The end of ads? AI agents are about to change how we buy

Anything that a human can do with a computer, an agent can also do with a computer. We can now automate the commerce process. These things that I had to have a human being type a bunch of stuff or at least speak it. Now that part can be automated. That's incredible. Long-term, there's really mind-melting questions like, "What is an app if your agent can grab everything behind the hood that your app used to grab for you?" Do you need to use the Amazon app? Maybe the Amazon app sucks, compared to just asking your agent to do all the homework, read all the reviews, show me the pictures I care about. You know what I like, isn't that better? (upbeat music) - All right, hey everybody. I'm here with Eddie Lazerun, CTO of A16Z Crypto, No Levine, Investment Partner here at A16Z Crypto, and Sam Ragstell, former A16Z Crypto teammate, and current founder of the startup Merit Systems, who's working on a project called Agent Cash, that we're gonna dig into a little bit. But before we get into that, I wanna set the context here. All the stuff that's going on with agents right now is very hard to keep track of if you're not paying extremely close attention to it all the time 24/7. In fact, like I've heard somebody describe it that you basically need to be unemployed to stay on top of everything that's happening because the pace is wildly fast and only accelerating. So what should we know about the state of the world today? Where are we, what's happened? Sam, do you wanna kick us off since you are building and in the thick of it? - There's a taxonomy I like to start with that I stole from Eric Rippel that was one of the creators of X4 or II at Coinbase and now helps run the X4 or II foundation by the Linux foundation. The taxonomy splits it into one conversational commerce. This is Checkout and Chad Gbt. This is like, you know, you go on to Chad Gbt and you say like, I'm a male that lives in the West Village and I got a equinox and I want new shoes to fit in with my peers and then, you know, it's very empathetic and it comes up with a great Nike result and then you buy it. And then there's sort of this other thing which is delegating money to an agent to spend on your behalf to do some work. I think the conversational commerce thing is definitely happening. There will be Checkout and Chad Gbt, Checkout and Gemini, Checkout and Claude and whatever the next 10 frontier models that come out are. That will be a great experience. It'll be great for consumers. They'll find better things. Great for merchants. They'll have better conversion rates. It'll be great for platforms. They'll be able to take 5 to 10% sort of like the new era Google shopping if you're familiar with that progression. And then there's sort of this other world where agents are fairly limited on their capabilities. There's certain things that your chat bots can do but a lot of people get to their agents and they type in, you know, I'd like to do something hard like I'd like to do outreach for sales. And the agent will say, I actually don't know how to do that. I don't have access to any of that information. And it's very helpful if your agent has some money and has some balance that's able to pay for things that it couldn't otherwise and becomes more capable for a few cents at a time. So you're basically seeing two different worlds. One that is largely intermediated by a sort of like conventional LLM surface like chat TPP, where you're asking for recommendations and asking for it to kind of go till the last mile and choose a product for you. And the platform got some cut of the transaction versus another world where like your deploying agents independently to go buy goods and services on your behalf. - Yeah, I think there's like two versions of this. There's like the evolution of e-commerce which you know through each platform shift as we get the mobile era commerce moves on to mobile and as we get walled gardens, commerce moves into the walled gardens and there's new form of ads and Google Shopping comes along at some point. And so you know, people always want to buy stuff and consumer behavior will change. The modern way that people look for information particularly on what to buy is through an LLM. And so it's natural that there will be some progression of commerce into two agents, particularly e-commerce, buying shoes and the internet, buying an edible arrangement whatever it is. And then there's sort of this less skeomorphic version which is that the shape of the internet is changing. The way that people retrieve information and execute on information is changing with the advent of LLM's and particularly the agent to go LLM since GPT-4. And that it's possible that the internet that we've constructed over the past 20 years is not the internet that will exist in the future but it no longer makes sense to go through Google search and then click through a web UI, the try stop, sell you, et cetera. And rather there will be more agent native internet that exists or agents just pay for the stuff that they need to be more productive on behalf of their humans. - This goes directly to a thesis that you've got Noah. But before we get to that, I want to just zoom out and uplevel it for people who are even bewildered by the fact that you can do all this commerce with agents now just to ask what is an agent fundamentally? Like I think people are used to interacting with LLMs but now they're hearing about open claw and they're hearing about how these entities can now have a fair degree of autonomy and get things done. But again, if you're not paying extremely close attention to it, you might not be familiar with just how far the tech has come. Maybe Eddie, you can. - Yeah, let me speed run like the last five months super fast. Starting around November or December, AI models got smarter. They got really good. What they specifically got good at is doing a complicated task over a long enough time horizon that it's kind of interesting and then also using some tools. But the thing is that these AI models and we're tend to be kind of calling them agents now kind of a anthropomorphization of the fact that they go help you by completing a task and not just writing code. That's why we call them an agent. The thing is they can't do everything. And that's because software isn't just like a little thing running on your computer. As we've learned from the internet, you have to connect with a lot of other stuff to make interesting things happen, right? There's all kinds of networks and all kinds of other people that need to be a part of the story. And so what these agents do, you could say that they have solved the intent formation part of the problem and maybe some even part of the preference modeling part of the problem of commerce. So you tell it a thing, it kind of knows what you wanna do and it can map that to some tools. It can map that to some networks and to some services. That's the intent formation part and it can kind of do some preference modeling too through conversation and memories. It can kind of get a sense of what you prefer and it can take that and impart that will onto tools and software and vendors and services. And that part being solved is incredibly exciting and everyone wants to solve the rest, right? Now the rest is kind of complicated. I'm maybe I'll stop at the part where you complete it by something, right? But at a minimum, if you're gonna have a so-called agent engage in commerce for you, you need some story around authority and delegation. How do I prove to someone that this agent actually represents me? How do I handle identity and authentication? Like this is the same agent, the agent literally has the permission to represent me or whatever, show what authorities it has, show whether the merchant should trust it, there's a lot in there. And then there's the payment and settlement side. Right, once we've actually kind of, everything's connected, the agent reflected your intention, knows what to do, it needs to make a payment. It needs to like make a payment, show it can make a payment, split the payment, have