Transcript: ‘How Stripe Is Building for an Agent-native World’

The transcript of AI & I with Stripe’s Emily Glassberg Sands is below. Watch on X or YouTube, or listen on Spotify or Apple Podcasts.

Timestamps

1. Introduction: 00:00:45
2. New rules for an agent-driven economy: 00:01:27
3. Compute theft is the new payment fraud: 00:03:57
4. How Stripe expanded fraud detection from checkout to the full customer lifecycle: 00:10:00
5. Why AI companies are scaling way faster than top SaaS companies: 00:19:48
6. Outcome-based billing is replacing seat-based pricing: 00:23:27
7. Where AI spending is coming from: 00:29:57
8. How the developer experience changes when agents are the builders: 00:36:45
9. The agentic commerce spectrum, from assisted buying to autonomous purchasing: 00:41:00
10. Meet Link, a consumer wallet for delegated agent purchases: 00:51:06

Transcript

Dan Shipper

Emily, welcome to the show.

Emily Sands

Thanks so much, Dan.

Dan Shipper

Really excited to have you. You are the head of data and AI at Stripe, and I feel like this is such a good time to have someone from Stripe on because you all famously are increasing the GDP of the internet. The internet is changing so much right now, and therefore the economy of the internet is changing from something where humans are buying and selling from each other to an economy where agents are buying and selling from humans, and agents are buying and selling from each other.

I feel like I want to know what that means for Stripe. But I want to understand, since you have this macro view of the agent economy, what does that even mean? And what are you seeing?

Emily Sands

A big shift I think we’re in the midst of is that the internet economy is becoming more autonomous. For a long time—for forever—the internet was built around an extremely simple assumption that the main actor was a person sitting in front of a screen. They’re browsing and they’re filling out forms and clicking through checkout. But also they’re writing code and setting up tools, and that assumption is starting to break in various ways.

Sometimes the human is still totally in control, but they’re interacting through an AI interface instead of through a website or a traditional app. Sometimes the agent is acting on their behalf. And then sometimes software now is just out interacting directly with other software. As all of that starts to happen at all of those layers, a lot of things need to be rethought.

There has been rethinking of how products are discovered and how products are bought, but also what should developer tools look like? In our world of Stripe, what is the underlying economic infrastructure—the payments and the billing and the fraud detection and the identity layer—that’s needed in this world where actors are no longer just humans?

For me, that’s the larger frame of the moment. It’s not just “AI is making search better” or “AI is helping people code” or “AI is evolving commerce on the margin.” It’s really that the internet has this new kind of actor on it. Over time, this actor—these agents—will become the predominant actors on the internet. As that’s happening, basically every layer of the stack starts to need an evolution.

For Stripe, it’s like, okay, how are we getting agent ready? But then also, how are we helping businesses get agent ready? Both of those are happening in a number of ways—yes, in commerce, but also in how builders build.

Dan Shipper

Can you give me some specific examples of the kinds of things you’re seeing? I’m almost wondering, for example—I know at Stripe one of the things you deal with a ton is fraud. I assume there’s a whole new type of fraud happening, but I’m also wondering what even counts as fraud now in the sense that it’s possible that my agent could go steal someone’s credit card and check out. I don’t think that Claude would, but you never know with Grok.

Emily Sands

No comment. No comment. But you’re right that AI introduces very different fraud problems. You asked, “What is fraud?” We used to think of fraud as payment fraud—someone was stealing money, someone was stealing your card credentials.

Increasingly, and I was in a meeting with one of our very large AI users today, fraud now is stealing compute. That’s a very different type of problem. In earlier software models, if you think of traditional SaaS, letting someone into a free tier didn’t cost you very much. And stealing a free tier wasn’t very valuable to the fraudsters. Now, giving someone credits, offering freemium, offering a free trial, letting them rack up a bunch of tokens and pay at end of month—except maybe they choose not to pay—actually is a major fraud vector and an existential risk to a lot of these businesses.

Because in AI, every prompt, every image that gets generated, every API request has a very real cost attached to it. People are talking about intelligence getting cheaper—yeah, but it’s still very far from free. And then when you look at the growth model for many of these AI companies, free compute is the new CAC. You used to spend a bunch on paid media. Now you spend a bunch on your free trials and your credits and your self-serve onboarding as a major lever for growth.

The abuse we see in that context—where compute is the new CAC and compute is very expensive—is threefold. One is multi-account abuse. Bad actors come in and sign up over and over again, creating a new identity every time on a new email address, claiming their new user credits, and staying ahead of detection by iterating across a bunch of different aliases.

Just to give you a sense of the order of magnitude—across the AI companies running on Stripe, about 7% of their signups are these multi-account abusers. Non-trivial share.

