Who's Liable When AI Agents Make Payments? | Abril Benhumea

Alex: Hi everybody, and welcome to the Curiosity Code podcast. I'm your host, Alex Khomyakov, and my guest today is Abril Benhumea, a fintech and payments executive and co-founder of Passport, where she helped build open banking and real-time payments infrastructure connecting financial institutions and fintechs across the Americas. Her career spans payments infrastructure, banking as a service, cross-border payments, and financial partnerships, including roles at Synapse and Veem. Today, her focus is the infrastructure behind agentic commerce. What has to happen between an AI agent deciding to transact and money actually moving?

Alex: Identity, delegated authority, risk, orchestration, settlement, reconciliation and accountability. So we're gonna be talking all about it. Welcome to the show, Abril.

Abril: Thank you, Alex. And thank you for sharing this time with your audience. And being able to talk a little bit about agentic commerce Banking as a Service, Open Finance, and what is behind all these agentic payments.

Alex: Let's start with Passport. You describe it as the open banking layer rather than banking as a service middleware with a payment switch that institutions connect through. What does that let a bank to do that it could not do on its own?

Abril: So, open banking for Passport is a layer built on top of the banking infrastructure that currently the bank has in order to enable banking capabilities through an infrastructure provider that is Passport to offer products like accounts, payments cards, lending products. So, at Passport, we were not trying to replace the banking infrastructure. We were actually working alongside banks to enable these payments and to allow fintechs to launch financial services without having a banking license. So I think Passport is less a banking as a service and more an interoperability and orchestration layer of payments and accounts and the financial infrastructure that we were building within the company. So, if you think about why is this relevant for agents, it's because agents will also need a standardized way to interact in a fragmented financial ecosystem.

Alex: Okay, we'll get to that point in a bit what I'm mostly fascinated about in this product that you've built that it covers the whole Americas from Canada to Latin America So we don't have many guests here in the podcast that has the exposure to the whole Latin and Americas region So, my question for you is, where are the real differences between those markets and how money actually moves?

Abril: There are some similarities and some differences, and it depends if the ecosystem is built on top of real-time payments or a batch payments. The US traditionally moves in batch payments, and Colombia, for example, is one of the markets moving on top of batch payments until very recently that they enabled Bre-B, which is the real-time payments infrastructure that Passport is powering right now. But that's mainly like the difference between the different markets. Like looking who's doing the settlement in real time and who's not. The different countries are trying to catch up in this real time infrastructure, but payment processing, card networks card infrastructure is kind of about the same.

Abril: Stablecoins infrastructure is also kind of about the same thing within the different markets.

Alex: Would that impact the way you build your product, your offer, your services? How do you adjust your offering in such a vast majority, vast differences in terms of solutions.

Abril: It depends who you are, right? If you are a financial institution meaning a bank, licensed bank, if you have an EMI license equivalent like in Europe that you are a payment network, if you are a fintech, if you are a merchant trying to offer financial services, the main difference between a real-time payments ecosystem and a non-real-time or asynchronous real-time, asynchronous payment ecosystem like ACH is about reconciliation and liquidation. Because liquidation and reconciliation sometimes is not happening at the same time. But even in real-time payments ecosystem, there are different banks where reconciliation and liquidation is not happening at the same time. In Mexico, for example, liquidation is not happening until the end of the day.

Abril: In other RTP systems such such as Brazil and Colombia, the liquidation is happening in real time. And that brings up other different considerations for a fraud prevention and refunds, right? But I'm not sure how deep you want to go into that because we can have a conversation of just one hour about it.

Alex: No I think just a high-level overview is enough just to get a grasp of differences in in the countries. But let's talk about back to your point about AI agents. said something on our intro call that I think is like it could be the whole episode. So decisioning will sit inside the agent, but the trust layer and the orchestration layer will not. Why not?

Abril: I have passed all my career building the layer of execution and the layer of intent. And that's common in frameworks such as open banking, open finance, and banking as a service. So this layer that is going to be used underneath the agentic commerce is not a new, it's not anything new. It has been already there for quite a while. What hasn't been there is the layer of trust.

Abril: And by the layer of trust, we mean the agent is going to get a mandate of execution and of executing transaction or any decision that you may want them to execute. The issue or what is missing, the missing piece is right now a couple of questions. Is the agent actually authorized to make a payment on my behalf or execute a decisioning on my behalf. The agent is actually representing Abril or Alex, or is not? And that's the layer of trust that it's missing still.