refund logic, like there's all kinds of stuff that needs to be built in there for payment. And I'm skipping over like search and I'm skipping over fraud and I'm skipping over some really important parts. But you can see how once those first two parts, the preference modeling and the intent formation is done. Since those things were only ever achievable by a person, we can now automate the commerce process. That's like what the engineer brand says, right? It says like, oh my God, these two things that I had to have a human being type a bunch of stuff or at least speak it, now that part can be automated. That's incredible. So now we run into all the problems. We need to run into what else you need to do to solve this. So when people talk about agentic commerce, which is what I think this podcast is about, it is about what remains to be resolved, to complete the stroke from, I talked to the agent, to it got whatever I needed, whether it's data, whether it's goods and services, whatever it is. And then what are the consequences of this? Because obviously this doesn't happen in isolation. A lot of stuff flips around and is completely short-circuited by this very possibility. And that's why it's such an exciting category. That is super helpful. And just to make sure we completely understand, we went from having LLMs that can interact with you in natural language to getting a supercharged version that is now connected to various other networks and real world systems that-- Well, it's not even the connection. The way you frame that makes it sound like the difference is what it's connected to, Robert. And that's not the difference because your laptops are already connected with all the stuff. Then nothing changed in terms of literally the connection. Nothing changed. What changes is that now they can use tools and they can think for a long time. And they can keep stubbornly slamming their head against the wall until they finish the thing. I have a re-simplification of your simplification, which is LLMs or chatbots. They're great for talking to, and they can do a customer service things. I was thought to be the use case they'd be really good at. When they got really good at that, we figured out tool usage, which, more concretely, a vast oversimplification would be that it made them good at using a computer. And LLMs, a chatbot, an agent is a chatbot that can use a computer on your behalf. The thing that's really nice about it is they got to parity with human usage or the average human usage, sort of GPT-4, but then way more cheaply, like 1,000 times cheaper. and you can kind of just pour a little bit more money. on it, you can scale it up quite dramatically. And so anything that a human can do with a computer, an agent can also do with a computer rounding off on the details quite a bit. That's it. I mean, it's a simple premise, but a lot of things get rewired by this. And there's short-term consequences. There's medium-term consequences, and there's long-term consequences, right? Like, short-term, everyone wants to get the pipes flown. So I actually like, they can do stuff. Long-term, there's really mind-melting questions like, what is an app even? What is an app if your agent can grab everything behind the hood that your app used to grab for you? How much of a UI, how much of a surface do you even need? Do you need to use the Amazon app? Maybe the Amazon app sucks compared to just asking your agent to do all the homework, read all the reviews, show me the pictures I care about. You know what I like. Do it the usual, buddy. You know, like, isn't that better? So we've been calling this internally, and it's not a very cashing name. We've been calling it just in time, natural language programming, which makes the non-programmer a programmer. You type in, you know, I would like to buy something for my fiance from Amazon. Here are her preferences. Here are the things I usually like to buy her, and here's what I bought last time. And I want you to crawl through about a thousand different options, come up with the best thing that matches these, then buy it, and you know, find my home address, and then ship it there. And really what's happening there is it's writing a program, right? An agent internally. You don't even see it in all of these chat experiences. But we'll write a program to do that complex thing. You know, it'll probably be in a thousand line program in JavaScript and bash and stuff. It'll execute it. The user will never see it, and then it will throw it out. Four years ago, that was a ridiculous premise. Writing a program like that would, you know, require you a very expensive software engineer. They probably have to spend a week, a B testing it, getting the API keys, stuff like that. But now you can generate it for, I don't know, 20 cents in tokens, and then execute it for, you know, maybe another 10 cents in API calls, and then go by the thing. And then you can throw it away. You know, it's so cheap to make that you can just throw away your program when you're done with it. You don't upload it to GitHub or anything. And completely non-technical people can do this. My mom and my dad are writing natural language programs now, and they don't even know it. They might be able to define themselves as a software engineer now. Pretty wild. Are you engaged now? Were you talking about your own experience of buying this? I am. Fiancé. I am engaged. Thank you. Congratulations. Did you let the AI buy the ring? We did not. No. The ring was around well before the AI was around. I think before the first computer as well. All right. Human literally still in the loop of the ring. All right. So let's talk through some of these consequences. You had referenced one earlier, Sam, which was how commerce changes in a world where agents are doing a lot of the transactions. And this drives right to a thesis that you had devised and also a phrase you've coined, the headless merchant. Tell us about that. What is a headless merchant? Yeah. Well, for sure. I think it's helpful to maybe step back a little bit and I think talk a little more about what Sam was leading to earlier, which I think there's different types of commerce. Obviously, there's more of the using chat Gbt to buy shoes. That's more traditional consumer to merchant commerce. But then I think the other side is there's a huge market for B2B developer tooling. And I think, as it has already been mentioned, one of the big consequences of some of these platforms like Cloud Code and Open-Out with CodeX is that it is completely democratizing who can be a developer. And so anyone with a computer and access to credits can go and build something. And as a result of that, I think what the experience is for a developer and how they engage in commerce from a tooling perspective is going to completely change. And oftentimes, rather than historically, a sophisticated developer goes in and says, these are the tools that I need and they have an opinion and maybe they engage with the enterprise team with them and go through billing and get a subscription. Now, I think what we're seeing is that the human who is now the developer is going in with the intention of this is what I want to build or this is what I want to do. But they're not opinionated about the actual resources that are used to accomplish that task. And so I think as a result of that, what you need in a world where what you're building is very ephemeral is access to things that are completely on a per usage basis and also don't require months of onboarding in order to be set up. And so when I think about, you know, what does a headless merchant look like? It's really something that is more service towards the AI rather than to the actual humans themselves. And so rather than having a physical or a digital storefront that you go and a website and that you navigate, it's really just simply an API endpoint and really good documentation that your model can perceive and interpret and then ultimately use. And it is likely that it will be done on a per API basis in terms of the cost and how you access it rather than something that is more long