AI agents get virtual credit cards

0:00Band-Dool is now available everywhere in the U.S. So you can get in on the sports action no matter where you are, including these places. May this year most rewarding season ever with Band-Dool. Agent location restrictions apply. Product availability varies by market. C-FanDool.com for eligibility, gambling problem called 1-800-Gemmar. It's football season and you can now get almost anything you need for game day delivered with Uber Eats. What do we mean by almost? You can't get a running back delivered, but you can get baby back ribs delivered. A strong defense? No. A strong deodorant? Yes. The six pack of abs? No. Six pack of beer? Yes. Get almost almost anything for game day delivered with Uber Eats. Official on-demand food delivery partner of the NFL. Order now. Alcohol 21 plus in Inselect City's Product Availability Various by Regency app for details.

1:00Propel Fitness Water. With Gatorade Electrolites, Zero Sugar, and Vitamins. Propel hydrates better than water to help you get the most out of your workout and get back to your best self. What propels you? Propel with Gatorade Electrolites. MasterCard is teaming up with a startup called Alchemy to issue credit cards directly to AI agents. Visa and American Express are doing the exact same thing. Thinking about the actual mechanics of that is... Yeah. It's a little strange. I mean, you have a piece of software that now just holds a virtual credit card. Yeah. And that card is linked directly to your bank account. Right. So the software can spend your money using your personal data. And it doesn't have to ask for your permission on every single transaction. You just sort of let it run. We are moving way past the model where an AI just gives you, you know, a shopping list or offers a bank account. It's not just an advisor anymore. Exactly. The algorithm is no longer just an advisor. It actually executes the transaction. It operates as an autonomous buyer in what the financial sector is calling a gentick commerce.

2:04A gentick commerce. Yeah, the machine is the customer now. So if the customer journey is essentially being delegated to algorithms, we really have to figure out what happens to trust and, I guess, accountability when a financial transaction happens entirely between two machines. Right. Because you just have a server talking to another server. And your money changes hands without a human ever looking at a screen or clicking a checkout button. The immediate problem is preventing these bots from just emptying our bank accounts because of a bad prompt. Yeah, because they definitely will. If you're not subscribed yet, take a second and hit follow on whatever podcast app you're using. It helps us keep making this. We appreciate you being here. So how are they actually stopping a bot from draining an account? The payment networks are essentially building a leash. They enforce strict rules at the network level. So they set total spend caps, restrict purchases to allow product categories. And they hard code a kill switch into the system. A kill switch?

3:04Yeah, they're building something called a gentick tokens, which rely on this concept of verifiable intent. Verifiable intent. Okay, what does that actually mean? Basically, the intent of the purchase actually travels alongside the purchase data when the card is swiped. I sort of struggle with the trust aspect of that whole mechanism. Because large language models hallucinate constantly. I mean, they make unforced errors. They misinterpret really basic instructions. Sure. A bank cannot genuinely trust an AI not to glitch and attempt to buy a car instead of a weak swarth of groceries. You tell it to stock the fridge and it decides you need a new refrigerator entirely. Well, that's where the technical separation comes in. That's how they are handling that specific risk. The spending limit lives completely outside the AI model. Outside the model. Right. It does not matter how smart the AI is or how confused it gets. It cannot talk its way past a network-enforced cap. Because the network isn't an AI. Exactly. The card simply clears or it does not clear.

4:05The restriction is structural, not cognitive. The network really does not care what the AI thinks it is doing. It's kind of like handing a teenager a $20 bill with a note to the cashier saying the money is only for milk. Yeah. That's a good way to look at it. You aren't trusting the teenager's impulse control or their judgment. You are trusting the cashier to read the note and enforce the rule. The entire system is basically built on the assumption that the agent will misbehave at some point. The only problem with the teenager analogy is that a teenager knows when they're breaking the rules. Right. They know they're trying to sneak a video game past the cashier. Yeah. An AI agent might genuinely believe that purchasing a completely unrelated item fulfills your prompt just because of a logic error in its processing. It literally does not know it is messing up. That makes the verifiable intent token so crucial then. Exactly. It forces the AI to declare what it is doing in a way the network can actually understand before any money moves.

