Stripe's AI Chief: How AI Agents Will Buy, Sell, and Pay

Yeah, I mean, a year ago we were talking about agents as buyers in a pretty hypothetical way. I think that the canonical consumer experiences, for example, weren't defined. We were largely reasoning from first principles about what this might look like. Fast forward a year, as you note, still early innings, still a lot to do, but we have actual infrastructure deployed. We have real companies building on it. We have real patterns to learn from. And to be clear, the shape of how this unfolds has become more clear, and it's going to continue to evolve.

So specifically, what we've come to believe is there's a full spectrum of how agentic commerce plays out. And that's actually really important for businesses in how they think about it. So at one end, and this is where our machine payments protocol lives, basically, you have agents that are out autonomously discovering a service, deciding to buy it, and handling the transactions entirely on their own, right? No human in the loop. Maybe what people think of when they say agentic commerce, but that's just one end of the spectrum.

There's also the whole other end of the spectrum where people are looking for shoes for flat-footed runners inside an AI surface, and the AI surface gives you an answer. And increasingly, that was true also in traditional search, but now increasingly that answer comes with a buy button. And so this is already how a huge number of people are discovering products. If you're a business, you need to show up there. And we've been building the infrastructure to make it, A, easy for businesses to show up, and, B, easy for agents to execute those transactions.

So you asked about what's become more concrete. We recently partnered with Google so merchants can sell right inside AI Mode in the Gemini app. So maybe you shop at JD Sports, because I was on the topic of running shoes, or Fanatics, or Quince. Those were all early adopters. Microsoft and OpenAI, we're doing something similar with them, helping businesses make their products discoverable inside Copilot and ChatGPT. Meta's another example, a little bit of a different flavor, but we're powering checkout right inside ads.

So there's the discovery, and then the one-click, and the agent actually goes and executes the transaction on your behalf. But really, I'd say what we've learned over the last year, the through line is whether we're talking about fully agent-led transactions with ACP or these very human-led purchases inside AI surfaces, there's just a new set of infrastructure that needs to work no matter where you are in the spectrum. And that includes, I mean, it's the premise behind our Agentic Commerce Suite, but briefly, businesses need to be able to expose their products and their catalogs and their prices.

And then consumers need to be able to authorize agents to pay on their behalf. And then agents need to be able to safely execute that transaction. And so that is the infrastructure we have built. Those are some of the partners that we've been working with. And I would say the companies building on it probably give you a good read of where commerce is headed, right? So companies like Wix and Shopify and BigCommerce and Commercetools, sort of on the platform side.

And then, on the brand side, Best Buy and Coach and URBN and Kate Spade. But again, it's still early, and what the interaction patterns will be and how they'll evolve. I'm particularly interested in how quickly consumers will give up more of the decisioning process and really move from consumer happening over here, where AI helps you find the product and the agent is the one sort of helping you avoid going through cumbersome checkout flows, to a time when we say, I don't know, I have a $500 budget for back-to-school shopping, and you already know everything about my kids and their school and where I live.

So, get it done. Or I don't even tell you, and you just go do it for me. But I think we'll need to learn that over the coming year or two.

We do, literally. Yes. Oh my gosh. It's like, we have level 1, level 2, level 3. And basically, if you think about it, the highest level is sort of the MPP version that I talked about, where the agent is truly autonomous, and sort of the level 1 is the human does basically all of the decisioning themselves, and it's the simple execution of the transaction. And I would say, sort of on the consumer side, we're mostly hovering level 2.

People are delegating a little bit of the selection or leaning hard on the AI to help find the product, maybe a hint of level 3. But we're not in the world where you're booking your summer vacation, one-shotting it with an LLM.

Totally. Or you're in Gemini, or—exactly.

Yes, we built it in partnership with OpenAI. The Agentic Commerce Protocol is just a standardized way for businesses to work with agents, and there's a couple different components of it. And this is sort of wrapped in our broader Agentic Commerce Suite. One is: how do businesses expose their product catalog, their inventory, their prices to agents? And you could argue, oh, the agents could go out and search or infer, but inventory is a thing that you want deterministically known. And we don't want businesses to need to register their product catalog or register their inventory with every single new agent that comes online, because in the same way you and I like to work with a lot of different model providers, or a lot of different models within those model providers—in many cases, both.

We similarly are seeing businesses not want to place bets on just one agentic surface. They may be selling B2C and B2B. They may be wanting to reach a wide swath of consumers across many different surfaces. And so, Agentic Commerce Protocol lets them expose their product catalog once and then opt into all of the agents who work with that protocol. It also includes the shared payment token. And so this is about making sure that, in the moment of transaction, the agent can securely pass the buyer's credentials over to the seller to execute the transaction.

And these are just tokenized credentials. So, the agent doesn't have access to the credentials in the way that you and I probably wouldn't want an agent to have our credit card. And one thing I love about both the catalog component of this, as well as the shared payment token, is it's platform-agnostic, payment-processor-agnostic. So, all this works. You mentioned that we co-created it with OpenAI. It works with OpenAI, but also other providers. It works if Stripe processes your payments, but you can also pass on that shared payment token to any other PSP.

And for us, this is really about making it easy for businesses to reach their customers where they are, which is increasingly through AI tools, and to not have to reinvent their commerce infrastructure to do that, right? We want to reinvent commerce infrastructure once, and then they, out of the box, can get these sort of new lines of demand.

We've actually seen a ton of demand from brands. So Best Buy's on it, Coach is on it, URBN's on it, Kate Spade's on it. We've seen Quince and Fanatics and JD Sports and a whole bunch more. We've seen a ton of demand from platforms, which sort of—you've probably long thought of platforms like Wix or Shopify or BigCommerce or whatever, you've probably long thought of them as building technology for small businesses to do commerce. And now an important part of technology for small businesses to do commerce is making sure those small businesses are appearing in AI tools and can engage in this wave of agent commerce.