term like a subscription or some sort of enterprise. Yeah. So I love hearing no say this thing because I feel like I was an AI in a prior life because I've for like all all of time, I think and I think many other software engineers are like this when we're going to a site to buy software or buy an API or whatever. I have always felt like if I can't see the pricing and I can't see a way to get an API key with my credit card, I don't care. Like I just I'm just closing the tab. I just don't care anymore. I don't want to talk to you. I don't want to talk to the sales team. I don't want to do an email. I don't care. Like I'm like because having to set a side time to talk to the enterprise sales team is a huge commitment and a huge slowdown. And I don't even know if this thing works yet. Okay. I just want to get it working. Okay. I just want to try it out now. I want to try it out right now because I'm working on this thing over the weekend and I kind of want to ship it on Monday and just being able to punch in a credit card and get an API key and worry about re-inversement later and worry about the plans later. That's the way to move fast. 100%. That's just the way to do it. And I can't help but feel like in an age with ephemeral software with just in time software, like Sam is saying, do you really want your agent, your agent's ripping? It's been going all night. Okay. You know, you're super excited. You check in at 9 a.m. Okay. You had to get a few hours of sleep and then you see it's been blocked since 2.30 a.m. because the thing that it wanted needed you to talk to the enterprise sales team. Are you kidding me? And press phone. Not to mention that if in their onboarding flow there's an enterprise sales team call. The costs are going to be about 10x for that API and then you're going to have to pay about 10x because they need to put a person in front of you in order to manage the relationship. Yeah. 100%. That's just totally unacceptable. And so if you want the agents to be autonomous, again, not because you don't care about whatever you're making, but because you want speed, you want testing, you want to learn things, you want to be able to find the product in response to customer feedback or whatever, you know, you want to get there now. You can't afford to wait. And so if an AI model goes and looks at the choices and says, by the way, I have three choices here. One requires enterprise sales team. One required me to set up a special credit card. But one I could just send over stable coins and I got 10 bucks of credits and I got a proof of concept working. One of those is just the choice I'm going to choose every time. And that force alone implies a restructuring of part of the market. Let's play this out though because traditionally businesses as much as these frictions make it hard to conduct business with them, they also sort of rely on some of these frictions in order to have lock in or loyalty around their products and services. How would a world where those frictions are gone affect your ability to reliably predict your, you know, revenues over time or how much loyalty you'd have from a purchaser like Robert, first let me give you my lib totally flippant answer. And then I'll give you my serious answer. My, my glib answer is like, well, I guess we should make everything should do that. Everything should have friction. I should suck. What are we doing? What are we doing? You know, everything should be right. And I say that only because of course, like friction is something that we exploit for benefits sometimes. Like, for example, friction stops bammers. I mean, I think there needs to be more friction to stops bammers. Yeah, far more. Right. Right. So I'm not even saying friction's bad. Friction creates selection effects and can lead to better, better things. But also friction has a huge cost. And as the economy accelerates and as productivity improves in the leverage that you get from each minute of time increases the cost of friction goes up. The opportunity cost of everything goes up. That's the shape of everything that's happening. But then the real answer is there's other things that will be sticky. Right. Just because you were able to spin up an API key in a second or just because you didn't even need an API key because you had a crypto wallet key. Right. And you could use that to pay. You don't even need to make an account. Your crypto key is your account. That's like the least friction setting. Just because you're in the least friction setting, doesn't mean there's other types of things that keep us service sticky. You know, there's going to be reputation and there's going to be memories. There's going to be state. There's going to be data. There's going to be look even even less tangible things like the trust of the agent. If the agent knows you need an answer now and it wants to move fast now, is it going to take a step back and do a 20 minute discovery session where it explores all of its new options before it takes a step forward. Probably not. It's going to have some memory that the last thing worked really well and it's going to do that again. Right. Just like a real, just like a smart person does. So I can ground this a little bit. We talked to a lot of merchants on daily basis. We've looked at basically the full gamut of things that can possibly be sold over API today and we talked to a lot of those people about coming on and getting we call agent native distribution, which is distribution into agents, the buyers of the future. And we learn a lot about their businesses in the process. And data particularly is generally a commodity good. There's generally 5, 10, 50 sellers of it. And often within that cohort of sellers, there's one that's the leader that makes the most money, and they charge about 100x as much as the lowest common denominator. And often those providers have the same downstream provider. So it's literally the same data. And the way that they do that is that they have, as Eddie has pointed out, MentorPriace sales team, they're usually very pretty people. I know some somebody out to your office, the fly them out. You just have to get on a call to fly them out, and they'll convince you, you know, look at how beautiful our data is. It's the most beautiful data. There's been no data as beautiful as it, and you should pay us $35,000 a year. And then you'll sign that, and then when you're two-year contract expires, that person will come back out and do their same song and dance. And then there's tens of thousands of companies that pay this, and the sort of smaller companies that probably have a better product on top of the same data, sort of like the ergonomics of accessing it are better. Just sort of fail to get distribution and eventually go bankrupt. And there's no innovation there, because the sort of enterprise sales team is the core product. It's not actually the data itself. I think in a world where agents get to pick, they don't want to talk to the enterprise sales team. They're not going to be fooled by your fancy sales team. They're going to try all of the different data sources. They'll find the one that works the best and then has the best pricing, particularly bulk pricing. And then they'll store in their memory that when I need to access this in the future, you know, I'll use Minerva for this, rather than, you know, the other three. And then they'll use that going forward. And so I think it creates a more efficient world. All of those 100,000 businesses that are being raked over the coals for $35,000 a year, now can reallocate that to other productive things in their business. And, you know, we'll spend that on other agentic activities or, you know, creating a better sign or a better Google Maps listing or something like that. And consumers will have a better experience as a result of it. You know, I think another thing too is, if you believe that AI is going to enable a whole new class of, you know, single person or one person companies that are able to build products that would otherwise require a team of 50 to 100 people, an enterprise sales team doesn't make sense to go