Conversation with a16z Crypto: What Will the Era of AI Shopping for You Look Like?
TechFlow Selected TechFlow Selected
Conversation with a16z Crypto: What Will the Era of AI Shopping for You Look Like?
After AI shops for you, you’ll no longer see ads.
Navigating Web3 tides with focused insightsTranslation & Compilation: TechFlow
Guests: Eddy Lazzarin, CTO of a16z Crypto; Noah Levine, Investment Partner at a16z; Sam Ragsdale, Founder & CEO of Merit Systems
Host: Robert Hackett
Podcast Source: a16z crypto
Original Title: The end of ads? AI agents are about to change how we buy
Air Date: April 28, 2026
Editor’s Note
This episode brings together Eddy Lazzarin, CTO of a16z Crypto; Noah Levine, Investment Partner at a16z; and Sam Ragsdale, founder of Merit Systems (and formerly of a16z), who is building Agent Cash. They dive into a high-density discussion covering the current state of AI agent technology, payment infrastructure, and the future—or demise—of the credit card system. Their core thesis: stablecoins’ instant settlement and zero marginal cost make them uniquely suited for microtransactions in the agent economy—transactions valued at $1–$2—whereas credit cards’ fee structure (2–3% marginal fee + $0.30 fixed fee) becomes untenable. Agent Commerce is dismantling the ad-driven business model that has defined the internet for two decades. Eddy Lazzarin goes so far as to declare: “The advertising economic contract is dead—and will vanish entirely within ten years.”
Key Quotes
The Essence of AI Agents
- “An LLM is a chatbot. An agent is a chatbot that operates your computer on your behalf. Anything a human can do with a computer, an agent can do too.”
- “Starting around November last year, AI models got smarter. They can now execute complex tasks over extended time horizons—and they use tools. We began calling them ‘agents’ because they don’t just write code—they help you complete entire tasks.”
- “Internally, we call this ‘just-in-time natural language programming.’ A user describes a need in plain language; the agent writes, behind the scenes, a potentially thousand-line JavaScript program to carry it out—costing roughly $0.20 in tokens and $0.10 in API calls—and discards the program once done. Four years ago, this would have required an expensive software engineer spending a week debugging, securing API keys, and more.”
Headless Merchants and Commercial Restructuring
- “What does a headless merchant look like? It serves AI—not humans. No website frontend: only well-documented API endpoints that models can read, understand, and invoke.”
- “Data industry leaders charge up to 100× the lowest-priced downstream data source—using that same underlying source. Their real product isn’t data—it’s their enterprise sales team. In a world where agents make decisions, agents won’t be swayed by polished sales teams. They’ll test every data source, identify the most effective and cost-efficient one—and remember it.”
- “You excitedly set your agent running overnight. At 9 a.m., you wake up—and discover it’s been stuck since 2:30 a.m., waiting for you to call an enterprise sales team.”
The End of the Advertising Model
- “The internet’s economic contract since 2000 has been built on distraction. Agents don’t get distracted. If an agent visits your site looking for a recipe, it won’t see the shoe ad beside it. This old model will die within ten years.”
- “In 2016, global internet ad spend totaled $60 billion—and everyone thought that was the ceiling. Today, Google alone earns $300 billion annually from ads. Yet after GPT-4 launched, traffic to tech news sites dropped ~80%; Stack Overflow saw similar declines. These early adopters have already chosen agents for information retrieval and code execution. Others will follow—because the experience is simply better.”
Stablecoins vs. Credit Cards
- “The average transaction on Agent Cash is $0.01–$0.02. Credit cards charge a flat $0.30 fee. That fee structure is absurd in this context. In 2026, loyalty should accrue to merchants—not to the card you used to pay.”
- “Credit cards predate the internet—and successfully survived its emergence. Though battered during the transition from offline to online, they endured. So the verdict isn’t final yet.”
- “If any credit card executives are listening: you hold money transmission licenses—you could instantly mint stablecoins on behalf of customers and let them pay with stablecoins. I strongly recommend considering this.”
