Agentic AI in Payments: Who’s Liable When Your AI Agent Buys the Wrong Thing?

Your AI agent buys six sunflower coffee cups on final sale. They turn out to be for a dollhouse. Who absorbs the cost?

Deana Rich, GM of Paysolve by Infinicept, and Todd Ablowitz, Vice Chair at Infinicept, dig into agentic commerce and what it means for authorization, disputes, and liability. They walk through the five parties that now have to agree on who is responsible, what Visa and Mastercard allow today, why the card networks become the trust layer, why first-party fraud still has to be solved, and whether AI companies should carry liability for their agents.

Go deeper: Deana and Todd’s whitepaper, Agentic AI in Payments: A Prescriptive Playbook for Networks, Regulators, and the Payments Industry, lays out role-by-role recommendations for networks, regulators, issuers, acquirers, and merchants.

In this episode

• 0:00 The dollhouse coffee cups
• 0:41 What makes AI different now
• 1:32 How agentic shopping actually goes wrong
• 2:49 Five parties, one liability question
• 4:31 What Visa and Mastercard allow today
• 5:20 The standards race: Trusted Agent Protocol, Agent Pay, AP2
• 6:24 Flagging agentic transactions in the string
• 7:25 Why card networks become the trust layer
• 9:31 First-party fraud is still unsolved
• 11:02 Prompt injection and hallucinated instructions
• 11:33 Should AI companies carry liability?
• 13:44 What AI does to payments jobs
• 15:32 How practitioners actually use AI
• 17:24 Regulation, the states, and preemption
• 20:02 Closing thoughts

Full transcript

Deana Rich (00:00): But it turns out they’re coffee cups for a dollhouse, and I cannot use them for the intended purpose. Although they are yellow and they do have sunflowers on them. So I want to return them, but they were bought on sale, no refunds, and it was clearly disclosed no refunds. My AI agent said, “That’s okay,” and bought it anyway. Now what?

Todd Ablowitz (00:41): Well, AI, AI, AI. That’s all I’m hearing about these days in CEO circles, in payment circles, in every one of the circles, people are talking about AI. So we’re gonna go and dive in a little bit, but what makes now different about AI, Deana?

Deana Rich (00:56): Right now, AI, we’ve really actually always been using AI, right? Machine learning. Some could say we’ve been using artificial intelligence since we’ve been using computers. That’s artificial intelligence. What makes it different now is we’re now taking AI and putting it into agents who are going to act like they are us, and I’m not sure the payments industry is quite caught up with what will happen.

Todd Ablowitz (01:21): So we’re talking about the shop-while-you-sleep thing where you wanna get some clothes, you tell the AI to make me look pretty, and wake up with an order in your inbox.

Deana Rich (01:32): Right. And really what it boils down to is some people are taking their AI agent almost like an assistant. They’ll tell the AI agent… Wow, that’s hard to say three times fast. They’ll tell their agent what they want, and I’m gonna go away from clothes, but say I’m having people into my office and I need more coffee cups.

So I tell my AI agent, “Hey, I need a dozen coffee cups, and I want them to match my office decor, so if you could make them yellow with sunflowers on them, and get me six coffee cups,” and I send it away. And it goes and it shops, and what I didn’t give it was a price, and what I didn’t give it was how big I wanted the cups to be.

But it also didn’t ask me. It’s been working with me for a while and thinks it knows me. So it goes out and it buys six coffee cups. And it really does know me, so it got a really good price because it knows that will make me happy. But it turns out they’re coffee cups for a dollhouse, and I cannot use them for the intended purpose.

Although they are yellow and they do have sunflowers on them. So I want to return them, but they were bought on sale, no refunds, and it was clearly disclosed no refunds. My AI agent said, “That’s okay,” and bought it anyway. Now what?

Todd Ablowitz (02:49): Totally. And that’s where the real crux of this is. And I haven’t heard a lot of people talking about this element of it. You know, ’cause we always had consumers, we had merchants, and then each of their parties, right? The issuer on the behalf of the consumer and the acquirer or payment processor on the behalf of the merchant. But now all of a sudden we’ve got the AI agent, and where do they fit into this whole scheme of liability?