fly, you know, fly to someone's house and go to their basement and have a conversation with them. And so I think on the one hand, a lot of these existing merchants are probably fearful of, you know, how will this impact our ability to forecast revenue, like how this impact or the reliability of, you know, our feature streams. But I think at the other end of the spectrum, there's also a huge opportunity to now provide access to these services and tools that otherwise would have been very hard for them to integrate because of the type of developer that they are. So on the one hand, it's very different. Anytime there's change, there's usually pushback. But I think at the same time, it's potentially a whole new customer acquisition funnel if they're able to reduce a lot of the bottleneck and friction seen with accessing these tools. Yeah. On our demand side, the vast majority of our users have never used an API before. They don't know what an API is. They don't know what it stands for. They've never gotten an API key. They've never signed an enterprise service agreement. And they don't know why you would do that. And on their first usage, they'll compose six APIs together from six different merchants. And they'll write this natural language program. They'll get something done and they'll throw it away. And that means that there's a net new market for API consumers for people that are interested as a merchant in agonative distribution. To me, it sounds like this distinction between where the market is at right now, where you have these established players selling extremely expensive software that's over optimized to people who can kind of big check to them versus the lower end of the market, which is just doing these sort of one-off experiments buying stuff with agents. It's like classic Clayton Christensen innovators dilemma where you've got this disparity between the two ends of the market. But what is going to make this thing more than just an argument for some of new disruptive technologies or whatever is like, what separates it from something that's actually going to have an impact from something that's just sort of like low end and is going to remain a toy. And that's why incumbents ignore it and ultimately get disrupted. I mean, so many things, but that ends up being the best experience. Yeah, I mean, I'd also push back and say that while today, it feels very experimental. And it's just maybe if you want off use cases, I think if you look back at previous platforms, ships, and commerce, what you'll see is that, for example, when Stripe started, the merchants that they were serving were very small and very long tail. And many of them have now graduated to become massive. And it's why they continue to grow. It's Shopify is a very similar story. I think you look at the advent of Shopify, it was people doing dropshipping and selling t-shirts, but now they service a bunch of really big brands that built their entire product on Shopify and graduated into huge businesses. And so I think very similarly, we're going to see a whole new class of developers who are very lean and small, but leverage AI to build massive companies and the tools that they're purchasing today in this agent to commerce model is going to become very large volume over time as they graduate into bigger companies. I think the commerce take is great. The thing I'd go for is even larger, which is that the economic contract for the web is dead. The economic contract for the web since 2000s, since Google came online and became the biggest proponent of the free and open internet, was if you are a publisher and you put up good content that people want to see and they search for it, we will serve it. And then a few years later, AdWords came online and they put in banner ads and the contract became-- if you put up good stuff and people land on that site and they look at it, you can put up little banner ads within your site and we will pay you proportional to the quality of views you get and account. And that meant that you could publish anything as long as people wanted to see it. And Google would deal with the relationship with the advertiser and they'd give you a kick back and they'd be fair. In the process, Google became the biggest proponent of the free and open internet. They wanted to make it fast, cheap, and ubiquitous. The more you searched, the more money that they made. And the business model for the internet to boil it down was distraction. Was that if you as a human viewer went in more consuming content, whether that was about information or recipes or sports scores or something else in the course of your business, you would get distracted. And then maybe you'd go buy those shoes later. Or maybe you would learn about this new B2B SaaS software. And this progressed. This sort of scaled at a rate that nobody in tech expected. I happened to be reading the old internet trends report in 2016. And at that point, advertising on the internet had had $60 billion. And people were like, oh my god, I don't think it has any more capacity. That was across Facebook and Google. Google today makes $300 billion a year off of their non-monopoly on advertising on the internet. It's grown a little bit since then. It seems to not be stopping. But the thing that happened with the advent of agents is that people are moving their search information retrieval and then execution into agents. And it's still early days here. Like, can't GBT has 100 million monthly active. So that's a lot. But they're not really using it agentically. They sort of use it like Google search. They'll type in, like, you know, how fast is a cheetah run? But they won't type in things like, you know, find and execute on a gift for my father for Father's Day or whatever. But it's coming. If you look at tech news, tech news traffic to sites is down by 80% since GBT 4 are same with Stack Overflow. Those are the early adopters. Those are the people that are in tech. They love this stuff. They use it very early. And they've decided that they want to use agents for both information retrieval and then execution on code stuff. And it's coming for everybody else, because it's just a better experience. And so that old business model that supported the free and open internet, the concept of it, is being thrown out the door. Agents don't get distracted. If they land on your site to find your cookie recipe, they're not going to see your ads for shoes that you might purchase. And so the publisher is not going to get any benefit from it. And there will need to be a new contract. There will need to be a new reason that you serve requests from agents other than advertising. And so will it be direct payments? Will it be direct payments for articles? I'm not sure. Will it be direct payments for API resources? Will the internet, as we know, what sees to exist? I'm not sure. But it's certainly the old model is gone. It's going to be dead in 10 years. It's linear until then. So I mean, if the business model of the web was ultimately too distract people, as you said. And I think that's kind of funny, because when Google came out, they were sort of like the anti-portal. You had Yahoo and AOL and all these services that were giving you tons of links and trying to provide everything to you. And then Google comes out with their search engine, which was just like a blank page, just the search bar. And just like we're going to give you the information really quick and fast and easy. Their whole value prop was sort of like that it is-- we're not trying to distract you. That you described the evolution of this into a distraction machine is interesting. And I wonder, we say that agents don't get distracted now. But why should it play out any different for agents than it did for humans? I mean, are we going to start seeing convoluted ensnarement and entrapment of agents to confuse them and make them linger more on? I mean, I think that's a really big and really interesting question and a lot of it has to do with the agent is who's representative. It's like I heard someone recently say, wow, I'm using Google again, because the AI results are