DAN HILL: My name’s Dan. I lead the consumer product team at Stripe, where I work on Link. So for this talk, I know this is a sort of business conference and you’re all here for your startups and businesses, but for this talk, we’re all going to be just regular consumers. So maybe get a show of hands. Who in the last three months has built some kind of personal agent for themselves, has used Claude or Cowork to make something for themselves, installed OpenClaw maybe? Raise of hands. Okay. I would say most of the room. That’s great. Okay. So today I’m going to talk about the growing consumer trend of personal AI. As you saw from a lot of the content yesterday from the keynote from Patrick and Sam talking about OpenAI, the last three months have been a real inflection point for AI and including I think for consumers. I think this rise of personal AI is going to be really impactful for both how consumers use AI in the future and what it means for agentic commerce.

Okay. So today we’re going to cover three things. I’m going to start by digging more into this rise of personal AI. We’ll look at a bunch of data from Link and some UXR we’ve done to kind of explain the trend and what we’re seeing. Then I’m going to talk about Link’s wallet for agents that we shipped yesterday. Give a demo of how that works and how we’re helping consumers buy with agents. And then finally, Kate Jensen from Anthropic will join us on stage and we’ll talk more about what this means for businesses and enterprises in the room. Okay, let’s get in.

So if you’d asked me six months ago, how do most consumers experience AI? I would’ve said one or two ways. The first way is chat. You go to ChatGPT, Claude, you ask it a question, it gives you an answer. You’re building a recipe, you’re doing some research. It’s kind of a back-and-forth chat interface. And the second way is AI coding. More and more people using AI to code, certainly at Stripe, it’s the primary way we build Stripe today. Over the last, I’d say three or four months, these two worlds have started to really, really intersect.

And they’ve intersected in two ways. The first is the rise of OpenClaw. Now what OpenClaw did was basically take all the power of the AI coding agents and bring it to a consumer experience. And it was one of the fastest growing, I think, consumer AI products ever. On GitHub, within 60 days, it’s the most starred software project ever. It has more stars than React, more stars than the Linux operating system. It was a huge sort of moment. And the reason is it gave consumers a feel for what AI can really do. It can do long-running tasks. It can open up a browser. It can do things. It can have access to your email, your desktop. It can do a whole bunch more than just Q&A back and forth. Now, the second way that I think we’ve seen personal AI take off is more and more consumers, as I said at the start of this session, using Claude and Codex and things to build their own apps, to build their own agents. Last weekend, I didn’t like how my family calendar was running, and so I had Claude in a couple of hours, build me a new calendar app for my family. I live in DC where nobody works in tech. Nobody knows what I do. But yet all my friends who are lawyers and teachers are in the evenings vibe coding, some little app, some game, whereas a year ago they wouldn’t have been.

So I think this tweet from Andrej Karpathy says it really well. For the first time, I think consumers are starting to really experience AI in the way that many of us who’ve been using it for coding have been doing. AI that can actually do things, not just question and answer, but take action on our behalf, do tasks, think for itself, and be a kind of personal assistant.

Okay. It’s not just individuals, I think, spinning up OpenClaw, though that’s been a huge trend. The same idea of personal AI doing more, I think, is spreading across all of the ecosystem. So some of the big AI companies like Anthropic launch Claude Cowork. So the idea of bringing the power of Claude, but to the desktop so it can do a whole bunch more there. OpenAI, of course, acquired OpenClaw. Perplexity in February, launched Perplexity Computer. Same idea. How can we bring the power of these models to the desktop where it can have so much more access to your data?

And we then have all of the sort of agent and app builder platforms, Replit, Lovable, Bolt, Vercel, have all seen huge growth in the last few months. Consumers just building their own little apps on these platforms. And startups too. I’ve spoken to dozens over the last few months. There’s so many new startups building personal assistance, travel planning apps, chief-of-staff-type operating systems for consumers. A third of the Forbes AI 50 are building some kind of app agent builders.