And so Wix and Shopify and BigCommerce and commerce tools on the platform side have all adopted the protocol. And then on the AI side, on the agent side, we're working with all the big ones. So with Gemini and with Google, Microsoft, OpenAI, and lots more coming online in all three dimensions. But we really think of the supply side as a combination of the large brands and the platforms who have the small businesses. And then on the agent side, it's what you would think of as the traditional agent players.

And then there's been some interesting nuance as well. For example, with Meta, maybe ads are just becoming agentic buying too. And so that's been an interesting extension.

So when I step back with my economist brain, I'm like, that would be really efficient, right? Agents are really good at discovery. We've already seen that agents are really good at integration. Agents are pretty good at finding optimal pricing, matching, negotiating. They're incredibly persistent. Their time is worth a lot less than human time, and they can get those back-and-forths done much more quickly. And then they're also actually really good with integrating and actually adopting the thing. So especially if you think of B2B buying, and maybe we can talk about Stripe projects a little bit later too, but actually not just finding the service and negotiating it for the price and contracting on it and buying it, but actually getting all the way to integrating it, using the product.

I think agents are going to help with a lot. And so I'm definitely imagining an economy that is much more efficient because you have agents on the buy side and, as you note, also on the sell side. And they're kind of hyper-efficient on all of those dimensions, which, of course, in the Ronald Coase, Nobel Prize winner Ronald Coase version of the world, would basically just remove frictions for firms to work with each other, would make markets more efficient, would make competition higher, would serve consumers, would spur growth.

I will say today, not a lot of agent-to-agent transactions are happening. So, like you asked at the top, what's become real and what's still in the future? I think that's still in the future. But I think you start with one side, you add the other, and over time, we probably land with both.

Yes. And actually, that is true on the consumption side, right? If we make it easier for folks to discover and transact and integrate, then that will spur growth. And by the way, we're seeing, I think, some productivity in the global numbers from AI, but I think right now it's not primarily about consumption because the numbers are still very small. It's primarily about, like, oh, we're flooding the economy with a bunch of AI capex. That's actually the consumption that's getting pumped into the economy.

But I think that's definitely going to be a driver over time. It is also true that there's a deeper change, which is agents are making it easier not just to buy things, but also to start and run companies. That's a whole other angle. But we see that very solidly in the macro data. I don't know if you've seen U.S. business formations over time, but during the pandemic they surged. That wasn't super surprising.

But then if you look, they plateaued over time, and then they're accelerating again now over the last couple of quarters. And what's interesting to me isn't just that acceleration, but actually the composition. What are all of those incremental new businesses being created? And the incremental growth is coming entirely from non-employer firms, which is the literal language that the Census Bureau uses, but you and I would just call them solopreneurs. And so, the number of solopreneurs who are earning more than $100,000 a year has just gone like this since 2022.

And now there's, in America alone, 5 million people making their living running solo companies. Not like, oh, I just said I was a solopreneur. Literally, that is my income supporting my family. And there are hundreds of thousands that are clearing a million a year. And so I think that's kind of interesting, because it's like, with AI, can you build something? And then with AI, can you run the business around it? And I think vibe coding and vibe deploying, by the way, are really important for, can you build something?

And then there's a bunch happening in AI, domain-specific agents that are really solving for, can you run that business on the accounting side and the customer support side? And that's making these smaller companies... So anyway, I think the economic enthusiasm I have around AI comes somewhat from the efficiency of markets and growing consumption and better matching and so on, but just as much, if not more, from the effect AI is having on business dynamism and the ability for individuals with an idea to get from an idea to a product that is in market and meeting real needs.

User needs. So anyway, I think they'll both play into the macro numbers increasingly over the coming years.

Yeah, I think the two primary blockers that we'll need to move through to really scale this up are, one, trust. And actually, we've done a lot on the trust side. We talked about the shared payment token. The agent doesn't have any access to the credentials. Every shared payment token includes Radar scores, right? Both, is this a legitimate buyer, and is this an agent acting in a legitimate way on behalf of the buyer? Maybe we'll talk a little bit about Link as the wallet for agents, but we've done a lot so that consumers can set guardrails around what the agent can spend, right?

So it's a little different than, like, a one-time-use virtual card, which are, like, pretty maniacally scoped credentials. But in the case of a Link wallet, you very much have the guardrails to set. But even with the sort of trust layer from a technology or infrastructure perspective, I just think it takes time for any market to build trust, especially when you're talking about making decisions for what I buy and spending my money. I think it's very natural for humans to kind of want to build their way up to that.

And so I think that's a big reason why, on the consumer side, what we're mostly seeing is people are discovering things inside AI apps, but they're still choosing the exact thing, and they're still disproportionately buying low- to mid-price stuff. And they'll need more trust and, honestly, also to some extent, an evolution of the user experience in some of those apps if they're going to get to a place where they're handing off major decisions. And by the way, I wasn't super close to how people moved their spending from stores to online, but I bet in the first few years of spending online, nobody was saying, like, "I'm going to go online and buy a couch or a mattress or a leather jacket," like a thing that I want to feel or I'm going to spend a lot of money on, or where quality is sort of hard to infer from things I can tell on the internet.

And over time, mechanisms built up for people to trust that that was the right product, that when they spent substantial money, it was going to arrive at their door in good shape. And obviously today, there's probably more mattresses bought online than in person. So I think we will get there, but I think trust is an important enabler. And I would say, again, a lot of the technology foundations are in place, but it just takes reps. Like, humans just need reps for trust to be built.