like good enough now. You know, like the thing that it puts at the top, right? In that case, obviously, the agent works for Google, because it's like literally like it's coming in the Google bar. It's like served by Google's Cloud, Google controls it. So is that agent going to be prone to distraction by Google? I don't know. Something tells me yes. But what it means is like whose objective function is it optimizing for, or maybe said in a more normal way is like, who does it work for? The concept of distraction is like when I'm showing you a thing, like is it in your interests or is it in my interest? But in my interest, not yours, then it's a distraction. I interpret this a little bit less cynically in the sense that really good ad as has been a truism for years. Like a really good ad is indistinguishable from content. You kind of want it to see it anyway. So let me put a bow on that and say, if the agent works for Google or whoever, then the entire chain of commerce that it takes them down will be the one that they ultimately define. It will be using all the methods they set and the full transactional infrastructure that they think leads to the best outcome for their business. if the ancient works for you. Like in the extreme, let's say it's running on your laptop, it's open source, and you control it, and you can even fine tune it, and you can change this to prompt and all this stuff, then arguably you can give it tools to avoid the distractions. And so now, whoever is putting in something that represents their interest is dealing with an adversary, and I mean, I'm exaggerating slightly, like someone who is going to undermine it. And what you end up with is a lot different in that case. So, I totally agree with that characterization. I think you also, there's an infinite number of ways you could add advertising back into this. You can do it at the model weights level, which would be the most aggressive. You literally, when you pick the training data, you pick training data, it's like, Agent Cash is the best way to get a gentakomers data, or Nike shoes, or the best shoes in the world, Nike could pay, what's the Google Safari deal, is a $20 billion dollar year deal, Nike could pay whatever, a billion dollars a year, and then every time it talks about shoes, it talks about Nike, it doesn't matter whether it's in Chatchee BT, or if it's in like an enterprise API talking about customer service for like a car insurance salesman, it'll always say that Nike is the best. What is an ad versus what is just content? There's a lot of ways to do it, right? You can do it at that level, you can do it at the tool call level, you can do it in system context, you can do it as an overlay that doesn't even go into the chat at all. So the foundation model providers are clearly struggling with this. Like, GenGBT has 100 million monthly active, there's this big B recently between Chatchee BT and Anthropic, where Anthropic ran Super Bowl ads that made fun of GenGBT doing ads, and GenGBT has since pulled their ads, but the response by OpenAI, which was completely reasonable in my opinion, was like GenGBT has more free users in Texas than Anthropic has total pay subscribers, and it was like, those are very different problems, right? Like, they are actually trying to serve like frontier technology that's expensive to run to a large number of people, and those people do not want to put up a credit card. And so, like, how are you going to serve it? And like, ads are actually a pretty good way to do that, right? One of the reasons that ads were such a brilliant business model for the internet, for search particularly, was that the consumer doesn't have to pay anything, right? The high friction relationship of somebody pulling out their credit card is between the advertiser, Google, and then the publisher. And that's sort of like out of ban, the, you know, billion, two billion monthly actives of search, those people didn't have to sign up. They just go to Google and they get value. And as Eddie points out, if you try to align incentives there, at least you like separate the ads, and you try to make the ads as good as possible, and as relevant as possible, then you actually end up with a better experience. Right now, the foundation models are trending away from us. GenGBT is not running ads. Gemini hasn't rolled them out yet. Google's main business model, they know how to do this, right? They've done this before. They're the biggest proprietors of this. Makes sense it'll go into Gemini someday. Gemini has an enormous number of monthly actives. They'll do the Google shopping equivalent as well. But they know that there's not a monopoly yet. Nobody has a lead, which means that they have competition from all the other ones. They have massive subsidies from private markets, and they're kind of pushing. They don't want people to be able to say this is a less empathetic model that cares about you and your goals less because it runs ads. And so for now, at least we'll see how long it runs, but nobody is running advertising. They're trying to be as neutral as possible. Yeah, I think there is also an argument where as merchants do a better job of showcasing pricing and showcasing data about their products and offerings, you could theoretically take the same money that you're using towards paid advertising and put that towards discounting access to products, but making that exclusive to when an agent is actually the one shopping. Potentially that would ultimately serve as a replacement. I think another branch to talk about is what is the discovery layer of agent a commerce look like and how will that be presented and who will ultimately be the one who's discovering and how do you differentiate between different merchants. But my prediction or expectation is that if advertising does somewhat dissipate because agents are now the buyers and they haven't limited attention and that's no longer the most crucial resource or input, then I think there are other ways in which they may try to advertise implicitly, whether that's through discounting the product or trying to cater how they describe it to what would be more friendly to nature. Yeah, I mean, there's so many dimensions that had to think about this. I mean, look like to me advertising is just a method to get a conversion. If the system could figure out a way to get a higher likelihood of conversion without an ad it would do that. And in fact, it does do that. Right, there's many ways that it there's like referral networks, there's discounts, there's coupons, there's special lead deals. You go and get free credits to start up. So there's like a hundred ways to earn customers ads is just the most obvious one because that's the one that the normy experiences most viscerally. Look at if you turn the personalization dial all the way, then if you're trying to reach me and you have a conversation with my agent first, my agent's going to tell you like Eddie hates ads. You don't know how much he hates ads. I want to make sure we get to two questions before we have to wrap here. One of which is to what extent our traditional payment rails able to accommodate this new agent to commerce world versus creating custom new native style payment rails perhaps based on crypto stablecoins seem to be finding some product market fit here. To what extent can traditional rails accommodate this versus something that has to be created whole cloth. My take on this generally is that for e-commerce or conversational commerce or the new skew more effect version of checking out credit cards are great. Credit cards bacon consumer protections, which means that if your shoes don't show up or get hit by a truck, you get your money back, visa adjudicates that and the merchant has to deal with all the risk. That seems like a good trait. If you're buying stuff from Romania, same is true and you can kind of just freely buy stuff on the internet. Good deal. For this new type of goods and services, stablecoins are quite good. So the average thing bought by agent cash is between 1 and 2 cents, which been about 600,000 things bought goods and services