Abril: We can talk about like the different protocols and what is going on right now and how the different companies are trying to solve. But that buying intent, the layer of trust, like the agentic part of it is the one that it's still not defined. How the bank knows that this agent is actually representing Abril, and Abril is allowing the agent to execute that transaction to buy a ticket. Or the agent has the opportunity just to make one purchase, three purchases per day but now having like 300 purchases per day. How the bank knows that this intent and this layer of trust can be enabled for the agent to execute the transaction.

Alex: Right, and is it connected to knowing your agent? Because you mentioned the sequence, know your agent, then know your customer then know your business, and then execution. So I'm wondering if the trust piece is actually connected to this knowing agent, and if so, what does it actually mean to know the agent?

Abril: Okay, so financial services is a fragmented industry that you need to put together a bunch of different pieces in order to execute the payment flow. So what we are adding to the regular payment flow right now is the know your agent or KYA. And this part of KYA is a new part of the layer of infrastructure that we are enabling. As I told you, we need to have the layer of trust, the layer of intent and the layer of orchestration of the payments. So for this layer of knowing your agent, there are a couple of things that we need to consider that it's the MCP server connectivity that it means that the protocol that we are connecting through is a secure protocol and it's not trying an intent of fraud to execute transactions without having been authorized and that it's a secure channel to execute the transaction.

Abril: And the second that you are saying and mentioning is that knowing your agent and it's this agent is in capacity to execute these transactions or not on behalf of Abril. So it's a matter of connectivity and a matter of a of mandate that it's representing the mandate that Abril is giving to the AI to execute the orders that I want them to execute.

Alex: How much authority you can actually give to that agent and how is that been passed to a bank so that the bank can validate it and I'm also wondering how would the revoke process look in this case

Abril: So that's a layer that at this moment everyone is trying to build, that's a missing piece that everyone is trying to sort out. It's not solved yet. There are a couple of solutions in the market trying to tackle this issue, but not necessarily like solving a 100%, I will say. It's in the path for enabling the agentic commerce, but this is the part that it's the one that it's still trying to be sorted out by different companies. You see Base Layer you see other companies, Stripe, even Anthropic, trying to solve it, enabling these protocols or these tools for the agents to be able to execute payments.

Abril: But that's still on the way of trying to solve who's the agent, what is the agent mandate, and how the banks receive and verify this agent's intent and the mandate that Abril is giving to the agent.

Alex: At what layer are they trying to solve that? Is there already like well established protocol that could be leveraged there or are they just going all the way down to finding the protocol, signing off on that, and then building solution on top of it.

Abril: So currently in the market, there are different protocols. They are open source protocols, for example, that they are promoted by Google. There are other protocols that are exposed by OpenAI even from Anthropic. So there's no one protocol that right now is the standard like for what it's gonna be happening in the agentic commerce. And we can make a full episode just talking about payments protocols and agentic protocols for this connectivity.

Abril: But I don't think that's the matter that we should be discussing here. It could be opening up also an opportunity for fintechs and for companies to enable agentic commerce in a layer of trust that you are building this connectivity to the different protocols, or if we see that one protocol is the one that is winning over the others and it end up being open source, probably you just need to connect to one. But that's still evolving day by day, and we are seeing that new ideas are coming up. And that's what Stripe, Adyen, Airwallex, and other different vendors are trying to solve right now.

Alex: Got it, makes sense. Let's talk about the mechanics then, more where Passport is involved, like before a payment goes out what has to be checked that would not be checked for a human if we're talking about again agents conducting payments and where do those controls sit in the agent or in your layer or at the bank

Abril: There are a couple of layers to unpack here. We mentioned so far who is the agent, who is acting for what the authority was delegated to it and if the authority is still valid, what are the limits and those specific transactions fall into those limits. There's a separate question on orchestration and once that the transaction is authorized, how should we execute it? Which provider? Which payment?

Abril: Which rail? The agent can help us out here by enabling the routing logic based on real-time decisioning. But the infrastructure, I will take it into three different setups, I will say, so far. One is the decisioning, the other one is the trust, and the last one is the execution. The one that we are seeing right now that it's the execution side is a part that it's resolved with banking as a service, open finance, open banking.

Abril: So that layer of execution is already over there. The layer of trust is the one that we have discussed so far, identifying the agent and the mandate and if it has the valid permissions. But the decision is coming from the LLM models that are enabled for making these decisions based on what you are configuring the different agents and what the models are deciding to. For example, you are authorizing your agent to make a purchase of a ticket to Mexico. And you are giving them a budget of $1,000 And in that spectrum, what it's going to do is probably make go out, look for a ticket and spend $700 on that flight to Mexico.