The Future of Consumer Experience
- “If an agent shops for you, and you equip it with a credit card optimization skill, you can now precisely track ROI per card. With zero credit card loyalty, all psychological lock-in effects vanish.”
- “One day, you’ll realize—you never actually enjoyed shopping.”
Architecture of the Open Agent Commerce Stack
Host: Hello everyone—and welcome. Joining me today are Eddy Lazzarin, CTO of a16z Crypto; Noah Levine, Investment Partner at a16z; and Sam Ragsdale, former a16z Crypto colleague and founder of Merit Systems, which is building Agent Cash—we’ll dive deep into that shortly.
Before we do, let’s lay some groundwork. So much is happening in the AI agent space right now—if you’re not tracking it 24/7, you’ll fall behind fast. So: what’s the current state of the world? Sam, you’re building on the front lines—why don’t you kick us off?
Sam Ragsdale: I like to start with a taxonomy borrowed from Erik Reppel, co-creator of Coinbase’s x402 protocol.
This framework splits agent commerce into two categories. First is conversational commerce—checking out inside ChatGPT. You tell ChatGPT: “I’m a man living in New York’s West Village, going to Equinox for fitness, and want shoes that fit my social circle.” It empathetically recommends Nike—and you buy.
The second category is delegating money to an agent, empowering it to spend on your behalf to accomplish tasks.
Conversational commerce will certainly happen. ChatGPT, Gemini, Claude, and every next-generation frontier model will add checkout functionality. It benefits consumers (better discovery), merchants (higher conversion), and platforms (5–10% take rate)—essentially a next-gen Google Shopping.
The other world is where agents’ capabilities remain limited. Many ask agents to perform difficult tasks—e.g., “Help me do outbound sales”—and the agent replies: “I can’t—I don’t have access to that info.” But if the agent holds even a tiny balance, it can spend fractions of a cent to acquire services it couldn’t otherwise access—making itself significantly more capable.
So today, two parallel worlds exist: one where traditional LLM interfaces recommend products and handle final checkout—platforms taking a cut; another where you independently deploy an agent to purchase goods and services on your behalf.
Noah Levine: I see two versions. One is e-commerce’s natural evolution—the platform shift. Commerce migrated to mobile; new ad formats and Google Shopping emerged. People always need to buy things; consumer behavior changes—and now, with LLMs reshaping how people access information, commerce naturally follows into agents.
The other version is less “physical-world analog”: the internet itself is transforming. How people access information and execute actions is shifting with LLMs. The internet we’ve built over the past 20 years may not be the internet of the future. Searching via Google and clicking into a heavily upselling web UI may no longer make sense. Instead, we’ll see a more agent-native internet—where agents directly pay for what they need, making them vastly more efficient for humans.
Host: That directly connects to one of your investment theses, Noah. But before we go deeper, let’s ground listeners with a primer. Most are familiar with interacting with LLMs—but now hear about OpenAI’s Codex and others, which exhibit significant autonomy and real-world agency. If you haven’t followed closely, you might not realize how far the technology has advanced. Eddy, why don’t you walk us through it?
Eddy Lazzarin: Let me recap the past five months. Starting around November–December last year, AI models got smarter. Specifically, they can now execute complex tasks over long time horizons—and use tools. We started calling them “agents,” a somewhat anthropomorphic term, because they don’t just write code—they help you complete full tasks.
But agents aren’t omnipotent. Software isn’t just a small program running on your laptop. The internet taught us that doing interesting things requires connecting many other systems—networks, participants, etc.
Agents solve intent formation—and partially solve preference modeling. You tell them something, and they infer your goal, mapping it onto tools, networks, and services. Through conversation and memory, they develop a rough understanding of your preferences—and propagate that intent to tools, software, and suppliers.
These two problems—intent formation and preference modeling—have been solved, and it’s incredibly exciting. Everyone wants to tackle the remaining challenges—but those are complex. At minimum, enabling agents to transact for you requires solving authorization and delegation: How do you prove to a counterparty that this agent represents you? How do you handle identity and authentication? Then there’s payment and settlement: Once connected, and once the agent reflects your intent and knows what to do, it needs to pay—to demonstrate payment capability, handle split payments, refunds, and more. I’m skipping search, anti-fraud, and other critical layers—but you can see: once intent formation and preference modeling—two things previously requiring human input or at least articulation—are automated, the entire commerce workflow becomes automatable. That’s the engineer’s reaction: “Wow—these two things humans had to manually input or articulate are now fully automatic. Incredible.”