Now, to back this up, we have some great insights from Link. If you haven’t heard of Link, it’s our consumer wallet. If you ever buy a subscription from OpenAI or Anthropic, you’ll see it there as a way that you can pay. It’s on a bunch of merchants. And we have some pretty unique insights, I think, into this trend. So Link is on 94% of the consumer companies in the Forbes AI 50, and we now have over 250 million consumers in the Link network. So we get this kind of vantage point where we can see how consumers are using these AI products. So I’m looking here at how much a consumer spends on AI in a given month. So this is consumers, not businesses, this is just individuals paying with Link for AI. So looking here at the P90, so this is the top 10% of Link users. You can see a fairly steady increase in spend over the last couple of years. And then starting this year, a huge inflection point, spending over $350 a month on AI. So this is not just one subscription, this is multiple subscriptions. This is paying for the pro plans from the big labs to get really great coding ability. If you look at the P75, so the top 25% of Link consumers, same thing, 42% year over growth in spend on AI. And the P50, the median consumer, also still a lower amount, but getting up to the $75 a month spent on AI. So again, we’re seeing this trend start obviously with the power users, but it’s coming down to the median consumer too, where they’re really starting to spend a lot of money every month on AI. If I just click on the builder platform, so how much they’re spending on Replit, Lovable, Bolt, Vercel, places where consumers can go to build little apps and build projects, they’re spending 5x what they were a year ago.

So in summary, the trend I think we’re seeing is that more and more consumers are running their own personal agents. They’re building their own apps. AI is no longer just a chat interface. It is a thing they can actually experience and have it do tasks for them and sort of operate on their behalf, and they’re willing to spend. They’re spending more than ever on AI. And so that brings us to what they want it to do. They’re sending agents out to browse and buy, and now we’re increasingly seeing them want to actually transact on behalf of the consumer.

So let’s talk a bit about how do we actually get these AI assistants to actually transact. So first, let’s kind of lay out the ecosystem. We have consumers, the people, us in this room as consumers who want our agents to be able to go buy things. We have the agent builders. So any of you in the room who are building personal assistants, building AI apps for consumers that you want to be able to go and buy on behalf of the consumer. And then we have the sellers, the places that the consumers and agents want to go buy from.

When we thought about Link, we thought, how can we actually address the needs of all three of these groups? How can we build a product that works for everybody? So what do they each want? Consumers want their preferred payment methods. I don’t want to spin up a new crypto wallet and figure out what Solana is and gas fees in order to go fund my agent. I just have a credit card. I want it to be able to use my credit card and buy things for me. They want controls and security. As great as AI is, I still don’t trust it with my bank account. I would like there to be some control in place that I’m going to be protected.

Those of you that are building agents, what we’ve been hearing a lot of is they want to purchase across the internet, right? You don’t just want to be able to buy from places that support certain protocols. You want it to be able to buy from all the places that consumers want to buy. And you want to be able to stay out of the funds flow. Stripe’s really good at complex money movement. You guys don’t want to be dealing with money transmission and figuring out how to move money around. And then there’s the sellers. Really excited about agentic commerce, but don’t want to keep making massive integration changes, figuring out how to support every new protocol. And at the same time, want agentic commerce, but from good humans with good agents that you can understand, not just a sort of sprawl of bots that are hard to understand.

So that’s why we launched yesterday, Link’s agent wallet. It’s a programmable wallet. It gives consumers the ability to grant agents the ability to spend on their behalf. It can produce one-time use virtual cards. So a 16-digit card number, you can plug into a website checkout, or it can create Shared Payment Tokens. So if you want to pay with a Machine Payment Protocol, it can produce those two. The human today approves every transaction, and it never exposes the raw card details. So if you’re a consumer, you’re protected. If you’re an agent builder, super easy to integrate and you can enable a whole bunch of commerce.

Okay. Let me show you how this works. I’m going to jump to a demo. Oh, great. Okay. So before I get to the demo, yes, we launched it yesterday, link.com/agents. Go check it out. If you are building an agent and want to integrate it with your agent so it can spend on behalf of consumers, come check out the GitHub, Stripe Link CLI. Okay. So I’m going to do two things. I’m going to first give you an end-to-end demo of an app I built last weekend to kind of show you how Link fits in and how Link enables payments. And then I’m going to do a deep dive for a couple of minutes on what we call a spend request, which is kind of the core new thing that we’ve created here.

So for this demo, I’m going to spin up Claude. That’s my agent here. And before I get to the demo, I should give you the heads up. I have three daughters, and they love getting physical mail. It’s just like a really exciting thing when they open, the things come in the mailbox, and they see their name on it. I love dad jokes. I’m a dad, and I just love them. My daughters hate them, but I love them. So I was like, what if I could mail my daughters a dad joke just by pushing a button with my agent? So with that in mind, let’s launch our dad jokes agent.