When it comes to financial stuff, basically nobody enters with an assumption of trust. You have to earn it.

Right? There is still a—I mean, I still feel extremely uncomfortable entering my bank account details on the internet. So I'm happy to enter my bank account details into my Link wallet on Stripe so that it can make payments on my behalf, but I am not happy, still today, to enter my bank account number on the internet. And I think this is actually really interesting to reason about: will the internet actually get safer to some extent as these foundations are being pulled out from under us because of agentic commerce?

But as stored balances and wallets and linked bank accounts and whatever become more of the norm, I don't know, I need my agent to be able to burn down my stablecoin balance with some guardrails. Like, does that actually reduce at least the types of fraud that we've seen in traditional online commerce? Maybe, maybe.

I think there's a few things. One, and part of what I'm working on with the Agentic Commerce Suite, is it should be really easy for these AI tools, LLMs, to accurately reflect and comprehensively reflect inventory. Absent a standard protocol, it's actually a little bit tricky to know exactly who's selling what, and is it real, and how's it priced, and how much is left, and what are the various parameters. And so partly the experience, I think, just needs to evolve to actually have the right inventory set to read over and the right deterministic metadata on that inventory set that the LLM can then do non-deterministic things on top of.

But I think some of the experience is just, we're all very familiar with the flow of going in and typing in search and getting some shopping results and choosing a thing. But I think that whole, what part of consumption do we want to delegate and what experience do we need, which is a little bit of the discovery and a little bit of the guardrails and a little bit of the trust, but also just the depth of understanding of us in order for that to be a great customer experience.

There's lots of great AI tools out there, and there's lots of ways that they're accelerating the efficiency of consumption. But I haven't yet seen—I guess what I was saying on experience is—I haven't yet seen an experience where I'm like, that is magical, and I know exactly the end things that I'm ready to offload.

Yes, so maybe even before we get to agents, Link is just Stripe's consumer wallet, and 300 million users use it today. We're making that the wallet for agents. So, the idea is you can authorize an agent to make payments on your behalf, but you get these built-in controls so you stay in the loop, right? So you're not giving the agent, like, a blank check or unlimited access to your checking account. You're very distinctly defining what it can do. And you can always pull it back.

And we talked about trust a bit earlier. I actually think the trust dimension here is as underrated or under-discussed as the product challenge. Like, when I go and ask my friends and family whether they'd be comfortable letting an agent buy things on their behalf, they usually jump straight to, like, well, is it going to overspend, and is it going to buy the wrong thing, and can I stop it, and what happens if it buys the wrong thing? And those are actually all legitimate concerns.

And so, the controls within Link Wallet are really what make the whole thing viable for consumers. I would also say that you and I and others have many different payment credentials, right? Like, we have bank accounts, and we have credit cards, we may have debit cards, and there are buy now, pay laters, and stablecoins and whatever else. One of the nice things about a single wallet that can be backed by balances and many different payment credentials, fiat and crypto and whatever else, is there's just one.

Like, I don't like having a lot of things to reason over, and I especially don't like having a lot of things to reason over when I'm passing it off to an agent and need to monitor it and make sure that what I think is happening is actually happening. And so, I think part of the beauty of the Link Wallet is also just the consolidation. You can back it by whatever the payment credentials are, but at the end of the day, your agent wallet is just a single wallet, which makes it much more streamlined to reason around.

Good question. Okay, so when we were talking 12 months ago, the very, very first version of agentic commerce on Stripe, and I think in the world, was these one-time-use cards. And actually, the genesis of that was one-time-use cards that we used for platforms and marketplaces, right? Like, when I order a salad from DoorDash, the driver has a one-time-use virtual card that can be used to pay for my salad. And that feels good to me because neither the driver nor the restaurant, which I don't have any affiliation with, needs to see my payment credentials.

And so in the first version of agent e-commerce, which for us, I believe the very first meaningfully live volume was on Perplexity Shopping, we used these one-time-use virtual cards. And basically, the human consumer would say they wanted the thing, their payment credentials were used to basically buy this one-time fund, this one-time-use virtual card. The agent was handed this one-time-use virtual card, and then they went off on the internet and purchased with it. And that was very scoped by definition of being one-time.

And it was generally scoped to a single provider and scoped to a very fixed amount. I would think of Link Wallet as much more flexible, right? So I can set a budget to be used across a set of providers or a set of providers in a domain. I can permission at the individual transaction level, or I can permission sort of above some amount in aggregate or individually. And so one-time-use virtual cards were super valuable for sort of getting us off the ground, especially for these sort of single-use consumer transactions.

But the Link Wallet is much more flexible. Now, I still think that we will be a few months hence before folks are actually agreeing to agents using the wallet beyond a particular scoped transaction, but making sure the infrastructure scales to that is really where we are with Link Wallet. And then, as the consumers become ready for it, the technology is already online, and there's no holdup.

It's a new payment primitive. We built it maybe six months ago specifically for agentic commerce. It is a way for a consumer to authorize an AI agent to pay on their behalf without handing over their actual card details. So the token encodes exactly what the agent is allowed to do, right? Which merchants can be charged, up to what amount, in what currency, for how long. The agent presents the token to the merchant at checkout. The agent never sees the underlying credentials.

And shared payment tokens cover more payment methods, including buy now, pay later options like Affirm and Klarna. So it's not just cards; it can really represent different payment methods depending on what the user has on file. And this is actually the payment primitive that makes Link Wallet for Agents work. It also powers our machine payments work. It's really about, okay, how can agents transact without taking on risk, and how can businesses remain in control as the merchant of record?