When people talk about “agentic commerce,” they mean everything between “I talk to the agent” and “it delivers what I need”—what remains to be solved, and the cascading implications, because so much will be rewritten.
Host: Extremely helpful. So we’ve evolved from LLMs that interact via natural language to enhanced versions that connect to diverse networks and real-world systems.
Eddy Lazzarin: It’s not just about connectivity. Saying it that way implies the change lies in what it connects to. That’s not it. Your laptop was already connected to everything—the connectivity hasn’t changed. What changed is that models can now use tools, think for extended durations, and persistently iterate until tasks succeed.
Sam Ragsdale: Let me simplify your simplification. An LLM is a chatbot—great at dialogue. Historically, people assumed its best fit was customer service. After pushing dialogue to its limits, we added tool use. Simplified: it learned to operate your computer. An LLM is a chatbot; an agent is a chatbot that operates your computer for you.
The key point: around GPT-4, agents reached human-average computer operation proficiency—at roughly 1/1000th the cost—and scaling ability expands dramatically with added budget. So broadly speaking: anything a human can do with a computer, an agent can do too.
Eddy Lazzarin: Exactly. The premise is simple—but the implications are vast: short-, medium-, and long-term. Short-term, everyone’s building pipes to let agents truly act. Long-term—if your agent can access apps, how many UIs or interfaces do you really need? Do you still need the Amazon app? Maybe not—maybe it’s better to let your agent do all the legwork: read all reviews, show only images you care about. Isn’t that superior?
Sam Ragsdale: Internally, we call this “just-in-time natural language programming”—though the name doesn’t quite stick. It turns non-programmers into programmers. You type: “Buy something for my fiancée on Amazon. Here are her preferences, here’s what I usually buy her, here’s what I bought last time—browse ~1,000 options, pick the best match, order it, find my home address, and ship it.”
What actually happens is the agent writes an internal program to execute this complex task—possibly a thousand-line JavaScript and Bash script. It runs, stays invisible to you, then gets discarded. Four years ago, this was science fiction. Writing such a program required an expensive software engineer spending a week debugging, acquiring API keys, etc. Today, execution costs ~$0.20 in tokens, maybe another $0.10 in API calls—and the program is thrown away after use. It’s cheap enough that you wouldn’t even upload it to GitHub. Non-technical people can do this. My parents are writing natural language programs now—and don’t even realize it. They might soon call themselves software engineers.
Host: Pretty wild. Are you engaged? Was that example drawn from your own life?
Sam Ragsdale: Yes, I’m engaged—thanks! Though I didn’t let AI buy the ring. That ring predates AI—and possibly even the first computer.
The “Headless Merchant” Thesis
Host: Great. Now let’s explore these ripple effects. Sam, earlier you mentioned how commerce transforms in a world where agents conduct massive volumes of transactions—leading directly to your concept of the “headless merchant.” Tell us: what is a headless merchant?
Sam Ragsdale: Good question. Let’s step back first. Beyond traditional consumer use cases like buying shoes via ChatGPT, there’s a massive B2B developer tools market. Platforms like Claude Code and OpenAI Codex are democratizing development—anyone with a computer and tokens can build.
Previously, experienced developers selected tools deliberately—often engaging enterprise sales teams and signing contracts. Now, new developers enter with only intent—“I want to do X”—no preconceptions about which resources to use. And what they build is highly ephemeral, demanding pure usage-based pricing—not months-long onboarding processes.
So what does a headless merchant look like? It serves AI—not humans. No physical or digital storefront for browsing—just well-documented API endpoints that models can read, understand, and invoke. Billing is per-API-call—not subscription-based or enterprise-contract-based.