Okay. So there’s going to be three parts to this. It’s going to create the joke. It’s going to mail it, and it’s going to use Link to help me pay for the postage. So the first thing it’s done is loaded our “create payment credentials” skill. This is the Link skill that tells the agent how it can spend with Link. Now it’s asking me what topic would you like the dad joke to be about and who should I send it to? I’m going to go with space for Iris, my middle daughter. Okay. So it’s going to go hopefully find me a very amusing dad joke. Great. It’s calling... I didn’t notice. There’s actually a whole public API of dad jokes if anyone wants to go integrate themselves.

Okay. I have three choices. I’m going to go with option one. How do you organize a space party? You planet. Okay. Option one. Yeah, that was the correct groan. Thank you. Okay. Now it’s going to... Yep, send it. That’s the right address. And it’s going to make the PDF.

Now, while it’s pulling that up, I’m just going to show you... We’re using this company PostalForm. They’re a great Stripe user, basically lets you upload a PDF and mail it. What they’ve done is integrate with our Machine Payments Protocol. And so basically you can, with an agent, send mail. It’s a super great product. The founders here, maybe in the room, it’s a super cool thing. It’s just so much fun to be able to send mail just by uploading a PDF.

Okay. So if I go back to my agent, it’s made the PDF. Let me just check it out. Yep. That looks like a great thing to mail. So yep, it looks good. Let’s go ahead and mail it.

Now, what it’s doing here is it’s kicking off the Machine Payment Protocol endpoint with PostalForm. So this is beginning the payment process using Machine Payment Protocols. Now, as I said, Link can support one-time-use cards or we can use Shared Payment Tokens for this authenticated, this flow here. Now it’s calling Link, and it’s checking if I’m logged into Link, and it’s realizing I’m not. And so in a second, it should prompt me to log in. So it’s decoding the Machine Payment Protocol response, and it’s generating me a link to log in here. So this is the link I need to basically grant access for my agent to use Link. So I’m going to pull up that URL and log in. Okay. So I can connect an agent. I can make agentic payments. I approve every spend, and my card details remain protected. That sounds good.

I’m going to continue, confirm the passphrase. And yet, my dad jokes agent on my Stripe laptop wants access to my Link account. It’s going to get my account details, and it’s going to get the ability to spend. Great. Let’s do it. Done. Okay. Now that I’m logged in, it should be able to complete the payment.

So it’s going to create what we call a spend request. I’ll come back to this in a minute, but this is the way the agent basically starts to set up the contract with the human to say, “Hey, I’d like to be able to spend this much money on this merchant. Can you get me a payment credential I can go use to complete that transaction?” So here we’re seeing it create the spend request. And in a second, if I come to my app, let me unlock my phone. I should have a push notification from Link coming in a second. Great. So you can see there, as well as a message from my kids’ school, a payment request for $3.40. So let me tap on that, open up the Link app, and I can see, okay, my agent here is looking to spend $3.40. I can see the merchant it’s looking to buy from, and it’s giving me some context on what it’s trying to achieve with this purchase. I’m just going to use my United Club card, which looks good, and I’m going to prove it with my face ID.

Okay. Now that’s done. Back to our agent. It should be able to now get the spend request, encode the PDF, and the letter’s on the way. Now, while it’s sending it off, I should say again, I built this into my own personal agent here. If you are building an agent, you’re building your own grocery shopping app, personal assistant, that kind of thing, you can absolutely integrate Link into that so that your consumers can spend with your agent. Okay. It’s got the 200 back, so in a second, it should be complete. Great. Letter’s on the way. I can pull up the confirmation. It’s been sent, and I can even see, it’s put a little space rocket. Great job, Claude. Okay. Thank you. Thank you. Just to prove it’s real, this is one I sent last weekend that my daughter received and she groaned heavily when she got it.

Okay. So let me for a sec, just deep dive more into what this spend request is because that’s the key thing that we’ve built here. So I have a couple of examples. This is using the Link CLI, which is what we have on GitHub. This is what your agent calls to be able to spend, and it creates a spend request. And so it has to pass a few things. The first is the payment method. So from the Link wallet, it can select which card or bank account it wants to use to pay. And then it has a choice of credential types. So it can say, “Hey, I need a one-time used card.” If it wants to go fill out a regular checkout form on the internet, or it can create a shared payment token. So if it wants to go be able to plug into an MPP server, it can create a shared payment token. We’re going to be adding support for more types of credentials in the future, so Link can support crypto-based things, stablecoin-based payments as well.