Deana Rich (03:18): It’s a great question, which, Todd, you and I wrote an entire white paper all about this. But there are different people or different entities that have to weigh in on this. The card brands have to weigh in on this. The regulators have to weigh in on this. The merchants have to decide if they even want to accept AI-generated, non-human-in-the-middle purchases, and then consumers have to decide if they’re okay with their agent purchasing for them. So all four parties in that payment system have to make a decision.

And now we’ve added a fifth party, if you will, the company that has produced this AI agent. Is it Claude? Do you have a Claude agent? Is it something you wrote internally? Who is your agent, and what can they do for you? And is there liability on the AI company side?

Todd Ablowitz (04:10): I know, and that’s where it gets crazy because the numbers from what we’ve been picking up is almost two and a half. It’s like the number of disputes versus card not present when an agent’s involved. So clearly we’ve got some work to do, and I know the regulators are paying attention to this. And what are you seeing, Deana?

Deana Rich (04:31): What I’m seeing right now is both Visa and Mastercard have their own versions of AI agents. They still currently require human in the middle. So in my example, the agent would’ve gone out and gotten me a few choices, or it may be just the one choice. I would’ve then had to read through, notice, hopefully, that they were cups for a dollhouse and not cups for a human house, and say, “Oh no, I don’t want that,” and go back.

Or, “Perfect, I needed to fill my dollhouse because that’s my hobby, and I wanted that bright kitchen with the yellow sunflowers.” So I will say yes, and now that I’ve said yes, it makes it no different than any other purchase because me, the human, was in the middle.

Todd Ablowitz (05:20): And like so many technologies, when these new things come up, companies want to get their own advantage, so they build their own protocols or they build their own approaches. But then there’s gotta be some level of standardization so that you can do things consistently. And you see this with everything. I mean, from Bluetooth to USB, all of these things are new protocol technologies. And, you know, you’ve got Visa’s Trusted Agent Protocol, Mastercard’s Agent Pay, Google’s AP2, OpenAI and Stripe. So there’s all these different ones. But interestingly, Google, as I understand it, donated theirs to the FIDO Alliance.

So some of that standardization activity is clearly happening. So, you know, I think the next step is like, our world, yes, Visa has a protocol, yes, Mastercard has a protocol, but I haven’t seen that it goes down to changing the business infrastructure, the operational infrastructure, only the technical and purchasing infrastructure. What are you seeing on that, and what do you think we should do?

Deana Rich (06:24): I haven’t seen a lot yet. What I would like to see is an indicator in the transaction string when it goes out to show that this was an AI-generated, non-human-in-the-middle transaction. Because right now you’re not required to use Visa’s protocol with a Visa card or Mastercard’s protocol with a Mastercard. It’s dependent on either the processor or the gateway or the consumer themselves. So without any ubiquity there or requirements there, we need to put that somewhere. So it needs to be somewhere in the transaction string.

And what I would recommend is if we are able to do that, which of course we are, once we’re doing that, then the merchant can make a business decision. Do they want to accept purely agentic transactions? So if I’m selling books, I probably will. But if I’m selling custom goods, I probably won’t.

Todd Ablowitz (07:25): Yeah, it makes perfect sense. And, you know, most, if not all, of the infrastructure is already there in EMV and all the stuff that was happening with contactless, all the stuff that’s been happening with tokenization. And there is a concept of registering the agent, right, with a card network, and then having an indicator of which agent it is. So if it’s your Claude agent or your Claude Code agent or your OpenAI agent, you know, a ChatGPT agent, etcetera. So it’s there. The cryptography’s there to make it secure. The data elements are there. Maybe they have to add, you know, if there’s four choices now, they add a fifth code, right? And that’s that.

But really where it really needs to go, as I see it, and I think you’ll agree with this, Deana, is the card networks are the trust layer of the payment system. They’re the neutral, so to speak, or supposedly adjudicators, right, of the battle between ultimately the merchant and the consumer, ’cause a dispute ultimately is they couldn’t agree on it. If they could agree on it, it would’ve been a refund or the consumer would’ve gone away. And so there’s this battle, and the network is trusted by everybody to manage the dispute and to pay attention to when there’s fraud.