Eddy Lazzarin: I completely resonate. I suspect I was an AI in a past life. As a software engineer, I’ve always been like this: if I land on a website and can’t immediately see pricing or find a credit-card-powered API-key-onboarding flow, I close the tab. I don’t want to talk to sales teams or send emails. Scheduling meetings with enterprise sales is a huge commitment and slowdown. I don’t even know if the thing works yet—I just want to try it now, immediately, because I’m building something over the weekend and want to launch Monday. Swipe the card, get the key, expense it later, plan later—that’s speed.
In the era of instant, temporary software—do you really want your agent waiting? You excitedly run your agent overnight. At 9 a.m., you wake up—and discover it’s been stuck since 2:30 a.m., waiting for you to call an enterprise sales team.
Sam Ragsdale: And if enterprise sales are part of the onboarding flow, that API’s price likely jumps tenfold—because someone must manage customer relationships.
Eddy Lazzarin: Absolutely unacceptable. You want agents to run autonomously—not because you don’t care about your work, but because you need speed, testing, and rapid iteration based on user feedback. You can’t wait. If an AI model sees three options—one requiring enterprise sales contact, one needing a dedicated credit card, and one letting you send stablecoins to get $10 in tokens for a proof-of-concept—it picks the third every time. That single force alone is enough to restructure parts of the market.
Host: For traditional enterprises, while these frictions make business harder, they also rely on them to lock in customers and sustain loyalty. If those frictions vanish—how do you reliably forecast revenue?
Eddy Lazzarin: Let me give you my blunt answer: then let’s make everything terrible. Add friction everywhere—make everything hard to use. What are we even doing?
I say this because friction sometimes helps—e.g., blocking spam, creating filtering effects. But friction carries massive costs. As economies accelerate, productivity rises, and time’s leverage amplifies per minute, friction’s opportunity cost rises too. That’s the trend across everything right now.
Back to the point—even in the lowest-friction environment, where you get an API key in one second—or don’t need one at all, paying directly with an encrypted wallet key, where your wallet address is your account—other factors still create stickiness. Reputation, memory, state, data—even intangible things like agent trust. If an agent knows you urgently need answers and want to move fast, it won’t waste 20 minutes exploring all new options. It’ll recall what worked well last time and reuse it—just like a smart human.
Sam Ragsdale: Let me give a concrete example. We speak daily with numerous merchants—having seen virtually all currently API-sellable products—and spoken with many sellers about “agent-native distribution”: native distribution designed for AI agents.
Data products are often commoditized—with 5–50 sellers. In this group, the top seller charges up to 100× the cheapest option—and frequently uses the same downstream data source. They achieve this via enterprise sales teams—staffed by polished professionals who fly to your office to demo: “Look how beautiful our data is—nothing compares—$35,000/year.” You sign a two-year contract; when it expires, that person flies back and repeats the demo. Tens of thousands of companies pay this way.
Meanwhile, smaller companies offering better products—superior usability wrappers atop identical data—fail due to lack of distribution channels, eventually going bankrupt. Innovation stalls, because the enterprise sales team *is* the product—not the data.
In a world where agents choose, agents won’t chat with sales teams or be dazzled by polished demos. They’ll test every data source, identify the most effective and best-priced (especially at scale), and remember: “Next time I need this data, use Minerva—not the other three.” This creates a more efficient world. Companies previously overpaying $35,000 can redirect that money toward productive investments.
Noah Levine: Another angle: if you believe AI will spawn countless solo founders or ultra-small teams—building products that previously required 50–100 people—then enterprise sales teams flying to someone’s basement to negotiate deals makes zero sense.
Yes, existing merchants fear revenue predictability disruption—and resistance is natural when change arrives. But this is also a brand-new customer acquisition funnel. Reducing tool onboarding bottlenecks and friction presents a massive opportunity for them.
Sam Ragsdale: On our demand side, most users have never used an API—don’t know what an API is, what it represents, have never held an API key, never signed an enterprise service agreement. Yet on first use, they can combine six APIs from six different merchants, write a natural language program to complete a task—and discard the program upon completion. This signals an entirely new API consumer market.