5:05The network acts as the cashier and it declines the transaction if the agent tries to buy something outside the pre-authorized intent that you established. So the network holds the agent to your original parameters? Right. The network level safeguards address the technical vulnerability but human psychology around money operates on a completely different axis. Yeah. Solving the math problem definitely does not solve the human hesitation. There is a very clear mental divide for consumers here. Like automating utility bill payments makes sense. Sure. Routing a set amount of money into an investment account every month, that makes sense. Those are highly predictable repetitive actions. But letting an AI choose and buy physical products, actual shopping. That feels like crossing a boundary. Because you're giving up control over preference. Right. Think about how you actually shop for something like running shoes. You do not just look at a spreadsheet of specifications. You know, you have irrational brand loyalties. Exactly. You know, a color, the aesthetic, maybe, you know, a commercial you saw three years ago that just made you feel a certain way about a specific brand.

6:08Shopping is totally subjective. And delegating the discovery, the comparison, and the recommendation phases to an algorithm entirely breaks the traditional retail model. If people actually start delegating those phases to an algorithm, the entire industry has to completely restructure itself. We are looking at a situation where the concept of marketing has to fundamentally change. Marketing has historically been about convincing a human being through emotional appeals. You want the consumer to feel something. Yeah. But an AI agent does not feel anything. It parses data. Right. So if you are a brand, how do you earn preference from an algorithm? You can't. Not the traditional way. You cannot run an escologic television commercial for a piece of software. A super bowl ad is completely useless. If the entity making the purchasing decision is just a string of code evaluating price to durability ratios. Marketing shifts from influencing human emotion to convincing a machine with parameters, data sets, and optimization metrics. You basically have to market to the API.

7:09Yeah. Brands will have to figure out how to structure their product data so that an AI agent prioritizes it. If your customer says buy me a good pair of running shoes, your brand needs to be the mathematical answer to the word good. But good is entirely subjective to the user, which means the AI has to learn exactly what you, the specific human, consider good. If the AI buys a brand of running shoes you hate just because the algorithm prioritized a discount over your historical preference, you're going to be furious. You'll just pull the plug on the AI entirely. Exactly. That is why trust has to be engineered into every single machine decision in that chain. The AI platform loses if the consumer unplugs it. The payment network loses the transaction volume. And the brand loses the sale. The entire ecosystem relies on the AI making the decision that the human feels good about after the fact. I am trying to imagine what that feedback loop even looks like. Do you have to train your shopping agent like you train a puppy? Basically, yeah.

8:09Do you review its purchases at the end of the month and say good job on the paper towels, bad job on the coffee beans? You probably will have to initially. You will correct it, adjust your parameters, and over time it will just build a localized model of your exact preferences. It will know you prefer organic coffee, but you don't care if the paper towels are generic. Right. But that level of personalization is incredibly difficult to achieve without making mistakes along the way. And in Thomas, mistakes cost money. The retail side of this is really just the visible half. It is what consumers can easily imagine because we already shop online. Right. The real volume of AI spending will not be on consumer goods like coffee and shoes. It will be entirely invisible. The end tool is now available everywhere in the US. So you can get in on the sports action no matter where you are, including these places.

9:09Make this your most rewarding season ever with Fandoo. Agent location restrictions apply. Product availability varies by market. CFandoo.com for eligibility. Gambling problem called 1-800-Gemmar. It's football season and you can now get almost anything you need for game day delivered with Uber Eats. What do we mean by almost? You can't get a running back delivered. But you can get baby back ribs delivered. A strong defense? No. A strong deodorant? Yes. The 6 pack of abs? No. 6 pack of beer? Yes. Get almost anything for game day delivered with Uber Eats. Official on demand food delivery partner of the NFL. Order now. Alcohol 21 plus in Inselect City's product availability varies by region. See app for details. Propel Fitness Water. With Gatorade Electrolites, Zero Sugar, and Vitamins. Propel hydrates better than water to help you get the most out of your workout and get back to your best self. What propels you? Propel with Gatorade Electrolites. You are talking about the plumbing of the digital economy.