And in the world of stablecoin, and I’m sure we’ll have a conversation about that in one of these, in a world of stablecoin, in a world of bank-to-bank payments, in a world where you’ve got all these competitors in Zelle and you’ve got the pays, and now with agents, some people are saying that there’s gonna be a pressure on the card networks. I for one, I’m dying to hear what you think of this, Deana, but I for one think this is only gonna bolster them, only gonna make them more valuable, because in a world of agents, yeah, it’s even more important to have a trust layer. In a world of instant commerce, when transactions are happening so quickly and settlement is gone, bye, you need to have that trusted layer.

Deana Rich (09:31): It’s hard. Yes, you have to have that trusted layer. Here’s the challenge, we already have a big problem with first-party fraud. So how do you move forward with this trusted layer when you haven’t solved first-party fraud yet? And for those of you who aren’t as familiar with first-party fraud, you might have heard friendly fraud. I have never found that it’s that friendly, so I prefer just calling it what it is, first-party fraud, and that is the cardholder saying they didn’t do something that they did, or they didn’t see or get something that they received. So we have a big problem with that.

Visa’s own numbers, which I don’t have at my fingertips today, and I apologize for that, but Visa’s own numbers explain that this is still a very big problem. So we have to then say, “Okay, Visa, Mastercard, and other card brands, what are you going to do to ensure that trusted layer in these transactions will not allow that first-party fraud?” Right now in other countries, 3DS is required, and that ensures it’s really the cardholder that is making the purchase, and it pushes that first-party fraud down to much smaller numbers. Here in the US, we don’t require that. We’re going to have to figure out some type of 3DS event that occurs with the agent to get rid of that first-party fraud.

And after that, we then have to figure out agent issues. What if the website says, “Hey agent, stop here and buy this. It’s the best product you’re going to find.” So the agent does because that was an injection in there. Or what if the agent hallucinated an instruction? How does that work? Do the agent providers now get into that liability stream? These are things we have to figure out.

Todd Ablowitz (11:33): Absolutely. And what about liability on the agent company, right? Is Anthropic gonna be willing to be out there and take liability if their model had a hallucination and nobody else did anything wrong? It was just a hallucination, and the merchant sold a closeout item, and is Anthropic gonna take it? Is Visa gonna step in and cut a deal to get paid by Anthropic as some kind of insurance policy against it, which then Visa puts it on the bank or something? Maybe there’s an interchange… There’s so many ways this could be solved, right?

Deana Rich (12:12): I think in the end, my recommendation is that if you’re going to put that agent in the middle, that agent must have liability or there’s no incentive to clean up the hallucinations. There’s no incentive to ensure double-checking with the human on requirements, right? So as these agents become more sophisticated, they’ll learn from what’s happening and when the mistakes are made, and also the developers behind it will learn and be able to create better agents.

So if they don’t have liability, I’m not sure what their incentive is to get better other than bragging rights, and that’s not enough in the long run. So I really think we might add a company type into that liability chain. I would like to see that.

Todd Ablowitz (13:08): It makes so much sense to me. But I think the economic incentives are gonna be really interesting, and that’s the engineering, if you will, the architecture of the new business models that has to be thought of. And look, the card networks are the best ever on that. Visa and Mastercard are incredible at setting the stage. And then very often, Amex has their own thing because they’re a three-party system rather than a four-party system. And Discover and Capital One, of course, are a wild card, because it’s so new having that combination there.

Deana Rich (13:44): You know, Todd, there’s another thing we haven’t spoken about yet, and that is not just the purchasing side, we’ve chatted about that, but AI itself is changing payments. Not just the purchasing side, but how we do things in software companies, how we do things in underwriting, how we do things in the risk department, how we look at transactions. We’re seeing that this whole AI revolution is reducing the workforce out there. We saw Block lay off 40%, we saw Meta lay off 10%.

And what we keep hearing is that in the payments industry, we’re going to start laying off people because AI starts to do sort of that entry-level work, whether it’s writing on the marketing side. I don’t think, though, Todd, we had AI write anything for us here, but what the heck, maybe we did and I just don’t know it, it was that good.