The Internet’s Existing Business Model Will Be Restructured
Host: This sounds like Clayton Christensen’s innovator’s dilemma—high-end markets served by legacy players selling expensive software to big-check clients, low-end markets served by agents conducting one-off experiments. But what transforms this from a “toy” into something genuinely disruptive?
Sam Ragsdale: Because it ultimately delivers a better experience.
Noah Levine: I’d add: though experimental today, historical platform shifts show similar patterns. Stripe began serving extremely small, long-tail merchants—many of whom later became giants—fueling Stripe’s sustained growth. Shopify did the same—starting with drop-shipping and T-shirt sales—now powering brands that scaled from zero to major success on its platform. Similarly, we’ll see lean, AI-powered founders building large companies—whose early agent-commerce tool purchases will scale massively as their businesses grow.
Sam Ragsdale: That e-commerce lens is great. But I’m talking bigger: the internet’s economic contract is dead.
Since Google launched in 2000—becoming the biggest catalyst for the “free, open internet”—the economic contract was: you’re a publisher; you publish great content; users find it via Google; a few years later, AdWords launched, adding banner ads. The contract evolved: publish great content; users land on your site; you run small ads; Google shares revenue based on impression quality. You could publish anything users wanted; Google managed advertiser relationships and paid you back. Google became the biggest catalyst for the free, open internet—needing it fast, cheap, and ubiquitous—because the more people searched, the more Google earned.
Ultimately, the internet’s business model was “distraction.” As a human user consuming content—whether searching for info, recipes, or scores—you’d get distracted—and perhaps later buy those shoes or learn about a new B2B SaaS. This model scaled beyond anyone’s expectations. I just checked the 2016 “Internet Trends Report”: internet ad spend was $60 billion—and people said, “That’s the ceiling!” Yet Google now earns $300 billion annually from ads alone.
But with agents, people are shifting search, information retrieval, and execution to agents. It’s early: ChatGPT has 100M monthly active users—but they’re still using it like Google Search, not agent-style (“Find my dad a Father’s Day gift and order it”). That’s coming. Look at tech-sector data: since GPT-4, tech news site traffic dropped ~80%; Stack Overflow saw similar declines. These early adopters have already chosen agents for information retrieval and code execution. Others will follow—because the experience is demonstrably better.
The old business model is being abandoned. Agents don’t get distracted. If an agent visits your site for a recipe, it won’t see your shoe ad—and publishers gain nothing. A new contract—and new reason to serve agent requests instead of ads—will be needed. Will articles be paid for directly? Unclear. Will API resources be paid for directly? Will the internet become unrecognizable? Also unclear. But the old model will die—within ten years, guaranteed.
Host: If the internet’s business model is fundamentally about distraction, that’s fascinating—because Google originally emerged as anti-portal. Yahoo and AOL offered cluttered link directories; Google delivered a clean search box and fast results. Yet your described evolution shows it becoming a distraction machine.
Now we say agents won’t get distracted—but why should agent evolution differ from human evolution? Could mechanisms emerge specifically designed to lure agents—making them linger longer?
Eddy Lazzarin: That’s a profound and fascinating question—centered on: who does the agent represent? I recently heard someone say, “I’m using Google Search again—the AI answers at the top are good enough.” In that scenario, the “agent” works for Google—it lives in Google’s search bar, runs on Google’s cloud, and Google controls it. Will that agent be “distracted” by Google? I suspect yes.
The crux is whose objective function it optimizes—or put plainly: who does it work for? “Distraction” means showing you something that serves *my* interests—not yours. If it serves mine, not yours, it’s distraction.
I’m less pessimistic. The longstanding industry consensus—that good ads *are* good content—has held for decades. Good ads are nearly indistinguishable from content you already want.
But let me clarify: if an agent works for Google—or anyone—it will follow their entire commerce chain, using their methods and transaction infrastructure—optimized for *their* business interests. If an agent works for *you*—running on your laptop, open-source, fine-tunable, modifiable system prompts—you can equip it with anti-distraction tools. Advertisers then face an adversary that sees through their tactics. I’m exaggerating slightly—but fundamentally, adversarial dynamics will emerge.