10:10Like, what an AI buys just to do its own job. Exactly. If you ask a research agent to compile a comprehensive report on some niche industry and it runs for 20 minutes in the background, it cannot just rely on free search results. It hits paywalls. Right. It might need to pay for access to a specific proprietary data set. It might need to make a handful of API calls to other software tools to analyze data. It might need to bypass a paywall to read a few heavily researched articles. Each of those individual digital items is worth fractions of a cent. Yes. So maybe accessing the data set costs two tenths of a cent. The API call costs 0.01 cents. A virtual credit card designed with a spending cap and a list of approved retail vendors was not built for hundreds of rapid-fire micro transactions. No. The math simply does not work on traditional payment rails. The current financial system is based on authorization. There is one card transaction per event. If you go to the store and buy coffee, the merchant pings the network,

11:11the network checks your balance, authorizes the charge, and takes a fee. But if an agent tries to clear a charge for two tenths of a cent. The processing fees alone make it economically impossible. The network might charge a flat ten cent fee plus a percentage for every authorization. You cannot pay a ten cent fee to process a fraction of a cent transaction. Exactly. So the machine side of the economy needs aggregation, not individual authorization. Aggregation. That is the only way the invisible economy functions. You would need a system where thousands of tiny cent fraction amounts are collected in the background as the AI browsers and works. And then those micro amounts would be settled at the end of the day, or I guess the end of the week, as one single lump sum transaction on your actual bank statement. Because the AI cannot pause to ask for authorization every time it hits a paywall or pings an API. The friction would render the agent entirely useless. Imagine telling an assistant to do research, and they text you 300 times an hour asking for a penny to read a new paragraph.

12:12You would fire the assistant. Instantly. The payment networks have to build an entirely new settlement layer, designed specifically for non-human traffic. We are basically talking about shadow ledgers. Shadow ledgers. Yeah, the network tracks the machine's activity, tallies up all the microcharges, and then executes a real financial transaction only when it reaches a viable threshold, like five dollars. When you mention an AI paying fractions of a cent to read paywalled articles, it highlights a collision happening right now between two very different industries. Yeah, the contrast is crazy. In the exact same week, the financial payments industry ships a token that carries an AI agent's intent to buy things. The content industry is furiously locking AI crawlers based on guesses about what they want. Exactly. The media and entertainment worlds are fighting a literal war to keep AI out. You look at the major record labels suing AI companies over what they are calling an AI slop pipeline. Yeah. You look at California passing laws mandating the disclosure of AI performers.

13:14They view machine access to human creation as theft, or at least unauthorized scraping. They are passing legislation and filing lawsuits just to build walls around their intellectual property. Meanwhile, the financial world is calmly building a toll booth to let AI in. Wall Street is looking at autonomous agents and seeing a massive new customer base. It is fascinating. One industry's existential threat is another industry's growth strategy. The media industry looks at an autonomous agent reading its content and sees a parasite. But the financial industry views that exact same action as a billable event. If the agent has a virtual card, it can simply pay the toll. That assumes the media industry even wants to sell individual articles to machines for fractions of a cent. Right. They might prefer the subscription model, which forces the human to pay $50 a year for access to everything, regardless of how much they actually read. But if the financial world successfully builds this toll booth, the economic incentive for media companies might shift entirely. How so? Well, if a single human buys a subscription, that is $50.

14:17But if a million AI agents pay a tenth of a cent every single day to access specific data points for their users, the whole revenue model changes. You are no longer monetizing human attention. You are monetizing machine queries. Right. And the friction comes from the fundamental difference in how the two industries view content itself. The tech and finance sectors view an article or a song as a piece of data. It has utility value, and you can attach a micro price to that utility. While the media sector views it as human expression, which has intrinsic value that should not just be carved up and fed into an algorithm for pennies. The California Mandate on disclosing AI performers suggests the objection is at least partly ideological, too. Yeah. There is a desire to protect human authenticity in media. People want to know if they are listening to a human singer or a synthesized voice. But in commerce, authenticity matters way less than efficiency. True. A consumer does not care if a human or a machine negotiated a better price on their electricity bill.

15:17They just want the lower bill. Yeah. If the AI can read 10,000 pages of regulatory code to find a tax loop, you do not care that a machine did the reading instead of a human accountant. Which brings us back to the toll booth. Yeah. The publishers have to decide if they're willing to accept machine traffic, if it actually comes with paycheck. A publisher might refuse to let an AI scrape their archives for free to train a foundational model. Because training models is one thing. But if an individual user's personal agent arrives at the website with a virtual card, offering to pay two cents to read a single article to answer the user's specific question. The publisher has to make a choice. Rejecting the agent means rejecting direct and immediate revenue. It really forces the media industry to decide if their objection to AI is strictly financial or deeply ideological. If it is purely financial, the micro transaction toll booth solves their problem. The AI pays, the publisher profits.