And so I would think of one-time-use virtual cards as a useful backstop as we went and built the shared payment token. Shared payment token is a new payments primitive that is wallet-agnostic and, by the way, also payment-processor-agnostic. You can pass it over to Adyen or whoever else you have. It doesn't have to be processed on Stripe. And then think of Link Wallet as the consumer experience, the consumer wallet for agents, that leverages that same shared payment token primitive, but is more fully featured for the consumer.

Yes, exactly.

Or you can literally set it to say, "I need to approve every single transaction," right? And when people are making reasonable-sized transactions, that's fine. I think we're gonna move to a world very quickly at the intersection of agentic commerce and what stablecoins enable in terms of microtransactions, and then reasoning about agents as the buyers. Microtransactions never actually really made that much sense, even in the context of content, because nobody wanted to put in, even if it was only 5 cents, no one wanted to put in their credit card to buy a 5-cent article.

That was just too much friction, A, and B, nobody wanted to process a credit card for 5 cents because you would have negative margins. But I think when we enter a world where the human is not doing any work to execute the transaction, you're not typing in any credentials, you're not navigating to any webpages, the agent's doing it for you, then microtransactions become viable. And you pair that with now it's burning down some stablecoin balance, and we can talk about the work we're doing with Tempo there.

Suddenly, microtransactions become very viable. And I think we're gonna quickly move to a world where, especially in buying inference or tokens or AI SaaS-y products, you are going to be using your Link Wallet to make a ton of microtransactions and not approve every one. But yes, if today you want to approve every one, you can. That's not gonna scale because no one's gonna want to approve the 1-cent transaction for the little bit of data or the little bit of research or the token that I needed for this job.

Yeah, we're thinking about this on a few dimensions. One is, to your question on what if the wrong thing gets sent or the thing's not good or whatever, businesses have to remain the merchant of record. That's a core design principle for us, which means we see our job as having agentic transactions behave the way human transactions do. Now, that doesn't mean that agents need to behave like humans or be constrained in the same way humans do, but from the perspective of the business, they remain the merchant of record.

They are selling; they happen to have that sale facilitated through an agent, but the business is still the business. Another design principle, which we've touched on a bit in different ways, is always provide appropriate, granular, programmable controls and guardrails to enable commerce to happen at scale. And that's basically saying, no matter how scoped in and in control you want to be, or how much you want it to be a free-for-all, you shouldn't have to adopt a new tool, move to a new wallet, change the underlying payment rails.

You shouldn't have to change your product catalog or where you expose it. So basically, out of the box, those controls and guardrails should ebb and flow. And I think in the consumer case, that looks like consumers remaining in charge of how Link authorizes agents to buy. And in the business case, I touched on this briefly in the context of Shared Payment Tokens, but it looks like making sure that the information we have about the goodness of the underlying buyer and the agent operating on that buyer's behalf are passed fully to the business to action intelligently.

And so, Radar's our fraud protection product. We've had it for over a decade. It used to be really about transaction fraud. Now it looks at all kinds of fraud and abuse and bot and multiple layers of goodness where there's not just the end customer, but also the agent operating on behalf of the customer. And so those Radar scores are actually included out of the box in the Shared Payment Token for businesses to reason about it. But basically, I think business remains merchant of record.

Consumer has as fine-grained guardrails and controls as they need. And then in the ecosystem, there is as much symmetric information as we can create. I mean, we now look across just about 2% of global GDP. In the context of AI, we look at a very, very large share of that GDP because basically all AI buyers and all AI sellers are on Link and on Stripe. And so that allows us to actually understand when something is going sideways, and we see very much our job to protect the ecosystem by providing that information upfront.

Yeah, not just in purchasing, right? Who is at fault for not just bad purchasing behavior, but bad behavior of all types when there's an agent involved? So I do think the landscape is changing there. When it comes to payments in particular, you mentioned a bit ago that back in the day, people didn't really trust just putting their credit card on the internet. I think it's interesting to think, with things like shared payment tokens, the agent never sees the underlying credentials.

Each transaction is scored in real time by Stripe Radar. Merchants get to handle the transactions. Credentials never travel through untrusted services the way they can when a human types a credit card number into a random website—not Stripe, but other random websites. And so, anyway, I think it's also interesting to ask: is there a meaningful payments trust and safety upside here? And I could imagine, yes.

Yes. Done right, in the limit, I believe it should be much safer. Today, I think the first-order effect is it's just more convenient, right? Usually it's not a fraudster on my credit card; it's me fat-fingering my own CVC. It's annoying, and I'm frustrated. But right now, I think most of what we're getting is convenience. But in the limit, yeah, you could totally imagine, okay, now nobody's passing these random payment credentials over the internet. And, by the way, people also aren't really necessarily reasoning about transactions coming in through 10 different payment credentials they have, right?

They really have a wallet, they see what happens in the wallet, and that wallet is used in a very tokenized, secure way across sellers.

It's what happens after you build the thing. And it's actually not talked about enough. But the coding part—a lot of people talked about AI for coding when we were talking a year ago—that makes sense because it wasn't solved. Now that's basically solved. And we actually see that it's solved in our data. These are semi-random facts, but I find them interesting. So, agent traffic to Stripe's documentation grew more than 10x since we talked a year ago.

Agent traffic is now about 40% of all our docs traffic.

Meaning the coders now are almost as much agents as they are developers, and in some segments they are basically all agents and not at all developers. And another example, actually, is our CLI, our command-line interface, which historically was a pretty niche tool used by a pretty small group of developers. It just exploded, and we were like, what is happening? And now 70% of its API resource requests are from agents. So you could think the majority of entities that are using our CLI today aren't people.