Abril: But there also one missing piece that we haven't discussed so far is what happens if the agent is buying this ticket in a wrong date? Who's going to be responsible for that? That's the decisioning piece. And that's also one of the missing piece that it's still building right now, OpenAI, different companies that it's going to be responsible if the agent makes a mistake. And as more decisioning we are adding, I think that more controls in the trust and execution layer should be enabled to prevent these accidents.

Abril: I will call them.

Alex: So like there could be so many different ways this could go wrong, right? So it could be back to your example about ticket to Mexico. It could be a wrong date. It could be a wrong person on the tickets. It could be completely wrong party where the money have been transferred.

Alex: So it could be, I don't know, a ticket not issued by an air carrier, but a train ticket whatever so You're saying there is no yet solution that would govern the way this money will be returned back to the agent or to a person who is deploying this agent

Abril: I won't affirm that there is not a solution because, for example, Alibaba with Antom is already a solution in the market working towards solving these different value chain. Stripe is working on this enabling agentic payments too, and Adyen and others are working. So I haven't tested out that much like to make an opinion of this is sorted out. But what I want to say is that authority is the part of the value chain that should be constrained. Abril authorized an agent, and therefore the agent can do anything Abril would do.

Abril: But that doesn't mean that they can take decisions so out of their mandate or the authority. So it shouldn't be binary, but we need to enable constraints for the agent mandate. And before moving the money, the financial system needs to verify that the required transaction falls within the mandate that Abril is giving to the agent. So that delegated authority needs to be verified and we need to remember that financial systems within the US or any different country or even worldwide, it's fragmented. So it's not going to be just one solution that fits everything and all the industries and all the different markets because we still need to build local payment rails and local infrastructure in each of the different markets that we are operating.

Abril: That's why, for example, stablecoins has become one of the payment methods that we are going to be using, probably using tokens to enable agentic payments for cross-border flows or even for local flows. But still, in an ecosystem that is fragmented, it can work in many different ways. You can issue a stablecoin, you can mint a token, you can provide a debit or prepaid card to an agent to execute the transaction. There are many different ways of solving these infrastructure capabilities for agentic commerce.

Alex: Yeah, that's actually one of my next questions, sort of. So when we talk about the whole payments infrastructure, reconciliation, settlements where volume usually breaks things. And when we think about agentic payments we imply that it's going to be very fast and at scale, at least eventually. Have you thought about what breaks first when the payers are agents rather than people.

Abril: What breaks first? There are many things that they can break at the same time. As we said, there are three different flows that we need to solve out. The decisioning, the trust and the execution. Within the decisioning it might be that the agent is making a poor choice when selecting the ticket and deciding a different date.

Abril: In the trust, it can be that the connectivity and the mandate is failing And in the execution It depends how you think about execution. Because if you have an agent, the same agent can route the payments through different payment rails, or you can decide to have already preset rules for the execution of payments, which has been and how traditionally it has been worked so far in financial services. But now with AI, you can route for the most effective for speed, for pricing, for time. And that's a part where we can use the agents and the LLM models to decide how to route these payments. What will break first?

Abril: I'm not sure.

Alex: We're yet to see.

Abril: I will have to see. There are different capabilities. and different problems to tackle within these different layers, the decisioning, the trust and execution. What are your thoughts about it? Have you ever thought about it?

Alex: I'm looking at where we are today and I think it's even too early to talk about it even, you know, this conversation is like we were trying to figure out where we are at. And it seems like there are multiple approaches right now from different big players and they're still trying to figure out it. So I'm an observer. I haven't thought about it. And talking about that.

Alex: So you mentioned, you know, Google working on their own agentic payments kind of protocol, open protocol and there are also model providers who are doing something on their side. I'm curious if that competition matters to the bank at all or it's just a plumbing that bank doesn't care much about. They will never use it. It's more on higher hierarchy of the transaction execution.

Abril: I think agentic commerce is not going to be a protocol problem. You have a discovery phase that it's how the agent is finding the product or the service. Then you have an intent and how does the agent communicates what it wants to do. The identity and the delegated authority, that is how we establish who the agent represents and it's allowed to do it. And then you have the payment credentials, the execution, the reconciliation and the liability.