The second trend we see as a new vector of abuse is free trial abuse. This is often the most urgent issue because the unit economics break really quickly. We had a large AI company who was seeing only 4% of their free trials convert to paid. Each free trial cost them $25 in LLM spend. So basically it was costing them $625 per payer before the first dollar of revenue was brought in. And when we double-clicked on those free trial folks, the vast, vast majority of them were actually abusers. They were stealing the compute. They never had any intent to pay. These weren’t people who were genuinely trying out your service and then chose not to buy. These were people literally abusing your systems.

Some companies just dropped free trials altogether. Of course, that’s not great because you’re throttling growth. Others responded by blocking virtual cards. I don’t know how often you’ve been marketed virtual cards. I’m often marketed virtual cards—get this one-time-use card, it expires after 24 hours so you never have to pay for the service.

In the hands of a good consumer, fine. In the hands of a fraudster, very much not fine. The problem with blocking all virtual cards is that for AI companies, about 15% of legitimate card transactions on Stripe are actually virtual cards.

Dan Shipper

We use those all the time. For Ramp, for example, we have a bunch of virtual cards.

Emily Sands

Totally. So in the same way you don’t want to be turning off free trials, you don’t want to be throttling virtual cards either. And just for order of magnitude—you can think of exponential growth in free trial abuse over the last six months. It’s four-Xed. And for one large AI user on Stripe, we’re currently blocking 250,000 fraudulent free trials a week.

The magnitudes here are quite high.

Dan Shipper

Is the volume of fraud constant? Is it just shifting shape, or is fraud actually going up because they’re more powerful now because they can use AI agents to do it?

Emily Sands

Fraud’s going up because the fraudsters have AI on their side—although it’s also on the side of the detectors. But also because the value of the services they can steal is higher. What do you get if you steal traditional SaaS? You steal some inference, you steal some compute, you can resell it, you can do all sorts of stuff.

Dan Shipper

Look, I love a good CRM seat.

Emily Sands

Don’t you? Who doesn’t love a good CRM seat? LLMs are for sure more tempting.

And by the way, the third type of new abuse we see is non-payment abuse. You incur overage, or you have 30-day invoicing except you never pay your invoice. In many cases, customers are consuming thousands or tens of thousands of dollars in compute during a month or a day or sometimes an hour. And by the time they get billed and fail payment, that loss has already happened. These AI companies are left holding the bag.

For us, fraud used to be a transaction thing. Now it’s a customer thing. It’s a full-funnel thing. It starts at the time of signup. Is this multi-account abuse? Should they get credit? Is this free trial abuse? Should we give them a trial in the first place? And then when they have overages—should we be throttling them? Should we be requiring top-up? Should we be blocking service completely?

It’s a whole new world because the thing to steal is much more valuable and the cost of having it stolen is much more existential.

(00:10:00)

Dan Shipper

How are you even able to do that? I totally understand how you need to be in that full funnel in order to detect fraud. But my understanding of—whenever we’ve integrated Stripe, it’s usually on the checkout. We’re not necessarily putting you in there when someone puts in their email address for a free trial.

Have you changed the product to do the full funnel, or how does that actually work?

Emily Sands

Yes. Radar, which is our fraud protection product, used to be at the transaction level—at the moment of checkout, as you note. But because so much of the fraud risk was coming up-funnel, AI companies are now increasingly integrating Stripe Radar at the time of signup. We see the metadata at the time of signup, we pass back scores at the time of signup, and every moment subsequently—because fraud is now a full-funnel problem, not a transaction problem alone.

Dan Shipper

If you’re—asking for a friend—if you’re running an AI company and you don’t even know what your fraud rate is and you want to protect yourself from this kind of abuse, what are the top things you need to do to make sure you’re reasonably safe?

Emily Sands

I would just adopt our highest-tier Radar plan. But the actual mechanics of that are: at signup, you want to know if your customer’s good before you give them any access to any credits. You want to make sure they’re good at the time they pay. You want to make sure that charge is good. And anytime they have an overage, you want to make sure they’re good for their money. There’s other stuff around refunds and disputes that we also support.

But I think those are the four major moments in the AI company customer lifecycle where we’re maniacally focused on protecting, because that’s where we’re seeing the biggest cost and the fastest fraud growth.

Dan Shipper

And at each point, that’s just a call to the Radar API?

Emily Sands

Yes, correct.

Dan Shipper

What if I’m sitting here—which I am—doing millions of dollars a year in Stripe transactions, but I actually have no idea what my fraud rate is other than there’s that little thing where it’s—I don’t even know if it’s necessarily our fraud rate. I think it’s our card chargeback rate. Anyway, our fraud rate is low enough as marked for me to not care about it. I don’t really know if there’s some amount of free trial fraud that I’m not totally understanding right now. So what are the things I should be looking for to know if I should dig deeper and potentially do a Radar integration?