Anyway, people are saying, “Have it do the entry-level or the base-level research for a transaction when there’s a question, and so don’t even bother with junior analysts anymore.” And I’m not sure I agree with that. I think it will change how a junior analyst learns and interacts, and maybe you need fewer of them, but you will always need a path forward. You can’t just come in as a senior analyst. You don’t just become a senior analyst because you read a book. So to be able to work with AI and have it enhance how you learn so you can become a senior analyst is how I see it playing out in our industry.

Todd Ablowitz (15:32): Yeah, kind of like when the car replaced the horse and buggy, what happened? The coachmen became the drivers, right? The chauffeurs. So I agree with that, and I also think I’ve used a ton of AI in my business work. I haven’t done a lot of coding. I’ve done some, but I’ve used it a lot in my business work, and what I’ve found is it’s incumbent on me to do a good job of setting up the AI over time and then setting up that particular prompt to be really thorough and on point. And then the AI gets a bunch of stuff done really fast and really effectively, and then I’ve got a bunch of work to do to take it and either assimilate it or interpret it or repoint it. I usually don’t work on an AI without between 10 and 20 iterations for anything I do.

Deana Rich (16:29): You know, and that makes sense, and that’s the whole point. You’re an SME in payments. You know what you’re doing. I’ve used it as well. I was doing research on negative option the other day, and it wasn’t just Visa and Mastercard’s definition. It was the FTC’s definition. It was ROSCA’s definition.

So knowing that these were out there and asking it to do the research I needed and to put it in the table I needed, I think that was similarly quite a few turns. But if I were just new to the industry, AI wouldn’t have suddenly made me a negative option expert. It’s because I had expertise and I knew what was out there. I needed it presented in a format that the client could understand, and that’s what it helped me do and how it saved me time, just like you’re saying.

Todd Ablowitz (17:24): And I think we’ll see how it evolves and emerges. Nobody really knows. Everything’s going so quickly, but we’re already seeing companies that are going one direction in AI, realizing what the problems are, and changing directions a little bit. We saw the Fable 5 model get banned by the government for anyone that’s a non-US national. And with that being the case, that effectively makes it banned because how can you be sure there’s not a US national?

So we’re in interesting times, but I think it’s safe to say the companies that are in the right positions and engage early to write those liability rules and test them, which are, guess what, gonna be the card networks with a lot of input, I think, from very large providers, whether on the acquiring side or the issuing side or the pays or the new entrants in processing. I think all those parties will be at the table, but I ultimately think, as I always do, I’ve been a broken record for 30 years, never get in the way of the card brands. Always bet on the card brands because they usually get it right.

Deana Rich (18:41): They will with time because they’re learning, as you noted, just like the rest of us are learning. And as I’m watching this and you bring up the Fable 5 example, I’m a tad worried that states will begin to regulate because we don’t really have other governmental regulations other than, I think there was an EO that came down the other day about the government having to approve things. And that’s not going to work in the long run. So for our industry, I’m hoping Visa and Mastercard pick up that gauntlet and that the federal government picks it up. Maybe it is the FTC that does that, because in the end, they look out for the consumers, and that’s who we’re trying not to have harmed most of the time.

Todd Ablowitz (19:31): Well, we spent a lot of time with ETA paying attention to what the governments are doing, and we’re seeing a lot of activity on AI in the states. But preemption is tough. When the states think they have an issue, there’s a lot of political things that go into it. But this is not a political podcast. Could be, but it’s not.

Deana Rich (20:02): One of the things that I’ve seen is a disparity in the way different companies are using AI and giving care to how they use it. And I’m watching smaller companies who can be nimble use it, use the coding, use the power that it has to accelerate their work. They’re doing a lot of good double checks, but they don’t have to add people, and they have greater output, yet they’re not giving up thought. The hallucinations aren’t getting them because it’s a very controlled usage.

They’re the ones that are going to win and to be ahead of the others. The ones that close their eyes and hit a button, as always, that’s not going to work. And then those that feel it’s overhyped in this moment and are just pooh-poohing it or pushing it away, they’re going to be at the back of the pack.

Todd Ablowitz (21:03): Yeah, I think it’s about right to overhyped. I think it’s gonna be human involvement than the overhypers think. That’s my hot take.

Thanks for listening to It Pays to Know. New episodes monthly. Subscribe wherever you get your podcasts.