Sam Ragsdale: Yes—there are infinite ways to reintroduce ads. You could bake them into model weights—the most aggressive approach. Choose training data saying “Nike is the world’s best shoe.” Nike could pay $1B/year—so whether in ChatGPT or a car-insurance company’s enterprise API, any shoe-related query yields “Nike is best.”
You could embed ads at the tool-calling layer, in system context, or as overlays that never enter chat. Foundational model companies are clearly wrestling with this. Recently, Anthropic and OpenAI clashed—Anthropic ran a Super Bowl ad mocking ChatGPT’s ads; OpenAI pulled its ads shortly after. Yet OpenAI’s response struck me as reasonable: “ChatGPT’s free users in Texas alone outnumber Anthropic’s total paid users.” This is a different scale entirely—they’re delivering expensive frontier tech to vast numbers of users unwilling to pull out credit cards. Ads are a rational solution.
Ads succeeded on internet search because consumers paid nothing. High-friction relationships—like credit card payments—exist only between advertisers and Google/publishers—not the billions of monthly search users. Those users get value instantly by opening Google.
If you align incentives, separate ads, and maximize relevance—you actually get a better experience. Foundational model companies are moving away from ads. ChatGPT runs no ads; Gemini hasn’t launched ads yet. Google is most likely to do so—it’s done it before and remains the largest ad operator. Gemini will inevitably add ads; its massive MAU and equivalents to Google Shopping will follow. But they know monopoly isn’t here yet—all companies compete fiercely, burning private-market subsidies. They don’t want to be labeled “this model lacks empathy for your goals because it runs ads.” So for now, no one runs ads—everyone strives for neutrality.
Noah Levine: Another angle: as merchants improve price and product data transparency, you can shift ad budgets toward agent-specific discounts. If the agent is the buyer, convert ad spend directly into discount budgets.
Another branch: what will the discovery layer for agent commerce look like? Who handles discovery? How do you differentiate merchants? My prediction: if ads weaken as agents become buyers—and attention is no longer scarce because agents possess infinite attention—merchants may try “covert advertising” via discounts or optimized descriptions that agents better understand.
Eddy Lazzarin: There are too many dimensions. Ads are merely one conversion-acquisition method. If a system achieves higher conversion without ads, it will. Systems indeed have many alternatives: recommendation networks, discounts, coupons, exclusive channels, free tokens for startups. Hundreds of acquisition methods exist—ads are just the most visible, because they hit ordinary users most directly.
Turn personalization to eleven—if someone wants to reach me, they’ll first talk to my agent—and my agent will tell them: “Eddy hates ads intensely.”
Stablecoins vs. Credit Cards in Agent Payments
Host: Before we wrap, I must ask two questions. First: to what extent can traditional payment rails adapt to agent commerce—or do we need entirely new, native rails like stablecoins, which appear to be finding product-market fit?
Sam Ragsdale: My overall view: for e-commerce or conversational commerce—“new physical-world analog” checkout scenarios—credit cards work well. They include built-in consumer protections: if shoes don’t arrive or get crushed by a truck, Visa arbitrates and refunds you—the risk falls entirely on the merchant. This is a sound trade for novel goods and services.
But stablecoins excel in another domain. Agent Cash’s average transaction is $0.01–$0.02—nearly 600,000 such transactions completed. Credit cards charge a flat $0.30 fee. Wire transfers cost ~$1. Marginal fees of 2–3% mostly fund cashback rewards. For e-commerce, you may love rewards and airline miles to Miami—funded by merchant fees. But when buying $0.01–$0.02 items—sporadic API calls—stablecoins offer zero marginal fees and sub-cent fixed costs.
Another key point: instant settlement. When purchasing goods/services online, settlement cycles are month-end—whether invoices, wires, or credit cards—meaning merchants extend credit to customers or agents. In the agent world, you often don’t know who the agent is.