Emily Sands

You can go to your Radar dashboard and see if you see anything that looks spurious there. If not, you can also ask the Radar assistant, which is in the dashboard. As you’re doing that, you can describe your business model—you can say, “I have a high marginal cost business,” in which case you care more about certain types of fraud than others.

But you can also just take a stab at integrating up-funnel and see how it performs. We can certainly share with you based on back-testing what we think the big issues are. But the fastest way to get a clean read is just to integrate.

Dan Shipper

Got it. So I would just go look at Radar and turn it on. I don’t think we’re integrated right now. Does it say anything? I’m doing that right now. It would be really funny if I found that we had a ton of fraud that I didn’t know about. We were at 0% fraud. How is that possible?

Emily Sands

Oh no.

Dan Shipper

0.02% early fraud warnings, total fraud rate 0.2%. So we’re doing pretty good, right?

Emily Sands

That’s pretty low. That’s pretty low. I mean, you’re a pretty good human. Maybe the fraudsters don’t want to come after you—until they hear this episode, and then they’ll be like, “Yeah, okay.”

Dan Shipper

That’s really interesting. Okay, so that’s fascinating. I want to go back a second to the AI economy because one of the things you said earlier is fraud is increasing overall on the internet. It’s increasing because the fraudsters have AI, but you all and everyone else on the side of good in the AI economy also have AI to defend against these sorts of attacks.

I think you’re getting an interesting window into the arms race that I think is playing out in lots of different areas that have this kind of threat vector. A really simple one is cybersecurity—not just for payments, but for hacking and stuff like that. But there’s all these other similar types of things where AI makes one part of the process much easier, and then another part of the process has to use AI to compensate, to catch up.

How is that race going? What is that like? What are the early reports that you’re seeing and feeling, being in a race with AI-armed fraudsters?

Emily Sands

I think the interesting thing about fraudsters is they don’t really care about boundaries. They don’t care about whether this transaction is processed on Stripe or off Stripe. They don’t care about whether this transaction is on fiat or crypto, whether it’s on a card network or a buy-now-pay-later. They’re just going to figure out how to work around the system to get through.

One of the important levers—and I appreciate you calling us the good guys—one of the important levers I think the good guys have for winning is to be comprehensive. A simple example in our world: Stripe Radar used to only work for card transactions, and then last year we added ACH and SEPA—other payment methods. But this year we’ve extended to all payment methods that have disputes, and we added crypto. We added the Radar API. So you can screen transactions even ones that aren’t processed on Stripe. You can process on Worldpay or Adyen or whomever, and through the Radar API get the same fraud signals.

Similarly—and we haven’t talked about agentic commerce yet—as we built out our agentic commerce suite, one of the new primitives we designed is the shared payment token, which allows agents to safely pass buyer credentials onto merchants for the merchants to process the transaction. As part of those shared payment tokens, we pass over the Radar fraud scores so that the merchant, whether or not they’re processing on Stripe, can action them appropriately.

When it comes to fraud, we really see fraud defenses and fraud mitigation as a public good. That allows us to invest disproportionately, above and beyond the direct value to Stripe, because protecting the internet is important for growing the internet economy.

I would say overall—yes, fraudsters have AI in their favor. Stripe looks at 2% of global GDP and is growing 34% year on year and sees a broader swath through our multiprocessor solutions like the Radar API. Luckily, not only do we have AI on our side just like they do, but we also have data on our side. The more comprehensive we’ve gone in our fraud protections, the more we’ve been able to eke ahead.

That’s not to say that we’re not constantly surprised by the new creative vectors they come up with, but you can have an agent every day or every hour taking a look at anomalous patterns on the Stripe network and identifying new vectors that are popping up across processors, across payment methods, across merchants, and burn them down pretty quickly.

I’m overall bullish, but certainly not complacent.

(00:20:00)

Dan Shipper

What about other parts of the AI or agent economy? We’ve talked a lot about fraud. What are the other things that you see as having this bird’s-eye view of what’s going on that people might not realize?

Emily Sands

I think the AI economy is broad. There’s a set of horizontal model providers that have a very interesting view into where AI is being adopted and with what intensity throughout the economy. There are a number of vertical AI solutions—people like to call them wrappers, and I say that not condescendingly, just as in it’s not their models, it’s someone else’s models, but they have domain-specific data and relationships and context, and they’re solving problems in healthcare or architecture or whatever—who have a pretty unique view into vertical-level adoption of AI.

But I guess I’d be curious—what do you have in mind on who has the best horizontal view?