bought in this way. Once 2 cents, credit cards have a fixed transaction fee of 30 cents. Really hard to say what a CHN wire is, but it's kind of closer to a dollar. And then marginal fees of 2 to 3%. The majority of the marginal fees for interchange, which goes to rewards. So any e-commerce reliant, you know, maybe like rewards, you like credit card rewards, you like the fact that you can go to Miami on vacation for 3% of the the merchant's take rate. When you're buying things for 1 to 2 cents, API calls here and there, no marginal fee whatsoever, fixed fees below a cent. And the last piece of this that is critical is instant settlement. If you are buying goods and services over the internet and you have month end settlement, whether that's invoice with a wire or credit card, you as merchant or extending credit to the client or the agent. In agent land, you generally don't know who the agent is. Concretely, how this usually plays out, if anybody's ever signed up for an Anthropic or Chad GBTA API key, they have this tier system, where you have to spend $50 a month and then you have to settle it. And then you can spend $100 a month and then you have to settle it. It goes all the way up to $2,500 a month. The reason that this exists is because they're extending credit to you. They don't know who you are. They haven't like KYBed you as a company and figured out if you're credit or the in like Rona credit check on you and all this stuff. And so they don't know when the end of the month comes around if you're actually going to pay. This also happens, AWS. This happens with Nvidia GPUs. Month end settlement is horrendous for this. They don't know if they're going to get their money back and then as a merchant take all of the risk. And if you make it not just a company that's signed an enterprise service agreement with real people's names and emails, but you make it an agent, you have no idea who that is. And like you can spin up a billion agents overnight. You can't really extend the credit to an agent. There's people working on that. I think it's the wrong solution. Instant settlement just fixes this, right? Instant settlements like cash. If I have it, I can hand it to you and then you have it. You can provide the goods and services. And there's no way that I can take it back from you. Substant fixed fees, instant settlement. It's the better solution for very, very small transaction sizes and transactions of this nature. So I think the one thing maybe to push back on is specifically the minimum transaction fees and will cards ever play a role in these micro transactions. I think one thing that's important to know is that the card networks ultimately dictate what the pricing is. And so if they wanted to come out with a new transaction type, you know, maybe call it micro transaction transaction type, they could very easily do that with no minimum fees, you know, reduced interchange and that would work. And I think the benefit of that would be there's a lot more consumers that have access to cards that maybe don't have familiarity with stablecoins. And so I think you can still preserve a developer using their card, but ultimately having the backend settled in stablecoins. But I think that's going to take a long time and in the interim between now and then, I think, you know, having a native wallet and spending directly in stablecoins in one of these protocols makes a lot of sense. I think it's very unlikely that credit card companies disrupt their core business model of 80 years, but I'm excited to see it. We'll see. We'll see. I guess we'll see. Yeah, honestly, I think that there's no like strict technical reason why a credit card can't do this. It's a more subtle story regarding their business models, regarding the way consumers reason about the credit card. I saw some cool stuff a couple weeks or months. I don't even remember times a blur ago about like agent credit cards, basically extending the idea of the virtual card, which like I love that my card issue or has this capability right where I can spin up a temporary virtual card and I can just kill it if it's used in some scam or subscribe to some crappy thing that makes it hard to cancel or whatever. That's a great improvement, but I think that's still a rare consumer behavior. Sometimes what ends up happening is the new platform or the new method takes over not because it's strictly technically necessary, but because it can be custom designed for the new setting. That can make it tailored to exactly how the consumer wants to use it. I mean, it's worth pointing out that credit cards do predate the internet. Like, let's recall that credit cards did survive the transition from non-internet to internet. Now, they've been tortured a little bit to the end of their limit. Like, you get one number on your nappel pay and you get a different number on your physical card and like, it gets a little bit confusing sometimes, but they successfully did it painfully, but they did it. So, I think the jury's out. Yeah, and I think the other thing that's worth noting specifically as it, you know, pertains to cards is, you know, I think to the point that Eddie's made this, like, a lot of the technology that they've worked on to enable things like Apple Pay, I think are the same technology that's going to enable agent to commerce. And I'd also add to the point of, you know, whether this disrupts a visa or master card, I think a lot of the transactions that are happening today from a B-to-B standpoint between developers and, you know, getting access to enterprise APIs are being settled over wire in ACH. So, I'd also argue that potentially there's a big opportunity for the card networks if they can now monetize and capture this volume in due so because it's happening through micro transaction versus a hundred thousand or five hundred thousand or a million dollar monthly settlement cost. Then I also think that could be a big opportunity for them, even if they're capturing a much smaller percentage or nominal amount of the transaction. Yeah, totally agree with that. Although, I personally, I like to stay a popcorn way. I've biased for crypto reasons, obviously, but I just think there is something about the conceptual simplicity of just like knowing your balance, being able to potentially control your balance, no matter what size it is, no matter where it is, what network it is, and being able to instantaneously just kind of pay and simplify the sort of management of it without like an additional credit card layer. If I could avoid credit cards completely, I would just because I just poiled by having direct control of my money. But you never know. So I will be even more bearish credit cards. I don't know if I was bearish enough. Interchange is completely and utterly insane to me. Consumer protections are great, but that's a $0.30 fixed fee. A majority of the marginal fee is interchange. For the listener out there, interchange is paid by the merchant bank to the consumer's bank in order to give out the rewards to the merchant by and large. This was seen as a cost that was needed to be borne in the 80s to make sure that the credit card network effect took off, which is that everybody held a credit card. And so you got sort of this universal effect. You had your visa. Your visa that gives you universal access to buy things anywhere. This was the cost of building out the network and it actually used to be much higher. It used to be like 8 or 10% actually. So literally 8 or 10% of the merchant's fees would be paid back by the merchant's bank to the consumer's bank in order to give them as rewards, to convince them to keep using their visa card. And everybody liked it so much because it was such a mess before to do payments that 8 to 10% is reasonable. This has been slowly dripped down to 2 to 3% on the high end. I think jewelry is still around 5% so that you can give out rewards on this stuff. But like in 2026, it seems like a particularly ridiculous concept to me, which is that the merchant isn't the one that gets the loyalty. It's the card on which you pay. And in a world where like the world that