So anyway, I think those are just two fun Stripe anecdotes, but there's lots of them across the ecosystem you can look at that tell you vibe coding is real, right? Just your own lived experience is the same, right? That part worked, but then what? So an agent writes you a complete working application in 20 minutes. Fabulous. We love it. Okay. But you've still got this pretty big friction before that app is actually live, right? So you've got to go, I mean, it depends what you're doing, but you've got to probably create an account with your database provider, then your auth provider, then your hosting service, and you're bouncing around.

You've got all these dashboards. I don't know how you do it, but copy-pasting stuff by hand and managing credentials and API keys and whatever else. And every one of those services has its own onboarding flow. So, a little different flavor than the payments flow we talked about, but it's a pretty inefficient fill-out-a-bunch-of-steps. And every one of them, just like the payments flows, was designed for you and me as humans sitting down and clicking through this weird setup wizard thing.

And I don't think it really bothered any of us that much because we weren't doing it that much, because the hard part was the coding part. But now that the app can be built and coded in 20 minutes, okay, the long pole is deploying the thing. And so anyway, vibe coding was easy. Vibe deployment has become more of the binding constraint. And so we actually recently, in the last couple months, launched Stripe Projects for this. But basically, it's like agents should be able to sign up for, configure, and integrate all of the services they need to deploy an app, and they should be able to do that right from the command line.

And it's not just Stripe services, right? It was like a whole ecosystem of, like, you get Vercel and Supabase and Cloudflare and Twilio, like whatever you need. Clerk. I think we announced, like, 16 more partners a week or two ago now. And anyway, there's still a lot of enthusiasm from the ecosystem because it turns out that everyone kind of has the same problem. Developers all have the same problem, which is now, like, the long pole is the deployment, and businesses have the same problem, which is, like, their onboarding was designed for humans, and now they need it to work for agents.

Okay, so I think there's a lot of things that Stripe cares about that on paper you'd be like, why does Stripe care? The honest answer is we care because it was becoming the bottleneck. Like, the barrier to building is gone, but the barrier to deploying is, like, a real friction. And if you just zoom back, like, okay, separate from being a payments company, our mission is to increase the GDP of the internet. But a big part of that is: how do we get more companies off the ground, and how do we help them get their first dollar faster and scale from there?

And so if a person with an idea can create an app but can't deploy it, they can't sell. And so removing deployment friction directly obviously expands the internet economy. And you can think of it primarily as just orchestration. All we're doing is orchestration. If someone else wanted to and could and did orchestration, we'd be cool with that. But we were looking at the developers trying to get the thing live, and we were looking at the businesses trying to get the thing used by developers, and we're like, okay, I think we can make this market a little bit smoother.

Yes, exactly. Exactly.

That's a big question, because what does it mean to monetize tokens from a world perspective is kind of like, okay, actually, how's the whole next generation of B2B and some B2C going to monetize? Look, the last decade-plus of SaaS had pretty beautiful and simple monetization economics, right? And in particular with SaaS, you build a product once, and then you get one more customer, and it costs you basically nothing to serve them, right? Marginal costs are near zero. And that's why SaaS margins are really good.

And that's why the fixed-fee subscription or seat-based license model works really well in SaaS. AI—and I say AI generally, because you could literally be selling LLMs, but you could also be selling some product that's a wrapper on top of LLMs, or a product that's heavily powered by LLMs and requires a lot of tokens—it breaks that model because, obviously, every prompt and every API call and every task has a real marginal cost all of a sudden, which it didn't have in SaaS.

The inference isn't free. And so you now have all these businesses where how your customers use your product directly determines whether you make or lose money. And that's a very different game. And from our vantage point, part of what that boils down to is the need for usage-based billing as really critical for AI companies. You need to be able to meter what customers are actually consuming in real time and then charge them in a way, obviously, that aligns with your—

I see very few scaling or scaled AI companies that are still exclusively subscriptions or seat-based. And I believe, from my conversations with them, that that is for the simple reason that the economics there don't make sense, because you have some people that are using a ton and some people that are not using very much. And these people cost you a ton, and these people cost you not very much. And it is very hard to separate the sheep from the goats and price them appropriately without a usage-based meter.

That said, many of those businesses have a usage-based offering, but it is complementary to what you can think of as a fixed-fee subscription. So a great example is Lovable. When they launched, they had a simple subscription through Stripe Billing, which makes total sense. They were early, they're moving fast, they wanted to monetize quickly. Subscriptions are also very familiar and easy for consumers. If you think about Lovable, some of their target users are not very technical. How are they going to feel about, like, oh, how should I reason about a token or a credit, right?

So they started with subscriptions, but then as the company grew and some of their costs grew, their billing needs evolved, and they needed to charge at least somewhat based on actual token consumption. And so what they did is a hybrid billing model, which actually we're seeing a large number of businesses, especially businesses that have a B2C component, do, where they have usage-based billing on top of their subscriptions. So customers hit some threshold, and by the way, many people have freemium thresholds, but then above that they often have, like, a fixed fee, $25 a month or $100 a month, up to some number of credits.

And then above that, you have usage-based billing. So once customers hit the threshold, Lovable, in this case, charges precisely based on the number of tokens consumed above that. And I think that alignment is good because you get a bunch of people in the door comfortable with subscriptions, but then sort of at any volumes that matter, your revenue is scaling very directly with how customers are using the product, and correspondingly with how much it is costing you to provide that product.

So customers pay only for what they get, and you make sure you monetize for the underlying cost that you're going to have to bear. And ElevenLabs is another example. They went through the exact same thing, literally started with subscriptions. They recently moved to this sort of pay-as-you-go plan, and this is just a pattern that we see playing out pretty much everywhere in the AI space right now.

So I think so, for a couple reasons. One is agents can consume at machine speed. And so even if you're doing a usage-based thing, but you're charging at the end of the month, by the time you've gotten to the end of the month, an agent can have spent—I bet a human can too, to some extent, with agents working for them—but especially an agent can have spent an egregious amount. And so actually, I think in the world of agents, what we're going to see more and more of is real-time metering, like, what have you consumed, and real-time billing, which is actually what we have co-built with Metronome.