The next thing it needs is a merchant name and URL. This is me ordering flowers for my wife for Mother’s Day, which is in a couple of weeks. And so you pass the merchant name in the URL. Now this starts to build the contract. What we then do on Link is make sure that the credential the agent gets back can’t be used outside of what is being specified here. You then have an amount, $84.98. Anything above that, we will decline. So the agent can’t spend more than you agree. And then a context. We ask the agent to provide a bit of information about what it’s buying and from where so that you as a human can approve. So what happens is once the spend request is made, the human confirms, and then the credential that gets back is scoped to exactly what they specified here. So this way, the agent can’t suddenly decide to go off and spend all your money. It can’t just go do things that you didn’t allow it to do. So yeah, that’s how we create a spend request.

Okay. In the last about 10, 15 minutes, I would love to welcome to the stage Kate Jensen from Anthropic to talk more about how businesses are adapting here.

Thank you so much for being here.

KATE JENSEN: Thanks for having me.

DAN HILL: Okay. So to get right into it, in your view from Anthropic, what are the best companies, the companies that are embracing AI in the loop with their customers or in their operations, what are the best companies doing right now with AI and their customers?

KATE JENSEN: I have a few serious answers, but one of them is you were just using your Stripe account on Claude Code to send that joke. And actually the best things companies are doing are encouraging people to just use the technology. It was funny to see the little disclaimer across the top as you were like, “I think it’ll be a space joke today.” But the reality is I talk to CTOs all the time who say, “I look at our power users, and 99% of them are using this technology for stuff at work.” And there’s always someone who’s writing a novel in their spare time and they’re like, that’s okay. How do I just help people get used to this technology and understand how it can change their day to day, whether it’s at work or outside of work because that sort of adoption is what really, really matters.

DAN HILL: Yeah. If you think about this huge shift from consumers in AI and enterprises and business trying to keep up, what do you think people are getting wrong? If you were to speak to 50 of your top users, what is the kind of common misconception where you think, “Ah, they’re not quite thinking about this the right way. I wish I could just reframe this whole thing for them.” What is the misconception?

KATE JENSEN: Two things. One, they treat it like an IT rollout. This isn’t necessarily just a new piece of software that folks are using. It should and will fundamentally change how most of their teams operate and how you build products. And two, they outsource it to someone who, and I will say this very kindly, many of you may be chief AI officers in this room. There is an important role for someone to be accountable and hold an organization accountable to KPIs with AI, but the CEO, the CTO, the entire executive team have to be fully bought in.

DAN HILL: So you kind of want to see leadership from the very top down embracing AI, driving that. Yeah, that makes a lot of sense. Since you’re the ones building this technology day in, day out, and we are all both employees and individuals in the evenings, how are people at Anthropic using AI each day? And I’m curious for like the kind of like more avant-garde things that teams are doing with AI.

KATE JENSEN: I can tell you about the percent of code that’s now written with Claude Code. Thank you for using Claude Code for that, by the way. But there are a few just standout examples throughout the organization that I think are pretty interesting. Our legal team, many of you feel this deeply, I’m sure, is wildly understaffed. Anyone who touches the legal team, probably feels like they’re not being prioritized fast enough. And what they’ve done is actually used Claude to help anyone who interacts with them, understand where you stand in the queue, exactly who’s going to help you, and about how long it will take. It’s amazing. It makes the whole experience so much more user-friendly. I had someone present on my team just yesterday, actually, a woman came and she’s a front-line leader in our commercial organization and she has about 2,000 accounts in her book and she has a version of Claude, an agent that she herself built managing about 800 of those accounts entirely on its own. She was a little nervous to tell me that she was effectively outsourcing a lot of her job, but at the same time, it was just unbelievable to see what someone who is a salesperson could actually go build on their own and how they’re doing that.

DAN HILL: Yeah. I’m curious, because I think at Stripe, we’re a culture where AI is everywhere, we use it all the time, but how do you... Obviously you mentioned the top-down part where executives are sort of pushing it down, but I’m curious, how does Anthropic create the culture and the freedom where people are willing to just try AI? Even if it maybe doesn’t work the first time and it produces worse results than had they done it manually. How do businesses create that sort of culture with their employees?

KATE JENSEN: I think culture is so much in large part tone at the top, but then being very deliberate about what you hope to get out of the technology. You asked me earlier what a company’s doing well, and this is very general, but we see companies bucketing their AI investments into one of three buckets, sometimes doing all three at the same time. One is just give tools to every person in the company. Claude Code mostly for developers, though not exclusively. Claude Cowork increasingly for every other knowledge worker.