Dan Shipper

You’re asking me?

Emily Sands

Yeah.

Dan Shipper

Well, I imagine the model companies have the best one overall because that’s where all the tokens are going.

Emily Sands

Yeah, I think they see a lot of the tokens. I think the AI gateways also have a pretty unique perspective into who’s buying what from whom.

As I step back and look at the AI economy from the Stripe vantage point—and we see who’s buying what from whom, for how much, who’s retaining and churning their subscriptions—there are a few themes that stand out. One is, and I think people feel this intuitively, but not everyone has seen it in the data: these AI companies are just growing from a revenue perspective faster than any previous cohort we’ve seen.

I was looking at the top 100 AI companies on Stripe, and the ones that reach $30 million in ARR get there in about 18 months—a year and a half. That is three times faster than the top 100 SaaS companies from 2018. And by the way, that’s the $30 million number. But even if you look at how fast they make it to $1 million ARR or $5 million ARR, they are scaling orders of magnitude faster than high-performing SaaS companies from less than a decade ago.

The second meta trend is this very fast iteration across monetization models. Traditional SaaS had a lot of seat-based usage, fixed monthly subscriptions. That made sense because those products were being used by humans primarily and their marginal costs were basically zero.

But we’ve talked about the very real inference costs in the context of fraud. Those also have very real implications for how you price. Usage-based billing has become very important very quickly. Companies are metering tokens and API calls, but they’re also metering workflows. They’re metering outcomes—whatever unit best reflects both the customer value and the cost structure. And then they’re charging with very high precision. They literally want to know every event, how it’s rated, and what’s all the metadata that sits on that rated event.

Way more hybrid monetization models too. I talked about subscriptions, but subscriptions aren’t dead. They’re just subscriptions with usage overages, or prepaid credits that burn down, or real-time top-ups—which gets to my comment earlier on the non-payment abuse issue—and very multidimensional pricing and monetization.

Lovable is a really good example. They used Stripe billing for their initial launch, which was fairly simple subscriptions—more traditional pricing—and allowed them to monetize very quickly. Then they added a bunch of products like Lovable Cloud or Lovable AI, and they moved with those into usage-based billing. Customers are actually charged based on token consumption. It’s a hybrid model above a certain threshold. That just helps companies like Lovable align revenue with usage, value, and the actual cost of running the models.

In the limit, we actually have a solution called token billing. Underlying model costs change a lot, sometimes very quickly. If you are a wrapper on top of someone else’s LLM and your pricing doesn’t keep pace, then basically your margins can disappear. Costs go up and your price stays where it is, then you’re in the red. Token billing is just: let’s in real time track and price to the costs of the underlying tokens with some markup as set by the business.

Missa, Ship, and Lovable are all examples of this kind of infrastructure.

(00:30:00)

Dan Shipper

I love all of these points. I want to go through them one by one. A big one you’re talking about is fast iteration across monetization. It feels like there’s this hyper-experimentation going on right now where people are like, “We could charge per token, we could charge per completed request.” I think Fin, the customer service platform, charges per case resolved, which has been a thing in customer service for a long time, but it feels like that could come for a lot more types of software as LLMs make it easy.

If we’re going to pick one new pricing model—if last year’s or last decade’s pricing model was just straight-up per seat—what do you think is the new standard pricing model that’s starting to emerge from the Stripe customers you see?

Emily Sands

If you are primarily a model provider—let’s say your customer’s primarily buying the model—I think you’re metering tokens.

Dan Shipper

Like an API. OpenAI API, Claude API.

Emily Sands

Exactly. For these vertical solutions, I think in steady state you are metering outcomes. But it’s going to take us some time to get there, not because of the billing infrastructure. That’s actually totally ready. You mentioned the Fin example—Intercom does the same thing actually on Stripe billing. They have an outcome-based meter for support tickets resolved.

Why do I say for vertical solutions it’s going to be on outcomes? Because I think end users are going to want to hold those vertical solutions accountable for outcomes, and they’re going to want to know that they have positive ROI on their spend.

When you and I buy a model, we feel like we ourselves are accountable for the ROI that we get on the whole plethora of applications we might have for that LLM. But if you’re a vertical provider—if you’re really focused on solving a concrete need in a given business domain on top of someone else’s LLMs—it’s on you to ensure the ROI is there. I think outcome-based pricing is the most efficient way to hit that.

Now, I don’t think all outcomes are created equal. You could imagine these complex objective functions—I’m an economist by training, so I’ll be a little nerdy—where it’s not just “did you resolve the support case,” but how complicated was it? With what quality? What was your CSAT? How expensive was the person that you were automating in that task? That’s why I say in the limit, I think it’ll take time for us to be very crisp on the outcomes we care about, how we measure those outcomes, and those outcomes will be multidimensional.