So, real-time metering, usage-based billing for very complex models, and Tempo blockchain, where agents are consuming tokens in real time and paying down the cost of those tokens. And that's important because agents can buy at machine speed. It's viable because of a bunch of the infrastructure between Tempo and Metronome. And agents are happy to just pay as they go in a very literal way, plus businesses need them to be paying as they go so that they don't rack up a bunch of spend and then go dark.

Totally. We have a revenue recognition accounting product, and it's mostly needed by—we provide it to a bunch of businesses who have traditional subscriptions. But where there's the most acute pain is actually, traditional accounting in spreadsheets does not work when you have this proliferation of rows because these microtransactions are truly happening. I also think it's changing what it means to be an accountant. I will avoid naming the company, but I was talking to the sort of number two in accounting at.

A pretty successful AI company. And a couple things were interesting. One is they were like a hybrid accountant-engineer, which they needed to be because of the scale of the data they were dealing with. And two, they weren't doing accounting in the traditional you-and-me sense of just closing the books. They have to close the books, for sure, but their job was also to find weird anomalous stuff happening. And I was talking to them because they were identifying some fraud patterns and wanted help with them.

But anyway, just very interesting. What does it mean to do accounting at one of these AI companies? Definitely needs new tools, definitely needs new systems, probably a different skill set. And then what you're accountable for really isn't just closing the books. It's looking across the whole revenue recognition stack and saying, what does this tell me about the health of our business? What does this tell me about fraud and abuse in our business? And what does this tell me about breakages in our product?

Yeah, it's a new form of fraud. It's, I think, one of the most underdiscussed topics in AI right now, maybe by a lot. Fraudsters have figured out that in AI, you actually don't really need to steal money or credentials. You can just steal tokens. And tokens have real value, right? You can use them to build things. You can resell them on marketplaces. You can wrap a new product on top without paying a cent and go and sell that new product.

So it's kind of like resale, but it doesn't look like you're selling the subscription to the thing. It looks like you're selling this other product that you've come up with on your own, but really, in the back end, it's completely powered by someone else's thing. And for AI companies, maybe this is implied by our earlier conversation on SaaS, but the risk is existential, right? If someone stole a little bit of your SaaS, it didn't matter because it didn't really cost you anything on the margin.

When someone steals your tokens, if your fraud rate is high, the economics of your product actually break really fast. So anyway, because we work with basically all the AI companies, we have this interesting front-row seat into what's actually happening and the fraud patterns playing out. And right now, there's three, and I don't think these will be the three forever because we've already largely burned down these three over the last three months. But new ones will pop up.

I think these three are just instructive for reasoning about sort of the breadth of the space. One is multi-account abuse. So this is like bad actors just hammering your signup again and again and again and again so that they can get those new-user credits. And the scale of this actually shocked me when I looked at the data. So more than one in six signups at AI companies are this kind of abuse. And by the way, that's also very confusing for the company.

Like, are these good customers? Who are these people? But it's very expensive for the company because I've spent a bunch of tokens on these free products for a relatively small number of people who are spinning up a massive number of accounts and actually consuming quite a bit of tokens. Okay, so that's very, very top of funnel. Another example that I think is interesting is free trial abuse. So these fraudsters come in, they create a free trial, they put down a payment method, they drain through all of the credits, but they never have any intent to convert.

And free trial abuse has always existed in various forms across the internet, AI or not, but it's more than doubled on Stripe in the last six months. Most of that doubling is coming from AI and AI businesses. And just to give you a sense of how lucrative this is, there's whole cottage industries built around it. So I don't know if you've ever been marketed a free trial card. I was recently marketed one. It's basically like, literally, it's like, oh, you can spin up free trial cards.

They expire in 24 hours, so you'll never have to pay. And if you or I had one, we'd probably use it for some legitimate purpose to try out a service and just not have to go through the pain of canceling. But literally, fraudsters can just explode these things and spin up a bunch of free trials and spend a bunch of tokens with no intent to convert. And then the third is a little further downstream of this, which is basically, like we talked about, the usage-based billing: folks are racking up thousands of dollars of costs and then being billed at the end of the month and never paying.

And so, like, whatever, the dine and dash that happens in restaurants is like a dine and dash, but it's for tokens. And unfortunately, at that point, again, this is AI, not SaaS. So the company has borne the costs of those tokens.

Yes, exactly. Okay, so what you do with it very much changes based on what's the service, what exactly were these tokens meant for? You hit the nail on the head for, okay, you just get the tokens from the underlying LLM, fine. For those businesses that are a layer above that, a bunch of resale abuse. So you literally sell it sometimes in other markets for a slightly discounted price, a discounted price on the base price that you should have paid but you didn't pay.

Yeah, exactly. So just say, I give you my login credentials, whatever. But then there's just this long tail—makes it sound small, but very domain-specific fraud pattern. So, for example, people steal tokens to create content that they then use to extract money in all sorts of scammy ways. So a simple example: we see people going in and mass-generating music tracks and then uploading them to Spotify and Apple Music and then getting fake streams and then collecting royalties.

Or then for basically all of the wrapper businesses, there's this whole other layer of wrapper on a wrapper where, rather than just resell the subscription on places like Taobao, they'll literally clone the AI company, right? You can just vibe-code a website. The backend is just spitting out exactly what you got from the service that you're stealing from. And then you sell the product as yours, but cheaper.

