, please.

I’ve sent you a checkout session link. You can use it to complete your subscription to the ElevenLabs Pro Plan. If you have any questions or need further assistance, feel free to let me know.

Cool. I’m just going through the process now. One second.

Take your time, Louis. Let me know if you need any help along the way.

So I should be subscribed. Can you see that on your end?

Yes, I can confirm that you are now successfully subscribed to the ElevenLabs Pro Plan.

EMILY GLASSBERG SANDS: And we use the Agent Toolkit ourselves. Actually, today, we’re thrilled to put in your hands our own agentic assistant right in the Stripe Dashboard. So I want to show you a sampling of some of what it can do. First, I’m going to ask it if it’ll create a product in my product catalog. I’ll state the price, $45, and also ask for a payment link; product’s created. I can ship this payment link off to a customer. And then let’s say I want to issue a refund for a particular customer’s particular transaction. Here, you’ll notice the agent asked me to confirm before executing. I appreciate that. And finally, here, I’m going to ask it for some business insights. It’s going to run over to Sigma and grab the data it needs.

The agent can do all that and much more from this single chat window and based on what we tell it in plain English. Zooming out, as AI agents become more common, Stripe is building the financial infrastructure to support them. I want to say this isn’t just for businesses, like Barista Agent, where the agent is the product. We’re making it easy and secure for any business to build, and deploy, and monetize agents on Stripe. Now, that requires rethinking how many moves when it’s an AI, not a human, completing the transaction, so we’re making checkout machine-readable. We’re allowing agents permission to access your product catalog. We’re enabling transactions, as you saw from chat, or any agentic interface you create. And, of course, we’re evolving fraud detection and disputes for an AI-driven world, all so you can make the most of the AI opportunities in your business today.

Then let’s hear now from one of the most exciting startups building in the agent space. Please join me in welcoming Scott Wu, CEO of Cognition. I really should have said founder and CEO. He’s been at it for day one. Maybe you could tell the folks in the room a little bit about Cognition, what it does, why you started it.

SCOTT WU: Yeah, absolutely, and thank you so much for having me here. We’re Cognition AI. We’re building the AI software engineer, Devin. So we’re all programmers ourselves, and we were super excited to figure out how to use AI to make programming 10 times more powerful. Devin is a full end-to-end software agent that will work in your systems. It’ll work on your code base. It’s able to test code itself, read documentation, browse the web, create PRs on GitHub or GitLab or whatever it is, and do all the things that a human software engineer would do.

So Devin really is basically your junior buddy that helps you implement tasks and build out software. Right now, we’re working with companies as small as two- or three-person startups and as big as the biggest banks in the world.

EMILY GLASSBERG SANDS:  Amazing. You’ve built a lot of product. Word on the street is you only have 15 engineers. I’d like to fact-check that. Fifteen human engineers.

SCOTT WU: Fifteen human engineers.

EMILY GLASSBERG SANDS: Fifteen human engineers. Okay. Do your human engineers use Devin? How many Devins does each human engineer use?

SCOTT WU: Yeah. I’d say the average engineer on the team is probably working with about five Devins at once. I mean, Devin is by far the biggest committer in the Devin code base in production. So it’s really—it’s changed our workflow, as well. I mean, we’re able to spend a lot more time thinking about the high-level stuff and just having Devin do a lot of the implementation and execution.

EMILY GLASSBERG SANDS: Okay, so 85 engineers’ worth of productivity from 15 FTEs, if you will.

SCOTT WU: Yeah.

EMILY GLASSBERG SANDS: Amazing. As you’ve gone about sort of implementing agents in the real world (and, by the way, using agents in your own day-to-day) I’m curious what counterintuitive—you know, people talk about trust, people talk about hallucinations, people talk about what are the tasks that are well-positioned to be automated or not—but what’s the counterintuitive thing that you’ve seen in agents that you think folks should have eyes for?

SCOTT WU: You know, one of the big things that I’ll call out, actually, is—we see this all the time—which is intelligence actually is not the bottleneck anymore in most of these situations. I think you guys have seen this all already with some of the really powerful reasoning models and the work that a lot of these labs are doing. But we’re already seeing that AI is able to solve some very, very hard problems. I think now the next phase which we’re in and really working on right now is how do I get AI to solve my hard problems? So for us in code, a lot of what that means is how do we get Devin to understand your code base better and better? How does Devin learn over time and pick up knowledge from working with your engineers? How does Devin use all the tools of the stack to debug its own work and things like that?

EMILY GLASSBERG SANDS: When do you think Devin will be able to use most of the tools in the stack?

SCOTT WU: So it practically—

EMILY GLASSBERG SANDS: And is Model Context Protocol (MCP) helping?

SCOTT WU: It is, for sure. Yeah. So it’s an interesting one where I think there’s kind of the first order problem of, I’ll call it, plugging in the integrations, basically, and having it use the tools. MCP is incredibly helpful for that. Then the second problem after that is—maybe I would describe it as using the tools intelligently, right? So if you give a software engineer a bug and you say, “Hey, this thing crashed and it’s not working,” probably the process that they’ll follow is something like, “All right, let me go run it all locally.” I’m going to go click around on the front end myself. I’m going to try to reproduce the bug. If I do, I’ll find the error on Datadog or wherever it is, right? I’ll read some documentation online and then try to use that to debug, and then I’ll test it all again to make sure it works after I’m done.