But I just have a hard time imagining that a year from now, most vertical providers are literally charging on tokens.

Dan Shipper

That’s really interesting. I am very curious to see that because what I’ve felt—and you can see this a little bit in the Lovable example you gave, but also in the Claude and ChatGPT examples and some of the pricing that we’ve ended up doing—is it’s per seat, it’s per user with overages.

Because we’ve started to exist in this world where we used to charge per seat so people know how to model it. It’s pretty easy to figure out how much I’m going to pay. But software used to be free to run, and now it’s not. We have to cover our butts basically, and protect our margin by adding the overage so that customers know what they’re going to pay unless there’s some special circumstance.

Do you see that? Where do you see that fitting in the examples you gave? And I guess you would say eventually that might go away. I’m curious why.

Emily Sands

I don’t think the charging for use or charging for overages will go away for most of the model providers. If anything, I think that will dominate and the seat-based billing will go away.

We can go back to the Fin or Intercom example. You and I would think it’s silly to charge based on number of customer service reps that are using the tool, because obviously a lot of what the tool’s doing is automating customer service reps. In today’s world, it isn’t perceived as silly to do seat-based usage of developer tools, but I think it’s a fair question since basically November or December to say, “Wait, why isn’t that silly?”

That seems a little silly because if what these agents are doing is making every developer 10x more productive, at some point don’t you need one-tenth of developers? And why would you want your revenue pegged to the count of developers as your base price?

I suspect that we will see seat-based disappear. Now, in the enterprise context, I think it’s quite different in the consumer individual context. I think with the exception of maybe some nerds on the call, most people are actually pretty uncomfortable as individual consumers with anything but a fixed-fee monthly, maybe with some overages if they want to spend like crazy.

But in businesses, I would be super surprised if six months from now we have half of the seat-based licenses that we have today.

Dan Shipper

That is fascinating. We’ll have to have you on again to talk about that one. I’m so curious to see, and I would love to see more Stripe data coming out about that.

One other thing you brought up before—you’re also seeing these companies scale faster. You said the time to get to $30 million in ARR is 18 months, which is significantly faster than any other cohort of companies you’ve seen. I’m curious—where is that coming from?

Presumably the spend or the growth from their customers is coming from somewhere. Either it’s spend that people weren’t spending before—it was on a company balance sheet just waiting to be deployed—or they’re pulling it from another provider and then going really rapidly into these new ones.

Do you have a sense for what’s happening here? Why are they growing so much faster, and where’s all the money coming from?

Emily Sands

I think a lot of the AI growth that we’ve seen is actually net-new spend being pumped into the economy. I think it has largely not been a substitute for traditional SaaS or for headcount opex, because it’s been experimental, because people are still learning, because organizations are somewhat slow to drop existing licenses often because they’re contracted into longer durations. But also because AI was starting not literally at zero, but at near zero. There weren’t other AI companies to go take market share from.

I would say now, going forward, I expect that some of it will be a substitute away from traditional SaaS. And by the way, I don’t say that in an old-company-versus-new-company sense. Some SaaS companies are doing an amazing job reinventing themselves as AI-first. You will have AI arms of traditional SaaS companies that are eating some of the revenue from the traditional version of the same company. But some will come from SaaS.

I think some will come from headcount opex. It is very hard to believe that companies will start spending single-digit, sometimes double-digit percentages of their headcount opex in LLMs and not step back and say, “Well, my headcount cost just changed. It used to cost me $300,000 for an engineer and now it costs me $330,000 for an engineer, because $300,000 is salary and equity and $30,000 is LLMs.” So I better reason about my budget on the plus-10% basis and make headcount decisions accordingly. And ROI decisions as well.

Then some of what we are seeing is definitely substitution now across AI providers. I was looking at retention rates for AI companies, and what you see is actually within the domain—for example, within AI dev tools or AI coding tools or AI model providers—the retention rate, both B2C and B2B, is higher than it was for SaaS.

Dan Shipper

Interesting. I’m shocked.

Emily Sands

But for the individual provider, it’s slightly lower.

Dan Shipper

Within—okay, got it. Yeah.

Emily Sands

Which is intuitive. Or, well, it’s ex-post intuitive, although I actually literally didn’t know and needed to query the data. But ex-post, it’s intuitive. Once you start using an AI dev tool, a coding assistant, you love it—you’re not going to stop using it. But you very well may iterate across providers as models vary in their quality.

Dan Shipper

Anytime a new model comes out, you’re just like, “I gotta try this.” And there’s a high percentage of curious travelers basically just hopping from one thing to the next within a category. But they’re definitely going to stick in using a tool like that for a long time.