Yeah, I don't know. I mean, they definitely put a lot of time into it. Tokens are also very valuable. And what's been interesting, as we sort of started seeing these trends maybe at this point six to nine months ago in various flavors, but then they escalated a bunch. Talking to AI companies, the large AI companies are all over this. It's one of the most existential things for their margins. They have been in the trenches with us identifying the issues.

They have been like, literally, we see a problem, we build a model: multi-account abuse. Okay, at the time of login, there's an API. You send us what you know, we send you back a score, you block if they're bad. At the time of free trial start, same thing. As people are accumulating usage, you send us all the metadata, we send you back whether they're fraud-y, and you can require a top-up or cut off service or whatever. Literally each of those, as we've seen them, we've, like, in the order of weeks gotten—generally this isn't even like a full-fledged product, it's like an API.

Like, you send us some stuff, we send you some stuff. And the adoption has just been like, okay, so the AI companies are all over this. I think what's interesting is every company is going to become an AI company. And I don't think the industry at large is thinking about this yet or has really even reasoned that, like, actually, most of the fraud that's happening is not traditional credential or payment fraud. It's like what we would historically have called first-party abuse, right?

Like resale abuse or account sharing or multi-accounting or free trials, like first-party abuse. And I think it wasn't that first-party abuse didn't happen before, it's just at least in sort of SaaS stuff, first-party abuse didn't cost you anything. And so, A, it wasn't that useful to get a little bit of Salesforce for free. The examples we talked about wouldn't be relevant in most SaaS. And B, even if it was useful to figure out a way to skim off the top, and you could get a little bit of profit for it as the fraudster, it didn't actually cost the business that much because their marginal cost was zero.

And so, as every business becomes an AI business, I think we've been, in the context of our work on Radar, really reasoning about our fraud prevention product as moving from transaction to full customer lifecycle and moving from traditional fraud to end-to-end abuse. But I don't think the whole industry is there yet. And you and I talked about much of the economic upside of AI, but I think that'll really only be realized if it can happen safely. So, for example, six to nine months ago, I was talking to some of these AI companies. They'd be like, oh, I know, I'm gonna solve my free trial abuse problem by cutting off free trials.

Like, I'm gonna solve my free trial abuse problem by only having a sales-led motion and only going after enterprises and not having PLG. And, like, I hear less of that today.

Not because the fraud's totally solved, but because everyone knows they need agents to also be their buyers. And if agents are going to be their buyers, they better have a self-serve motion. They better have a PLG motion. There's no way they want to siphon off that source of growth and sort of only double down on a highly secure sales-led motion. But it's been interesting to see what's happened with token theft. And I totally agree, the fraudsters are creative, but I think that's a manifestation of how valuable the tokens are.

Yeah, it's real-time, it's AI-driven. And then I think actually most importantly here for its differentiation is that it just looks across the Stripe network. So there's basically no good AI buyer we haven't seen before, and there are very few bad AI buyers we haven't seen before. And so that combination, it's, yes, the size of the network, and there's 2% of global GDP flowing through Stripe, but really, when it comes to AI, it's the density of the network. And we talked about Link briefly, but to give you a sense, Lovable's a good example as an AI company.

58% of Lovable's volume flows through Link. Link is an extremely dense network when it comes to AI. You can sort of extrapolate with them. But if we know who all the good buyers are and we've seen the bad guys be bad somewhere else, then combine that with AI and real-time APIs, and you get pretty good fraud defenses.

So I think there's a couple components of our work with Tempo that I think are really interesting. So when you think about agentic commerce on the business side, we launched the Machine Payments Protocol, or MPP, and we built it with Tempo, and it's an open standard. And the way it works is quite elegant, right? An agent requests access to a service, and it can be whatever, an API or an MCP server or whatever. And then the service responds with a payment request.

And then the agent pays, and there's no account creation, checkout UI, human in the loop, or the way you and I would traditionally engage on the internet. It's just this very machine-readable, standardized way for agents to buy from businesses. And MPP is really the primary mechanism that we're seeing businesses use for agents to buy from them, agents as buyers. The other collaboration with Tempo that I am super bullish on is more related to fraud because agents are increasingly becoming the users of AI products, and agents can burn through tokens very, very quickly.

And so we were talking a little bit about this dichotomy that a business faces where either you can turn off self-serve and be really safe, but grow slowly, or open it up, including to agents, but then be putting yourself at risk of quite a bit of abuse and monetary losses. Neither of those is great. What you actually want to do, especially when the agents are the buyers, is track the tokens as they're consumed. And you mentioned the infrastructure, you want to track them in real time at substantial scale.

And then, as importantly, you don't just want to track them as they're consumed, you actually want to collect payments as they're consumed. And so we call this streaming payments, and it's what Metronome and Tempo, which is this blockchain optimized for payments that Stripe helped co-build, are making possible together. Metronome's job is to track the usage in real time, and Tempo's job is to enable fast, low-cost, high-volume micropayments that settle instantly. Obviously, they settle in stablecoins. And so put together, AI companies can charge their agent buyers as tokens are consumed instead of having to choose between closing off business or getting stiffed on the invoice.

So we're very bullish about Tempo and stablecoins in general in the agent economy writ large and for the purposes of agent commerce.

So it's interesting. When we talked last year, we talked about the growth of AI startups and how they looked different than traditional startups. And what I would say today is there's definitely a delta between AI startups and non-AI startups, but what's more striking to me is how AI is just changing the startup ecosystem generally. And so, in the vein of vibe coding and vibe deployment and all of that, new business registrations are—well, I think in general, the business formation story is underappreciated.

So new business registrations are up basically around the world, at least for advanced economies. They're up like 40% in the Netherlands and 70% in Finland and 80% in France. And so there's this sort of surge in dynamism that, yes, we see and feel and hear in the U.S., but it's happening across advanced economies. When we look at this with the Stripe lens, the pace of new businesses launching has doubled since we talked last year. And not all those businesses are AI-based businesses, but many, many of them were made possible because of AI.