16:17But if they view AI as fundamentally detrimental to human creation, then no amount of micro payments will resolve the tension. They will just keep building walls even if the machines are throwing pennies over them. Using the financial world successfully builds this infrastructure and the agents are out there actually spending money, what happens when the autonomous driver crashes the car? Customer protections and fraud risks are the immediate concern there. There are already warnings being published about this exact issue. Demos of AI agents are always perfect. The executive stands on stage, types of prompt, and the agent flawlessly finds the item, checks the price, and completes the checkout. But real-world rollouts mean real-world exception handling. I wish the boundaries of accountability on that for a second. If an AI buys a non-refundable item because it misreads your prompt, who's fault is it? Let's run a specific scenario. You tell your agent to book a flight to Paris, Texas, for a family reunion. The agent misinterprets the context, goes online, and books a non-refundable first-class ticket to Paris, France.

17:20Right. So who is liable for that mistake? Is it your fault for giving a prompt that lacks sufficient geographical clarity? Is it your company's fault for building an AI that hallucinated or failed to ask a clarifying question? Or is it the payment networks fault for clearing a transaction that you obviously did not want? We basically have to completely redefine fraud in this new era. Because in traditional finance, fraud is very straightforward. Someone stole your card or stole your number and bought something you did not authorize. It is a crime committed by a third party. But in agent commerce, no one stole your identity. You authorize the entity. You gave the software the virtual card. You gave it the prompt. Yeah, but the entity did something you did not actually want it to do. It is like giving an employee the company credit card and they accidentally buy the wrong office supplies. That is not fraud. That is incompetence. Traditional credit card dispute mechanisms are not built to arbitrate misunderstandings between a human and their software.

18:20If you call your bank and try to issue a charge back for the Paris France ticket, what is the bank going to say? The payment networks will almost certainly unear that the transaction was entirely legitimate. The agent presented the verifiable intent token. The purchase was within the total spend cap you said. He was in an allowed category, travel. Right. As far as the network is concerned, the system worked perfectly. The dispute is between you and the AI developer. And the airline is also going to refuse the refund. They sold a ticket in good faith to a verified buyer. They held up their end of the transaction. They are not going to absorb the cost because your digital assistant needs better geography training. That pushes a massive amount of liability onto the software companies building these agents. If Alchemy or whoever builds the interface is on the hook for every non-refundable mistake their models make in the real world, the financial risk of deploying an agent becomes astronomical. They will undoubtedly bury binding arbitration clauses and liability waivers in the terms of service.

19:22You will have to agree that any action taken by the agent is legally your action. If it buys a thousand dollars worth of decorative gourds because of a processing error, you own the gourds. You will waive your right to sue the software company for the cost of the goods. But that creates a massive chilling effect on adoption. How so? If consumers feel they have zero recourse when the machine makes an error, they will restrict the agents to entirely consequence free tasks. They will let the AI research the product, find the best price, put it in the digital cart, but they will insist on clicking the final purchase button themselves. But that defeats the entire purpose of autonomous commerce. The value proposition of an agent is that you do not have to be involved in the execution. If you have to supervise every transaction, it is just a slightly more advanced search engine, you are still doing the work. To get past that chilling effect, the software companies might have to offer internal insurance or guarantees. Like, they might have to absorb the cost of erroneous purchases just to build trust with the consumer base, at least in the early stages.

20:27If our agent buys the wrong thing, we will refund you directly. Right, but absorbing those costs would require massive capital reserves. It would mean only the largest tech companies could afford to deploy fully autonomous agents. A small startup cannot afford to cover the cost of a thousand misordered international airline tickets. If their model has a bad day, and hallucinates a sale. This kind of circles back to why human psychology will be the slowest part of this adoption curve. A user might accept the risk of a liability waiver for a $10 weekly grocery order. If the AI buys the wrong brand a cereal, you are out $4. You can live with that. But you will not accept that risk for booking international travel, buying electronics or signing contracts. The technical separation we discussed earlier, the leash, is really the only way to mitigate that fear right now. The kill switch and the network-inforced limits are there to ensure that even if the AI completely loses its mind, the financial damage is contained to a predetermined amount. You only give the agent $100 limit.