Emily Sands

Yes, exactly. A lot of the crazy-fast AI growth we’ve seen is net-new dollars spent. But I think businesses are going to start to reason about that as a substitute for SaaS, or a substitute for headcount opex, or a substitute for other AI companies. It will be less purely additive in the go-forward year than it was in the past year, when people were really just starting to ramp up on their AI spend.

Dan Shipper

Does that imply anything to you about the valuations of current hot AI companies? Let’s except the OpenAIs and Anthropics of the world, but the ones in the $30 million cohort and the coming-up ones—does that say anything to you about their prospects or their growth rates or their valuations?

Emily Sands

If you look at the top 100 on Stripe, there are little pockets of twos and threes that are directly competitive, but a bunch of them are solving totally disjoint vertical problems with no competitor yet in the space. I do think there’s enough blue ocean vertical solutions that overall AI valuations are probably okay.

I think there are a couple of crowded spaces that you and I could intuitively reason about where you might think it would be a little frothy. And by the way, you see this in the micro view too. If you look at the sales-led growth contracts—when you are the first AI dev tool, you basically charge people sticker and you do very little negotiations, and enterprises pay you sticker and whatever.

Then all of a sudden you have to have these much more complex sales motions. You hire a bunch of sellers, you have your CPQ—configure, price, quote—system, and you have this nuanced billing because you’re competing against two or three other providers who have competitive-looking monetization models and you’re reacting to that.

On the micro, you start to see some of those competitive reactions creeping in as well. But I think the overarching next year will continue to have a bunch of blue-ocean vertical stuff that didn’t exist before. There will be some pockets where it’s a little more heated.

(00:40:00)

Dan Shipper

Fascinating. I feel like I’m learning so much. This is amazing. I want to go into Stripe. Instead of talking about the AI economy, I want to go into Stripe a little bit. Specifically—Stripe serves developers and is built for a world where humans are the ones buying and selling and also making the software.

Now agents are buyers, they’re sellers, they’re builders. You have to serve agents. I’m curious how that has changed how you think about the products that you offer, and maybe moving from just thinking about developer experience to agent experience.

Emily Sands

Do you want to start with agent experience or agentic commerce? I think they’re both really interesting, but they’re kind of different.

Dan Shipper

Which one are you most excited to talk about?

Emily Sands

Maybe agent experience, and then we can work backwards to agentic commerce.

Dan Shipper

Yeah. Let’s talk about agent experience.

Emily Sands

The whole idea of developer experience is changing. Historically, when I said developer experience, you thought: making it easier for a human engineer who’s at a keyboard. You need clear APIs and you need better docs and you need less setup work.

All of that still matters—it’s not going anywhere. But I think the developer is now a broader swath of persona. It could be a non-technical founder who’s in Cursor or Replit, describing an app in plain language. Or it could be a coding assistant who’s scaffolding an integration. Or it could be an agent who’s out trying to provision infrastructure on a human’s behalf.

I think it’s less about just “how do we help a human developer write code” and more about “how do we have a coherent and trustworthy product experience end to end” that acknowledges that at some moments the actor’s a human, at some moments the actor’s an agent, and at some moments the actor’s a human working through an agent.

You see this shift in some really concrete ways. Very simple example: LLM traffic to Stripe docs is up 10x year over year. That’s just a useful signal that machines are becoming users of developer infrastructure too, including Stripe’s developer infrastructure.

Dan Shipper

What about human views of Stripe docs?

Emily Sands

Human use of Stripe docs is actually flat to climbing. It’s not a straight substitute. I think there is just more developer activity happening, and LLMs are growing dramatically within that share.

Dan Shipper

That makes sense. Cool.

Emily Sands

I would also say the humans continue to check on the docs to sanity-check what the agent is coming up with, because your payments integration is actually a pretty big decision that you’re making.

Dan Shipper

I’ll say, better humans than I are sanity-checking. But I’m glad that someone is sanity-checking.

Emily Sands

Are you YOLOing it?

Dan Shipper

I’m YOLO vibe-coding my payment infrastructure.

Emily Sands

Okay. Amazing. So maybe you’re YOLO vibe-coding, but even if you’re vibe-coding, there’s still an important step around provisioning your modern software stack, and that is still very manual. You as a human are still creating accounts across multiple services. You’re managing credentials, you’re clicking through to do a lot of setup. You’re probably bouncing between dashboards. The coding is getting easier a lot faster than the setup is getting easier.

That’s actually the idea of Stripe Projects, which we launched—I don’t know, maybe two weeks ago.

Dan Shipper

That looks amazing. Tell people what that is.

Emily Sands

Yeah. Okay, if you want in, let me know. We can use it.

Dan Shipper

Yeah, I want in. I absolutely want it.