So Devin obviously needs to be able to think, and plan, and go through these processes the same way. I think a lot of the hard part there actually is in having it decide how to use its tools intelligently to increase its likelihood of success, if that makes sense.

EMILY GLASSBERG SANDS: Interesting. Fun fact or embarrassing fact, I’m actually an economist by training. I know you’ve thought a lot about the economic impact of agents— 

SCOTT WU: Yeah.

EMILY GLASSBERG SANDS: —in the software-engineering domain and more broadly. I’m curious, sort of, what you think general equilibrium will look like, what you think the overall economic impact of these agents is positioned to be.

SCOTT WU: Yeah, yeah, yeah, absolutely. No. There’s a lot of fear out there, honestly. We really are—we’re all programmers ourselves. We love programming. We love building software. I think one of the big things to call out is there really is so much more software to build. I’ve never met an engineering team that says, “All right, we’ve got this project, and then we’ve got this project, and then we’re done. No more software ever again.” It’s always—

EMILY GLASSBERG SANDS: It’s not a zero-sum game, right?

SCOTT WU: Exactly, and there’s always 20 more projects and you have to pick two or three because that’s all you can take on. I think that maybe one way to put it is, I think as software gets more and more powerful, I think, if anything actually, programming at its core is just telling your computer what you want it to do, you know? As computers get more and more capable with AI, I think this act of software, and building software, and telling your computer exactly, “Here are the problems that we’re going to solve, and here’s the solution that we’re going to put together to solve it,” I think is only going to be more valuable. We’re really excited, I think, for the golden age of software engineering with this.

EMILY GLASSBERG SANDS: I heard a lot of rumbles in this room. Nobody here thinks there isn’t more stuff to build.

SCOTT WU: Yeah.

EMILY GLASSBERG SANDS: As I was actually walking in, there was a whole crew that was sitting in the chair pretending like they were waiting for the talk, but really just coding, like trying to get their next thing built in between the passing periods. Well, zooming out from coding for a second, I’m curious what the public is still heavily underweighting or underrecognizing about the impact of AI or how agents are going to change the landscape.

SCOTT WU: Yeah, I think we have a long, long way to go with AI. I think it’s crazy to say because, of course, there’s so much hype in AI already. But if anything, I still think it’s—yeah, I think folks are often underestimating it still.

EMILY GLASSBERG SANDS: And underestimating the end state, or underestimating the pace, or both?

SCOTT WU: I would say both. I think over the next couple of years, there’s going to be a lot of very exciting things going on. A lot of it in my mind comes down to actually agents as the core, I think. I really liked what you said about transitioning from knowing to doing, right? I think one of the things I think about all the time is we talk about this whole Generative Artificial Intelligence (GenAI) wave. I think, really, there’s been two GenAI waves. There was the first one, which I would say is kind of like this full training on the Internet, right? That was ChatGPT in November 2022.

EMILY GLASSBERG SANDS: ChatGPT moment, yes.

SCOTT WU: Exactly, yeah. And machine learning is what we would call imitation learning, basically, where you read as much text as you can and you try to train a model that will talk like the text, right? So, basically, we have a model that sounds like a smart person on Reddit or something because we’ve read all the Reddit comments out there, right? It was really amazing and it did a lot. But now, really, a lot of the advances are coming from a very different thing, which is high-compute reinforcement learning (RL), which I think, for us, maps a lot more to basically getting to try problems, work on problems, interact with the real world, and get feedback from what’s working and not working and learn that way. I think that’s been the name of the game for the last maybe one year or so. But it really is going to unlock a lot more in terms of autonomous systems that are actually going to go and take on chores and do a lot more.

EMILY GLASSBERG SANDS: So we all experienced the ChatGPT moment, right? 

SCOTT WU: Yeah.

EMILY GLASSBERG SANDS: Is there going to be a ChatGPT-like moment for agents? I don’t think it’s Will on stage, necessarily, showing agentic commerce. But what’s going to be the moment when people are like, “Oh, yeah. I get it. Agents are here. Agents are real”?

SCOTT WU: Yeah. It’s interesting because it’s all already possible. I mean, Devin is very much focused on software engineering, but as a simple example, we order all of our DoorDash with Devin. We order all of our Amazon packages with Devin. It’s like, “Hey, Devin, could you buy us some more whiteboards or something?” 

EMILY GLASSBERG SANDS: Wait, hold on. How does that actually work?

SCOTT WU: Yeah, so we have a lot of trust in Devin. Perhaps more than we should, but we give Devin our cards and stuff.

EMILY GLASSBERG SANDS: Yeah, tell me about the card credentials.

SCOTT WU: Yeah, exactly.

EMILY GLASSBERG SANDS: Are you using Stripe Issuing? Like should we be concerned? 