The second big bucket is how do you change how you operate internally? So you think about, we published a study where Goldman Sachs has completely rewritten their KYC process that touches Stripe in a lot of ways. You are too. Stripe is doing a whole lot of this. How do you think about risk management yourselves? How do you think about merchant onboarding? How do you operate the teams that look at all of those things? Can you completely rewrite what maybe used to be 200 steps and make it 30 and have agents do 15 of them? There’s so many opportunities for companies to be run more efficiently. And then the third, and this is something Stripe is also investing in a lot, is how do you build AI into the products that you’re shipping? And my answer to how does it work well kind of depends on which one of those you’re doing. In the first bucket, it’s find the champions, find the people who are just amazing at this and really adopt it and celebrate how creative they’re being. The second is have the right organizational leaders really running it and learning the results. And the third is just try it and experiment.

DAN HILL: Yeah. So something you just said there, I think we’re seeing consumers clearly more and more familiar with AI. It’s no longer a thing they don’t understand. It’s like they’re using it every single day, everywhere. I guess what I’m curious about is like, how are the best businesses and brands leveraging AI and their products in a way that feels right to consumers? Whenever I come to a chatbot and it feels very AI scripted, it’s personally not my favorite thing. So I’m curious, how do the best businesses build AI experiences in a way that consumers actually feel like this is good?

KATE JENSEN: I think the thing that is so exciting about the technology is how fast you can iterate. So the best companies that we’re working with who are doing this really well aren’t planning for three months and then shipping something. They’re planning for three days and then shipping a hundred things over the next three months. And you get user feedback really quickly when you reach consumers. You get hundreds of data points a day, if not thousands, if not millions, depending on what you’re shipping. Use that, change what you’re doing, and then figure out how to make it great. And the technology today allows you to do that more easily than ever.

DAN HILL: Yeah. I mean, certainly on Link, I think we see the same thing as Stripe’s consumer-facing brand. Thanks to AI, we can build things in a day or week and get that feedback loop really going.

KATE JENSEN: And you don’t necessarily have to build one thing and ship it anymore. If you used to have 10 user profiles, now you can have a thousand and build for all of them and manage it all because it’s a little bit easier to do that.

DAN HILL: So if you fast forward a year, and to your point, we should only be thinking a couple of months ahead because it’s where the future goes. But if you think ahead a year, what do the companies that have really sort of got exponential benefit from AI look like? And what separates them from just the averagely better company with AI? Well, what are the things that really differentiate those who will really succeed, especially the ones that are building consumer products, but just in general?

KATE JENSEN: I think it goes back to that culture question you were asking. Don’t be too conservative about what you’re letting people do. Celebrate if someone is really mailing a lot of jokes, maybe after 5:00 p.m., but celebrate if someone’s really mailing a lot of jokes from their account. How do you go and make sure that that desire to be really creative and to really push the limits of the technology is infused throughout the entire organization? And be very deliberate about those three different buckets. Make sure you’re doing something with all of them.

DAN HILL: So it’s less about trying to predict exactly what AI or the technology will be in a year and more a culture of rapid iteration, adaptation, learning, and being able to embrace it when it comes.

KATE JENSEN: I think so. The companies that I think are moving a little bit too slowly are exactly the ones that you might expect, but in some ways, those companies that are generally traditionally more old-school enterprises, heavily regulated enterprises that have long procurement cycles, some of them are actually best positioned to go be very creative because they have so much information about their consumer bases. Where I find that companies are partnering with us very well is when they’re asking us, “What do we think is coming?” You’re right, we have no idea what’s going to be released in the next year, but we do have an idea of what will come out in the next month or two, and they’re building for that.

DAN HILL: Great.

KATE JENSEN: They’re always trying to stay one step ahead. Stripe does this very well.

DAN HILL: Great. Well, in the final minute, any other thoughts, words of wisdom, ideas, things that you think would be useful for folks in this room to take away?

KATE JENSEN: I was at Stripe for eight years. I left three years ago. I’m very familiar with Sessions. I know this is a very, very technical audience. It’s funny, two years ago when I was here, I asked people to raise their hand, who was using AI at work? And it was about 20% of the room. Now, I guess if I asked you this now, please just humor me and do it. It’s almost all of you at this point, right? And it’s amazing how quickly the technology’s being adopted. The one thing I would challenge each of you though is I bet your jobs day to day haven’t fundamentally changed. And I think that that will not be true a year from now. So the challenge I have for everyone here is go figure out how to make that true sooner rather than later, so it’s not really decided for you.

DAN HILL: Thank you so much. Thank you, Kate, for joining us.

KATE JENSEN: Thanks for having me.

DAN HILL: And my final thought, if you’ve been inspired, want to give your agents a way to pay link.com/agents to get started. Thank you.