Specifically, anyone using Anthropic or ChatGPT APIs knows the tiered system—first $50, then $100, up to $2,500. This exists because providers extend credit—they don’t know you, haven’t conducted KYB or credit checks, and don’t know if you’ll pay at month-end. AWS and Nvidia GPU rentals operate similarly. Month-end settlement is terrible for this use case—merchants bear all risk. If the customer isn’t a real company signed to an enterprise service agreement, but an agent—you have no idea who it is. Overnight, someone could generate a billion agents—but you can’t grant credit to an agent. Some are building agent credit solutions—I think that’s misguided. Instant settlement solves this directly. Instant settlement is like cash: I have it, hand it to you, you have it. You deliver goods/services—I can’t reclaim the money. Tiered fixed fees with instant settlement are superior for microtransactions and this transaction nature.
Noah Levine: One point worth challenging: minimum transaction fees and credit cards’ ability to support microtransactions depend ultimately on card networks’ pricing decisions. If they launch a new transaction type—e.g., “microtransaction”—with no minimum fee and reduced rates, it’s entirely feasible. Advantage: far more consumers hold credit cards than understand stablecoins. So developers could pay with cards while stablecoins settle in the backend. But that takes time. Until then, using native wallets to spend stablecoins directly on these protocols makes strong sense.
Sam Ragsdale: I think it’s extremely unlikely credit card companies will disrupt their 80-year core business model. But I’d love to see it happen.
Eddy Lazzarin: I agree there’s no strict technical barrier for credit cards. But the issue is subtler—rooted in business models and consumer perception of credit cards. Recently, I saw concepts for “agent credit cards”—essentially virtual cards extended. I love my card issuer’s virtual card feature: generate disposable card numbers anytime, kill them instantly for fraud or tricky subscription cancellations.
Yet sometimes new platforms or methods win—not because they’re technically indispensable, but because they’re tailor-made for new contexts. Credit cards predate the internet—and survived its emergence, albeit battered. So the verdict isn’t final.
Noah Levine: Also, technologies enabling Apple Pay will similarly enable agent commerce. Will this disrupt Visa or Mastercard? My intuition: many B2B transactions between developers and enterprise APIs already settle via wire transfers. If card networks capture this volume—monetizing via microtransactions rather than $100K–$1M monthly settlements—it’s a huge opportunity—even if per-transaction fees are tiny.
Eddy Lazzarin: Fully agree. Personally, I prefer stablecoins. I’m biased toward crypto—but stablecoins offer conceptual elegance: knowing your balance, controlling it regardless of amount or network, enabling instant payments. If I could avoid credit cards entirely, I would—I’m spoiled by direct control over my money.
Sam Ragsdale: I’m even more pessimistic than Eddy. Transaction fees strike me as utterly insane. Consumer protection is great—but that’s the $0.30 fixed fee. The bulk of marginal fees is the interchange rate—the money merchants’ banks pay consumers’ banks to fund rewards and cashback. This originated in the 1980s to ensure credit card network effects took off—getting everyone to carry cards. Visa gave you global purchasing power—this was the cost of building the network. Early interchange rates hit 8–10%, with 8–10% of merchant fees flowing back to consumers’ banks for rewards—keeping users on Visa. Pre-credit cards, payments were chaotic—so 8–10% felt reasonable.
Now it’s down to 2–3%—5% for luxury jewelry. But in 2026, this concept seems especially absurd: loyalty belongs to the card you pay with—not the merchant. In Eddy’s world—where you generate disposable virtual cards per account, with zero credit card loyalty—the network effect is established; everyone already has cards. That 2–3% should belong to merchants—enabling them to run their own loyalty programs: discounts, “buy 10 get 1 free,” etc.
Today, Visa/Mastercard will say this already exists—if you’re Lululemon, negotiate a BD partnership, endure a three-year legal process, and your Amex points can be spent once monthly at Lululemon for a free pair of shorts. But that doesn’t work for small merchants. Joe’s sandwich shop still pays the interchange fee. And those fee schedules run seven pages—no one understands them. A famous Stripe article notes that Stripe’s world-class payment experts spent four years building a predictive interchange model—because it was previously impossible to calculate.