SCOTT WU: As it turns out, I think a lot of it comes down to just—as we’re solving a lot of these fundamental systems of obviously planning and long-term decision-making, but also working with the browser and managing secrets, and credentials, and things like that—I think a lot of these are becoming more and more capable. But yeah, I think there will be a moment, honestly, pretty soon, where a lot of folks’ day-to-day workflow changes. You can imagine there’s a lot that folks are doing in shopping. There’s a lot that folks are doing in—you know, Deep Research is a really amazing product, as well. But I think the big kind of switch in people’s heads will be when they get to a point where they’re really comfortable having agents do a lot of their day-to-day chores for them.

EMILY GLASSBERG SANDS: Okay, so total selfish sidebar. I have a three-year-old and a five-year-old, and they grow like weeds. And they’re very picky about what they wear, and I just don’t have time for it. Should I actually be working with Devin to make sure they stay clothed? Pink rain jacket, minimum sparkles, neck size up, less than $30. Is Devin going to be my play?

SCOTT WU: We’ve seen folks who do it. It’s actually not the work—I mean, it’s not what we built Devin for or what we typically endorse as a workflow, but it does happen.

EMILY GLASSBERG SANDS: You never know, when you build primitives, how they might be used, right?

SCOTT WU: Yeah.

EMILY GLASSBERG SANDS: Great. What is your mental model? I know you’re not that worried about me clothing my toddlers. What is your mental model for how AI agents are being adopted in companies?

SCOTT WU: Yeah. And back to the point we were saying about intelligence not being the bottleneck in a lot of these cases. A lot of it actually is now really about thinking about how do I give these agents all of the knowledge or the decision-making that we have here, right? So, for example, it’s how do I onboard Devin to my docs? How do I get Devin to run my code locally? Or things like that. I think practically—you know, what we often see is I think there are a few folks or a handful of teams within the company that are the real kind of early adopters and are pushing things forward and are setting up these systems and teaching agents how to use it.

I think one of the really cool things with these agents is that they learn from every interaction that you have, right? So you get to a point where – with Devin, for example, Devin knows all the files in your code base and has talked to your engineers about all these things. It makes it much, much easier for the next wave of the rest of folks at the company to come and onboard. We’re seeing this happen basically everywhere, but I think of it as a parallel of getting a lot of the world’s businesses online, for example, in the Internet era, the mobile era, and so on. I think one of the big phases that we’re really getting into right now is getting to a point where the big question is figuring out how to get a lot of this information and a lot of this decision-making power to the AI so that it has the same context that you want it to have.

EMILY GLASSBERG SANDS: Do you actually have any tips for businesses that are looking to adopt agent software, presumably from a bunch of different providers for a bunch of different use cases, like what should they be reasoning about from an infrastructure layer or a data layer, or a security layer? Anything top of mind?

SCOTT WU: Yeah. It’s funny, but I really think that the right parallel here is actually just thinking about how you’d onboard a human. It’s like here are the docs, here are the onboarding guides, here’s a good first task to give them that’ll teach them about whatever it is—the context that they’re working on—and going from there. But that’s often the way that we’ve seen go most effectively. Yeah.

EMILY GLASSBERG SANDS: Big misconceptions about where all this is going?

SCOTT WU: Yeah. I think, honestly, the biggest misconception which comes to mind is the idea that this is going to be bad for us. I think it is going to be amazing for us.

EMILY GLASSBERG SANDS: Agents in general?

SCOTT WU: Agents, and AI, and this whole wave. Maybe one way I would describe it is I think it’s always been the case with computing and automation that AI or whatever you want to call it takes on the repetitive and tedious parts and it frees up humans, obviously, to spend all their time on kind of the more strategic, high-level decisions, right? I think the only thing that’s really changed over time is we’ve really just kind of kept pushing the border of what we count as repetitive and what we count as high-level decision-making, right? Obviously, there’s such a long way to go, I think, in terms of how much more any one person can do or any one person can accomplish. I’m incredibly excited for AI to push a lot of that forward, so…

EMILY GLASSBERG SANDS: We talked about how you’re creating more leverage and more capacity in your organization with Devins, and that there’s always more code to write and more products to build. I’m curious. Nonwork question. If you could just magically free up 15, 20 hours a week because of agents, what would you go do?

SCOTT WU: Oh, that’s funny. I feel like I would—

EMILY GLASSBERG SANDS: He would build more software, yes.

SCOTT WU: I would probably write a lot. You know, one that comes to mind is I’ve always wanted to be like a Twitch streamer celebrity. The thing is—

EMILY GLASSBERG SANDS: Really?

SCOTT WU: —I’m not actually talented at like video games or whatever, so my thing would be coding. So I’ve done Twitch streams of me coding and stuff, and it has some viewers and stuff, but that would be probably the— 

EMILY GLASSBERG SANDS: The creator economy has not yet taken off in his niche, but we will start it. Amazing. Amazing. Okay, well, we will all try to follow you. Thank you, Scott, for a fascinating conversation and for giving us a glimpse into where all this is headed. Thank you all for joining us. I know it is getting too late in the day and late in the conference, but it is certainly an exciting time to be building new things in AI. I hope you take forward and enjoy the rest of your time at Sessions.