Emily Sands

Okay. You’re in tech. I won’t Slack right now, but I’ll Slack right after this and get you in. But basically the idea of Stripe Projects for those who haven’t explored is that you or your agents can go create and manage parts of your software stack right from the command line. Resources are provisioned in accounts you own and credentials sync back to your environment and so on.

One of the things that stood out besides your enthusiasm for it—which I appreciate—is just how overwhelming the interest has been in general from the ecosystem. We launched with Cursor and Supabase, PostHog is there, Neon, Runloop. There are a bunch of great companies involved. But then immediately after launch, over 100 other great companies reached out wanting to join, which I just think reinforces that the friction is real.

You talked earlier about how some things get easier with AI, but there’s a counter effect. I think coding gets easier, but code reviews become more burdensome because who’s reviewing all the AI code? This is another example: building gets easier, but you still kind of have to provision everything.

That’s just an example of how we’re building for this world where the developer is no longer just a human.

Dan Shipper

Got it. And then tell me about agentic commerce.

Emily Sands

Agentic commerce is a bit of an overloaded term. I think a mistake that people make with agentic commerce is they jump straight to the most extreme version. They hear the phrase and think: some system that knows everything about me and decides what I need and goes off and buys it for me. And then they’re underwhelmed with the world we’re actually in. Maybe we get to that extreme eventually in some form, but we’re not there yet.

I prefer to think about it as a spectrum. The economic infrastructure you need is actually pretty similar no matter where you are on the spectrum. But the spectrum also brings some realism to it.

At the first level, AI is just removing friction from the internet we already have. It helps you research and compare options and fill out some forms and narrow down your choices. But you, the human, are still making the decision. The agent is just making that experience easier.

Then you move to where search is descriptive. No more blunt keywords and filters. It’s like: I have little kids, I need a summer camp for my kids in this budget, on these dates, with this driving radius. That’s already a better commerce experience than search plus filter.

Then you get to real delegation—and I think this is what most people would consider the minimum viable bar for agentic commerce. I give some constraints—some budget, some dates, some category, maybe a few preferences—and then the system goes and makes the purchases on my behalf.

But then there’s the further-out version, the ambient version. I don’t prompt anything and the system knows me and my seasonal needs and knows that summer camp planning is happening. That would be music to my ears. That’s the most futuristic thing.

The point is that no matter where you are on that spectrum, the economic infrastructure the internet needs starts to change. Even the earlier stages force a redesign of payments infrastructure because the old model—humans sitting in front of a browser, creating an account, choosing a plan, filling out forms, clicking purchase, entering card details—not all those steps are happening anymore.

I think there are two worlds I reason about preparing for. One is agent-assisted buying—I’m ultimately in charge, but the discovery and checkout and payment happen inside AI interfaces instead of on a merchant website. I’m not going to Nordstrom; I’m buying within Gemini or ChatGPT or Meta.

What’s challenging here is two things. One, the AI agent needs to be able to understand the merchant’s products and prices and checkout flow so that they can act on behalf of the consumer. Two, trust can break down. As a consumer, I don’t want to hand off my credentials to an agent. As a merchant, I don’t want to let every bot through—I want to know if it’s a good bot acting on behalf of a legitimate customer.

The agentic commerce protocol, which we co-created with OpenAI, is the shared technical language between AI systems and businesses. It shows up across a lot of surfaces. We built it with OpenAI, but Microsoft Copilot uses it, Meta’s in-ad shopping experience uses it.

How it works is: the merchant only has to integrate once with Stripe for their product catalogs, their prices, their checkout flows. Then they can literally from the dashboard turn themselves on through a whole host of agents and be exposed through those shopping experiences.

Importantly, the merchant remains the merchant of record, and that part really matters. Businesses want access to these new storefronts, these new channels, but they don’t want to give up the customer relationship. They don’t want to give up control over trust or fraud.

Category one is: the human is still leading the buying, but the agent is facilitating the transaction. You could call it agent-to-commerce, you could call it facilitated commerce.

Dan Shipper

How does that actually work? Is the experience something like I’m in ChatGPT and it says, “Here’s a thing you might want to buy,” and I can click checkout from OpenAI, and that’s using that protocol to then go send my information to the merchant and then send me back, “Hey, your thing’s on the way”?

That’s kind of what you’re talking about?

Emily Sands

Exactly. Yeah. Same thing—you’re in Facebook, you get an ad in Meta, you do a one-click checkout. One of the primitives we built for this is the shared payment token, or SPT. It just lets your payment credentials be passed securely from the AI agent to the merchant so the merchant can process the transaction. The merchant processing the transaction is important because that allows the merchant to remain the merchant of record.