I estimate dozens of entrepreneurs have already tried fighting interchange—and failed. But I think it may finally fall in the next decade. Talk to credit card folks—they’re 100% certain consumers love this psychological manipulation—thinking 3% equals 10% value, loving their points. But what you actually get is worth far less than 3%—and depreciates over time. I believe this era is ending. We finally have a better rail. No reason for these massive companies to persist. No reason for five companies to handle each credit card transaction—each bearing distinct risks. Stablecoins are like cash: I hand it to you—you have it. No justification for marginal fees; fixed fees under one cent.
Historically, everyone predicting credit cards’ demise looked foolish. But I’ll be that person.
Host: I must interject with one of my favorite business stories. Ron Johnson—former head of Apple Retail—joined JCPenney, scrapped all coupon/discount systems, simplified pricing to “what you see is what you pay,” and assumed consumers would love it. It flopped spectacularly—because consumers genuinely enjoy the coupon-and-discount game. So I wonder: will agent commerce repeat this? Do people truly love credit card points and silly games—or do they want a simple, transparent experience?
Sam Ragsdale: If an agent shops for you—and you equip it with a credit card optimization skill—it calculates precise ROI per card. With zero credit card loyalty, all psychological lock-in vanishes.
Eddy Lazzarin: Fully agree. If the “game” itself is the selling point, I won’t pay an agent to play dumb coupon-cutting games. If consumers truly want that experience, those games will migrate elsewhere. Ultimately, if consumers need that experience, you can reallocate savings from that layer to deliver it elsewhere.
Sam Ragsdale: Maybe I’ll have to eat those words later.
Noah Levine: I think you’re spot-on. Interchange persists partly because it’s ancient—and credit cards carry enormous inertia. My counterpoint: it’s not about whether merchants think it’s ideal or not—it’s about what consumers ultimately want. For most merchants, payment priority isn’t fee cost—it’s appearing on channels consumers use, ensuring acceptance of preferred payment methods. If storing a card and spending existing balances feels easier than setting up a wallet and funding stablecoins, consumers will use cards—and merchants will accommodate them.
Sam Ragsdale: If any credit card executives are listening: you hold money transmission licenses—you could instantly mint stablecoins for customers and let them pay with stablecoins. I strongly urge you to consider this.
Noah Levine: If credit cards play a role in agent commerce, I suspect the outcome is “AI mullet”—business up front, stablecoin in the back. Consumers spend via cards and balances; merchants receive stablecoin settlement on x402 or MPP—combining both advantages.
Consumer on-ramps are improving rapidly—acquiring stablecoins grows easier daily. On the enterprise side, most modern banks are launching native stablecoin support. Friction in acquiring stablecoins approaches zero: open a bank account, complete KYB and credit-risk processes, and directly buy and use stablecoins—via Mercury-issued cards or direct account withdrawals. Any entity holding an MTL (Money Transmission License) can easily add stablecoin support. Conveniently, credit card companies are exactly that type of entity.
Host: We’ve noted consumer preference may determine the path forward. Sam—what product or service do you think will truly drive mainstream adoption of agent-based transactions?
Sam Ragsdale: I’m most excited about a new class of companies—built by ultra-lean teams leveraging AI to create excellent products. Their construction method itself is fascinating: “composable commerce”—assembling diverse APIs, where end customers place orders without pre-purchasing capacity or pre-signing agreements—triggering sequential API calls to deliver the final result. This enables building more efficient, leaner, better-structured companies.
Eddy Lazzarin: Absolutely agree. An intriguing consequence of instant settlement is atomic settlement across complex transaction chains. Why can’t the merchant I transact with instantly—and atomically—settle with their upstream supplier per transaction? This radically transforms merchant cost structures. Composability also improves: with pre-funding, you can access vast resources without users configuring or purchasing upfront.
Ultimately, you get more complex payments: multi-party, multi-merchant payments. We don’t contemplate these today—because we assume payments go to one party, with others’ settlements being that party’s problem—requiring relationship-building and extensive risk management. Now, all these seem solvable.
Host: Brilliant. Thank you all—and we’ll bring Sam back in five years to check his predictions!
Sam Ragsdale: See you in five years.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
深潮TechFlow