21:29It creates a really interesting dynamic where the payment networks are effectively acting as the final layer of governance for artificial intelligence. We are relying on traditional financial institutions, literally the companies that issue plastic cards, to keep large language models in check through strict credit limits. The transaction data itself becomes incredibly valuable in this ecosystem too. If an AI is constantly pinging APIs, paying for microservices and executing retail purchases, the entity processing those transactions has a complete map of machine behavior. They will know how algorithms make decisions better than the people who actually coded the algorithms. If MasterCard is processing millions of agent tokens every hour, they can see exactly which models are buying what, how fast they make decisions and what their optimization paths look like. Think about how that alters the whole concept of the consumer economy. We measure economic health by consumer spending, consumer confidence and human purchasing habits. But if a significant percentage of economic activity shifts to autonomous agents executing micro transactions and optimizing retail purchases in the background, the metrics we use to understand the economy will just warp.

22:40FanDool is now available everywhere in the US. So you can get in on the sports action no matter where you are, including these places. Make this your most rewarding season ever with FanDool. Agent location restrictions apply, product availability varies by market, see FanDool.com for eligibility, gambling problem called 1-800-Gemma. It's football season and you can now get almost anything you need for game day delivered with Uber Eats. What do we mean by almost? You can't get a running back delivered, but you can get baby back ribs delivered. A strong defense? No, a strong deodorant? Yes. The six pack of abs? No, six pack of beer? Yes. Get almost almost anything for game day delivered with Uber Eats. Official on demand food delivery partner of the NFL order now. Alcohol 21 plus in Inselect City's product availability varies by agency app for details. Propel Fitness Water with Gatorade Electrolites, Zero Sugar and Vitamins.

23:44Propel hydrates better than water to help you get the most out of your workout and get back to your best self. What propels you? Propel with Gatorade Electrolites. You could have a scenario where an AI agent detects a price drop in a specific commodity, let's say a specific brand of battery, and autonomously executes thousands of purchases on behalf of its human users before a single human being even registers the market change. That already happens in high-frequency trading on Wall Street. Algorithms detect a microfluetuation in a stock price and execute trades in milliseconds. What is unique here is applying that autonomous execution to the retail and personal finance space. It is high-frequency trading applied to buying running shoes, paying utility bills, and accessing digital content. The infrastructure required to support that volume of activity is fundamentally different from what exists today. We talked about aggregation versus authorization. Building a settlement layer that can handle billions of microtransactions daily requires a complete overhaul of how ledger systems operate.

24:46The intent token is the crucial piece of that puzzle. If an agent is executing hundreds of tasks a minute, the network has to be able to instantly verify that task number 72 was authorized by the user's overarching parameters. The token acts as a digital passport for the machine's action. It carries your permission signature along with the routing data. But that digital passport has to be standardized. Right. If MasterCard has one format for verifiable intent and Visa has another, and American Express has a third, the software companies building the agents will face a nightmare of integration. They cannot build a different checkout protocol for every single card network. The financial industry will have to agree on a universal protocol for machine-to-machine commerce. They manage to do it with EMV chips for physical credit cards. When you insert your card into a terminal, it works regardless of the bank. They will likely form a consortium to standardize egenic tokens. The financial incentive to capture this transaction volume is simply too high for them to let fragmentation ruin the market.

25:47If they figure out the protocol, they capture a percentage of an entirely new invisible economy. The liability question is still going to generate some very complex case law over the next few years. Yeah, if an agent operates strictly within its spend cap and purchases an item from an approved category, but the human claims the agent misinterpreted the nuance of the prompt, the merchant is caught squarely in the middle. The merchant shipped the product in good faith. The payment network cleared the transaction based on a valid token. The software company points to the binding arbitration clause in their terms of service. The consumers, the only one left holding the bag. And if the consumers left holding the bag too many times, the trust completely evaporates. We are officially crossing the threshold where AI goes from being a fast researcher to an independent economic actor with its own wallet. If trust has to be hard-coded into every machine decision, how long will it take before we realize we trust the money? If we realize we trust the machine spending habits more than our own. If you're not subscribed yet, take a second and hit follow on whatever app you're using.

26:51It helps us keep making this. We appreciate you being here. Also, check out our YouTube channel for more business and tech updates. There's a link in the description.