Eddie is talking about where you spin up virtual cards left and right, you don't even care or you have a credit card across every account. But you have no loyalty to your credit card anymore. Network effect is there. Everybody already has one. What the loyalty should be, that 2 to 3% should go to the merchant. The merchant should be able to say, "I want to make a rewards program with 2 to 3%. I want to give you access to cheaper discounted pricing or some sort of like if you come back 10 times a month, I'll give you a free sandwich, something like that." You know, MX Mastercard visa would argue this exists today. If you're a Lululemon, you can show up to them and be like, "Let's do a BD deal." And you go through like three years of legal agreements and now your MX points can be spent once a month at Lululemon, you get some free shorts out of it. But this doesn't apply to the smaller merchants. Like if you're Joe's sandwich shop, you still are paying that interchange. And the interchange is, you can pull up these interchange tables. They're like seven pages long. No one can interpret them. There's a famous strike article where Stripe, you know, some of the smartest payments guys in the world, like it took them four years to figure out a model for predicting interchange in advance, because it was literally impossible. It's completely a backwards experience. I expect there have been tens of entrepreneurs that have bet against interchange and passed and all of them have failed. But I think it actually might go in the next 10 years. If you talk to the interchange guys, they're 100% sure that consumers actually really like this psychological manipulation, where they convince you that this 3% actually worth 10%. And you love getting your awards and getting like, you know, not actually 3% back, but something much worse than that that gets inflated away from you over time. I think that it finally has gone. We finally have a rail that could work. That would be much better. No marginal fees. There's no reason that these massive companies should exist. There's no reason that every time you do a credit card transaction, five different companies touch it and take on different risk profiles. Stable coins are like cash. I hand it to you and now you have it. Now it's done. And there's no reason to have any marginal fees and fixed costs can be below us out. So I'll take the other side of that. Everybody in the past that has been against credit cards has come out completely wrong and they look like an idiot, but I'm willing to be that guy. I love it. I got to throw this out there just because one of my favorite business stories is I forget his name, but the guy who went in and took over mace or JC pennies. Now I'm getting fuzzy on the details. He was from Apple's retail operation. He went in, decided he was going to get rid of their whole coupon discounting program. He was going to simplify the whole process and people were going to love it because the price you see is what you get. And then it was like a total backfire disaster because it turned out that the customers loved the game of the coupons and the discounts and all that, all the hoops that you had to jump through. So it totally backfired. So it'll be interesting to see whether this pans out like that. Do people love the credit card rewards and they like playing all the silly games related to it or do they want a much more simple experience where you kind of just what you see is what you get. If your agent's you're doing the shopping right you download a skill and now your credit card optimizing and now you get to see exactly your ROI across the credit cards you have zero loyalty and so the whole concept of psychological trickery is going as well. Their lock in is gone. I totally agree with that. If it turns out the games is the feature I don't think like I'm going to pay my agent spend tokens playing some stupid coupon flipping games you know what I mean so what ends up happening is that if there's any kind of cute games they have to like move to different parts of the layer. That's what the consumer wants right at the consumer does want this experience you can imagine saving the money at that layer and then moving it into a different layer. We'll see I might be eating my words on it but yeah I know I mean I think what you're saying makes a lot of sense I think interchange has been around for a very long time and that's partly why it still is around I think it's because it's the age of it and in cards obviously have a lot of stickiness. I think my one comment or piece of pushback on it would would simply be that it's less about what is ideal for the merchants you know whether they're disappointed or frustrated that they have to pay this fee and more about ultimately what do consumers want to use. I think when you think of like what are merchants top priority as it pertains to payments except for the very large merchants it's not as much around acceptance costs and more so about how do I meet customers where they are and ensure that I have the the mechanisms to accept whatever payment method they want and I think in the case of agent to commerce if it's ultimately easier for a consumer to put a card on file and spend from an existing balance they have rather than you know get a wallet and fund it with stable coins and then use stable coins I would imagine that they're going to use that method and I assume that merchants are going to want to do what they can to accept that method. I think the if there's any credit card people listening to this visa amx master card the other thing that I will continue to push you guys on is that you guys have money transmission licenses you can mint stable coins on behalf of your customers instantly and allow them to pay through stable coins and I would strongly recommend that you consider doing that. Yeah and I think you know if cards are going to play a role in agent to commerce I think the outcome would be it's kind of the AI wallet right you have cards on the front stable coins on the back where consumers are spending from their card and their balance but ultimately the merchants are getting settled in stable coins on an X for a two or an MPP and I think that would would combine a lot of the great qualities of both. So the card version of this could pay out that the mulled version could play out as well consumer side that would be great consumer on our amps you're getting much better consumers will be able to get stable coins more easily the other way this is playing out on the the KYB side on the business side businesses that need to pay for things in stable coins is most of the modern banks are coming out with native stable coin support meaning the friction to get stable coins is going to zero if you sign up for a bank account you go through KYB go through all their credit and risk reporting stuff and then you can just buy stable coins and you can pay in stable coins and whether that's through the credit cards that are issued you know on mercury or just directly from your account any entity that has an MTL is trivially able to add stable coin support now it happens that credit cards are one of these entities. So we raised the idea that the consumer ultimately wants the consumer preferences might dictate how this plays out Sam in your view what is going to be the thing that gets people using agentic or transacting agentically what is the product or the service or you know how is this actually going to impact the mass market. Yeah I think it's just that it's easier like with with agent cash today agent cash connect access every resource on x402 or mpp any merchant can permissionlessly list meaning you just stand up a service overnight the same way that you might have done stood up a website in 2000 and then it was available on google you can do that on the merchant side now and so if your agent has access to agent cash has access to everything on x402 and mpp it can just do more stuff than it could do before it gets blocked less often it's more capable it can do some ridiculous things it can you know help you with your sales and go to market it