And they're not just getting started; they're also scaling. So Atlas is our product for founders to incorporate, and Atlas startups from the 2026 cohort—and it's only June, so it's early in their lifecycle—are tracking to like five times the revenue of last year's class at the same number of months.

Five times. And some of that is there's more of them, but a bunch of it is they are getting to their first dollar faster, and then they are scaling up more quickly. And getting to their first dollar faster, a lot of that is for sure AI. And then scaling up more quickly, a big part of this is actually how they're going global. And I don't know how much of that is AI or not, but the old model people had was you get big, and then once you're big, you deserve to go global.

And what we're seeing increasingly over the last year is you literally go global from—it doesn't necessarily literally mean every country—but it's like you're in dozens of countries on day one, like your launch day. And that is how you get big. You get big by being global. And so, I guess we're talking about AI, so I could use Emergent Labs as an example. AI platform for—you build and deploy these kind of full-stack apps. So they were founded in 2024 in the U.S.

Seventy percent of their revenue comes from international sales, and they do material business in 16 countries. A substantial share of their revenue comes from many, many countries. So anyway, I think there's this sort of, yes, there's AI companies, and it's moved from just being sort of the underlying providers to a lot of wrapper businesses, proliferation of wrapper businesses across every single vertical. But I think what's kind of more interesting from the top-down macro perspective is just, you used to have to be a developer to be a builder and therefore build a business.

Now you kind of have to have an idea plus vibe coding, plus vibe deployment, plus reasonable economic infrastructure. And then we see this proliferation of new businesses, and their development is not easily arrested, right? Even with one employee, they start monetizing early, they grow quickly, they expand across a bunch of markets. And I think that's almost certainly helped by all of the operational work that can also be done with AI, which is less in the day-to-day core wheelhouse of Stripe.

We help with some of that on the accounting or RevRec or whatever, but much of the customer support and other operations are obviously done by other businesses. But I think it's an interesting time, and there's a lot of discussion of, like, is AI going to lead to a small number of firms with heavy market share and not a lot of competition? And I think one of the reasons I'm bullish on the AI economy is, at least so far, for sure there are big guys who have things that are highly complementary to AI.

We don't need to go through that list. Everyone knows them; they're exploding. But there's also an explosion of little guys coming into the market and not just being created, but reaching customers and growing quickly. And so I think that bodes well for competition and for economic growth.

So, we've all read the stories of companies who have accidentally gone bananas on token spend because they had no control over what their employees were doing. I think the companies who have truly gone bananas are, by and large, the companies with pretty deep pockets, which isn't to say that it's not going to be a problem in the economy if that spending continues. But there are smart companies that are well-run, well-managed. They can handle a month or two of poor decisions and runaway costs.

And by and large, I don't exactly know what we're talking about at these companies, but if we're talking about, I don't know, 2%, 3%, 4% of their headcount costs going into tokens, and 30% or 40% of that is inefficient, I think they can pretty quickly get back to an efficient frontier, and it's not really anything existential. For the little guys, I actually don't think that's happening. Many of them are sort of making fixed-fee purchases, and they're still on small plans and whatever.

So the stories are out there. And by the way, just to be clear, I think we have a lot to learn about the efficient frontier of AI use. And there's a component there around model routing and what model do you select for the job. And there's a component around just observability, which I think many companies have found themselves to be behind on. And then there's just norms and controls and guardrails. And I think we all want high-ROI usage of LLMs, but we also want employees to know when they're racking up really substantive costs, and we want them to be able to tell us whether they anticipate the ROI of those costs is going to be there.

I think there'll be a little bit of recalibration, but I don't think anything existential has happened that's going to annul the sort of long-run upside here.

I think the most interesting thing in the next 12 months is—I think we'll move up, and I don't know if it's going to be to four or to five, but I think the more interesting thing is actually that we talked about agents as economic actors mostly in the context of buying, but I think we'll start to see—and I'm not saying this will proliferate everywhere—but I think we'll start to see agents that are multifaceted economic actors.

They're buying and they're selling and they're provisioning infrastructure and they're running businesses and they're doing the thing end to end. And again, I don't know how many of these there'll be or how they'll operate or whether it'll be with each other or in these weird niche silos. But I think all of that's just going to continue to demand more purpose-built infrastructure, including financial infrastructure, versus just the sort of old human-centric commerce stack. And so that's kind of where, like, what does it look like when, okay, I'm vibe coding and I'm vibe deploying, and then my content, my offerings, are by default exposed.

And actually, we just quietly went to public preview on Stripe Directory, which is just—

A really easy way for agents to discover providers, and then through Stripe Projects, they can integrate them directly. But then an agent's also discovering everything and integrating it and buying it, and then creating a service out of the combination of things it has provisioned and integrated and bought and is starting to sell a thing. Anyway, this idea of an agent as a micro firm, I think, would probably be the most interesting thing to see 12 months from now. And again, I don't think the median firm is going to be—or forget a solopreneur, what would it be?

A solo agent. I don't think that's the world we'll live in, but I think 12 months from now we could totally see some examples of that that sort of pave the path for...

Like, the whole thing end to end. And this always happens with a new technology, right? You take your current processes or market or whatever you have, and you figure out how does the new technology make that 5% more efficient or 10% more efficient, keeps me from having to type in my credit card number, right? But where it actually gets interesting is where we start to reimagine how the system works. And it's not Emily permissioning an agent to buy on her behalf.

Emily has an agent who's tasked with running a business, and that includes buying some things and selling some things and making some profits. And that'll be the world that I would like to be talking about 12 months from now.

Awesome. Thank you.