But you don’t want your credentials viewed by the agent, which is why it’s a token and not your actual payment credentials. And the merchant needs to know that you and the agent are good, which is why as part of the shared payment token, we pass over a whole host of fraud scores.

Dan Shipper

Can I integrate this? We have a bunch of software. Can I offer agentic checkout easily, or does it have to go through the OpenAIs and the Facebooks of the world?

Emily Sands

Yes, you can. And I think one of the premises here is—just like to date we haven’t seen one model provider to rule them all or one model to rule them all—we don’t think there’s going to be one agentic shopping experience to rule them all.

Merchants will literally break if they have to integrate with every single potential new storefront. When they integrated with the internet, they built their own storefront and iterated on it, but basically they built it once. If you tell them, “Hey, you need to build your storefront for agent shopping startup X and Perplexity and OpenAI and Meta,” their eyes are going to get bigger than their heads and they’re not going to be able to handle it.

We really want to abstract away that complexity for businesses. We spent the last decade-plus helping businesses sell wherever their customers are. First that was on their websites, then it was in apps, then it was through platforms and marketplaces, and actually some in person too with our Terminal product.

But now, where are the consumers? Where are they wanting to buy? Increasingly through AI tools and agentic flows. We just want to make it really easy for merchants to agnostically participate in those different storefronts. They can choose where they want to sell, they can turn it on—a little toggle in the dashboard. But it’s not a different integration, which is the whole idea of the protocol.

(00:50:00)

Dan Shipper

How often is this happening? What’s the volume of agentic commerce right now?

Emily Sands

The volume of consumer commerce is still relatively small as a percentage of all of the commerce we see. But it is growing quickly, particularly for what I would think of as commodities.

What is the first thing people are comfortable buying through agents? It’s things that are reasonably known, reasonably observable, not super high-priced. When people started buying online, you didn’t imagine they were going to go online and buy a $2,000 couch. Or a mattress—oh my God, these mattress companies that have blown up. It took time for them to build comfort making higher-price purchases, making more quality-dependent purchases.

Today it’s predominantly commodities.

Dan Shipper

Give me an example of one of these commodities and also what the order of magnitude we’re talking about when we say it’s relatively small.

Emily Sands

An example of a commodity would be a Halloween costume.

Dan Shipper

Got it. Agents are buying Halloween costumes for themselves.

Emily Sands

Agents are buying Halloween costumes. How many lazy parents are there in the world?

I think the consumer side is interesting too because we talked about what businesses need—they need a fast, easy way to safely expose their products, their prices, their inventory, their checkouts, understand fraud, and be in control of the relationship. From the consumer angle, the question’s a little different. Even if I’m a lazy parent, I’m not so lazy that I’m willing to give someone my payment credentials and let it rip. The question for me is: how do I safely let an agent buy on my behalf?

Have you heard of Link?

Dan Shipper

Yeah, I’ve used Link.

Emily Sands

Amazing. Link is our consumer wallet. What did you use it for? Do you remember the first thing you used it for?

Dan Shipper

I mean, I use it all the time. It’s everywhere.

Emily Sands

Amazing. Yeah, it’s everywhere. You wouldn’t believe where. I was getting soccer lessons for one of my kids from a local guy, and I was on their website and they only accepted Visa and Mastercard—neither of which I had on me—or direct debit from my bank account, which I wasn’t going to put in this very janky website, or Link. And I was like, “Oh, amazing, Link is here.” Great problem solved.

Anyway, a lot of people know about Link as our consumer wallet for buying soccer classes. It speeds up checkout. But it’s already used by about a quarter of a billion consumers. It’s not a small network. What I think is most interesting about Link is it’s a very dense network when it comes to AI.

Lovable is an interesting example. 58% of their payment volume runs through Link. You are hyper AI-pilled. It is not surprising that everywhere you are, Link is.

What’s changing now is that we’re evolving Link for the AI economy because so many of the Link consumers are already AI consumers. Acknowledging that agents themselves are becoming economic actors, the model isn’t “give a random agent your card and hope for the best.” Instead, it’s delegated authority with guardrails. You as the consumer decide which agents are allowed to request credentials and under what conditions and with what limits, and whether those purchases require approvals before they go through.

You do all of that through Link. It’s just a much more sensible model for delegated purchases.

Dan Shipper

That makes sense. Emily, this was a fantastic conversation. I learned so much.

Emily Sands

Awesome. Thank you for having me.

Dan Shipper is the cofounder and CEO of Every, where he writes the Chain of Thought column and hosts the podcast AI & I. You can follow him on X at @danshipper and on LinkedIn.

To read more essays like this, subscribe to Every, and follow us on X at @every and on LinkedIn.

For sponsorship opportunities, reach out to [email protected].