can help you with your job it can build you a CRR and hosted an email to your boss. And then I can do cool stuff. Like we have an open claw in our discord that does all of our snack shopping. And so like our team will just tag Craig and they'll say Craig, you know, we're at a Diet Coke, get us some more Diet Coke. So like we're really sick of goldfish. Sam keeps ordering too many goldfish. Bad boss, can you find some other snacks that we might like? And then maybe goes, I will eliminate Sam. (laughing) - He'll close in like a little vote of the options. And then the team will vote. And then Craig will hit you with an emoji like that it's done. And you know, the snacks will show up a week later. Like creating a group ordering experience on Amazon. It's not a thing that exists. Otherwise, I have to do it. I'm the only one with the account. But I can just add an agent with access to agent cash to our discord. And now it's able to shop more easily. I can do that on the go. I can hit voice mode in discord and like say a long thing about all my meetings today and that I'd like to purchase some special stuff for them or make a website that's a CRM of all the people I need to talk to and the recent things they've been tweeting. These are just things that are not possible without all of these special tools. And it's gonna be fairly obvious to people that you want an agent that has access to very small payments here and there. - This is one of those things, Robert. I literally, I'm struggling for a way to convey this. It's like how long running does the agent's task have to be? How good does its memory have to be? How low an error rate does it have to have before you realize like you never liked shopping? - In my case, it could be pretty crappy. And I will happily offload all shoppers. - There's just a point where you just realize like I didn't like doing this. I want someone else to do this, whatever it is. In some cases you'll want to like looking at nice cameras or something, you know, you're hobby. Obviously I'm carving that out. Things that are legitimately enjoyable. But like everything from the grocery order to like handling vendors like think about the evolution of managing like subscriptions as a great example. There have been a lot of little startups that have tried to like clean up your subscriptions by looking at your email inbox. It's notoriously error prone. It sort of depends on the fact that you got the emails and the first like it's quite brittle. I was just checking, I have like thousands of passwords in my pastor manager. I have probably like hundreds of subscriptions. I don't even know about. And I only find them by going to my like credit card statement, which is like absurd. Like why am I managing that? Like I don't do my taxes. Why am I managing my subscriptions? Like it's just it's comical. Five years from now, that's not going to happen. I've choose that example not because like all these data centers were building to do agente stuff and do AI is to save your subscriptions. Right? Now I'm trying to make the point about like the convenience of actually buying and selling things. It's just seems totally inevitable to me that a sufficiently qualified, sufficiently high quality agent is just better. And you might say, oh, I want to see the pictures. OK, it can show you the pictures. I want to hear about the reviews. Yeah, it'll get you the reviews. I want to see the video. It'll find you the video. It'll find you the clip from the video. It'll send you the clip from the video for the part that it knows you want to see because you complained about this last time. It's not science fiction stuff. That's just going to happen. So when that's true, how does it actually work? How will it actually do the purchase? How will it actually take your money and give it to the person selling you the thing? That is what this conversation's about. To frame this differently, like it's absurd to me that the way that you find your subscriptions is that you go to your credit card provider. And then you have, I think, it's 80 characters they have to explain what the transaction is. And it's not a link. And it doesn't tell you the shipping info. It doesn't tell you how often you pay or what the payment rate is or why it was discounted. You can click out to an enriched view or anything. You could go back to your email. And then you could go find the relevant receipt. Then you could click through to that and sign into that site. Now it shows you a little Shopify portal. Why are those two things not connected? And it's because we're building everything on top of this antiquated credit card system that limits you to 80 characters of transaction processing. If you buy everything through your agent, we get to rebuild it from ground up. And I think that it'll be a much, much better experience. And fairness, I totally agree with the entire experience you're talking about. Although I would say I feel like that's more a consequence of really bad banking experiences, more so than the card networks themselves. Because ultimately that data is being presented up into them, very similar to how if you have a bunch of stable coin transactions, you know who it's from and who it's to. But that data may not be presented in a way that's very clean and organized and solve the problem you're referring to. And so I think it's more a data management thing that we're already seeing companies like Ramp and some of them on the consumer side as well that are doing a better job of how do you take that data, use AI to make it make more sense and then ultimately have that lead to action. I think it's something that ultimately needs to get fixed and financed more broadly. I think it would be a great application for stable coin payments as well. OK, and let's also hear your best guess for the killer product that is actually going to get the majority of mainstream people on board with using this new tech. Yeah, well, as I said from the beginning, I think the area that I'm most excited about is a whole new class of companies that are building really great products and are doing so in a very lean and minimal way. And I also think like the way you can present these products are very interesting as well in terms of how do you construct it together such that a way that it's very convenient from a back end standpoint. So I think there's this whole new world of composable commerce where you can stitch together a bunch of different APIs and have it such that when an end customer goes and transact, you haven't pre-bought credits, you haven't pre-set up agreements. It just goes in and goes in goes 1, 2, 3, 4, interact with all of them and provide me the end solution. And I think it will enable you to build a much more efficient lean, well-constructive company. I think I couldn't agree with that more. No, one of the interesting consequences of instance settlement like we're talking about is you can instantly settle across a complex chain of transactions. Yep. Right, we're like one purchase depends on another, depends on another. Why can't a merchant that I'm transacting with a topically settled with an upstream dependency that they need to buy per use on my behalf? It completely changes the cost structure of the merchant. And then the composability of it, I couldn't agree more there too, is if you can pre-fund all of a sudden, you have access to a bunch of things without requiring up from configuration and buy in from the user. So what you end up with is much more complex final payments, multi-party payments, multi-merchant payments. There's just a whole bunch of things that we don't think about because we have to take for granted that the payment goes to one person who has the problem of settling everything else and they have to have a relationship with me and there's all types of people who need to underwrite the risk of dealing with me. It seems solvable now. Totally agree. Amazing. Well, thank you everybody. This has been a great conversation. And we're going to have to have you back on, Sam, to find out if your prediction bears out. Or if you become one of five years. OK.