Abril: So I don't expect that one protocol will own the whole stack. we end up with multiple agent ecosystems, merchant standards, payment providers and financial rails. Then something will have to translate this into agent intent and financial execution. And that's where is probably the orchestration opportunity and the orchestration becomes extremely important. That's what we are seeing companies such as Stripe trying to solve.

Abril: Stripe is working right now with OpenAI trying to enable merchant payments and they are working on something that it's really interesting right now, that it's the positioning of the catalog that it's servicing the first part of how the agent is finding the product. And I think that once that the agent has the full catalog in order to make these recommendations of what to buy is when the execution of the transaction is going to come true. But there are two different problems that they are trying to solve right now. Alongside you can see from the news what Shopify how Ulta Beauty is powering their catalog with OpenAI. I'm not trying to make any promotions here to any of the companies, but those are the most public ones that I can talk about it like so far.

Alex: I'd like to wrap this episode up by looking out at where are we heading with all of this because you know like we've been talking about the ways technology can support it today, things that are still being developed. I personally work with LLMs and AI on a daily basis and I have delegated a lot of things that I couldn't have imagined even a few months ago that LLMs will be capable doing that. So I wouldn't be surprised that agentic payments will actually be part of our lives like in a year probably less than a year, and that will be just one of the things that we'll be using day-to-day and won't as a consumer won't even think about how they actually works. Do you have any idea how big the market is for this and where we're heading with all of it. Just trying to really kind of paint the big picture in terms of where this technology is heading.

Abril: That's a good question. I am not sure because scaling this market will require that different customers get on board it. So probably it's going to be adopted early for the new generations, for Gen Z and everyone that is like digital native but it might take some times to be adopted by my mom, for example, although my mom is making decisions for purchases of a car, based on what ChatGPT is suggesting right now. So I will not be surprised that she will be also buying and executing these payments.

Alex: Yeah, that's what I'm saying. Behavioral changes are happening so fast. It's too fast sometimes.

Abril: But I think it's going to be defined about like scalability and autonomy and also before execution, how you can program all these transactions. And if you have the controls and detecting the problems like fast, because you have merchant restrictions, limits, control. So sometimes the agent is going to follow its mandate perfectly, but the outcome will be still wrong, right? That's what we were discussing within the questions that we thought that we will be relevant for this episode. So I am not sure how big the market is, but I know that we are at 1% right now of payments processed through agentic commerce.

Abril: And from data that I have here from other conferences the expectation is that this grows to 30% by 2030. So that's a huge opportunity right now for the different companies. going to be also merchant adoption. How do you position your company for recommendation? So not just the agent is referring your product.

Abril: I am seeing that happening with beauty because beauty is always ahead of any other trends. But for purchases that are above $20, you might not be right now deciding to buy a car, right? Within OpenAI or within the Anthropic interfaces. So maybe for low value transactions, and then later, we are going to see the high value transactions. But so I am looking forward and excited about what is coming in the next coming months.

Abril: Because as you said, this is changing so fast.

Alex: Yeah, yeah, yeah, listen, the way I look personally at this is more scenarios, like a personal assistant scenario where you would say hey buy me a ticket for tomorrow afternoon to that location or to that event or buy flowers and ship them tomorrow to a person X. I think that is kind of where the most adoption will happen. I don't think agents are going to be purchasing cars or houses or you know high value transactions. I think it's those kind of roughly a couple hundred dollars transactions. I think that's where the volume will be.

Alex: At least that's my feeling. That's how I would use this technology. The way I'm doing that with my a personal assistant at the moment who is just a human being same thing. You have certain guidelines, you have certain way to put your task and certain acceptance criteria. It's just gonna be replaced by AI agents, I think.

Abril: I mean, we are seeing it so far within reservations hotel reservations, right? The stock market punished Airbnb recently on their stock based on what we are seeing that now you're using OpenAI or Anthropic to book a hotel directly, and you are not going through the payment flow of Airbnb, which is the one that they are using to make money. So Airbnb is not making money on the reservations by just booking, it's making money by processing the payment and holding the balances for the hotels or the renters. So if they are starting to lose their revenue, we are seeing that OpenAI and the other different LLMs are winning markets that we didn't expect to get this fast in just one year.

Alex: Yeah, yeah, and that happens across the board across many industries. We've spoken about it in different episodes, about different industries, about, you know, development, software development services you know, you name it, it's happening across the board. So it's not a surprise.

Abril: But the question that we didn't make is who's going to profit about it? It's going to be Stripe. It's going to be the layer of trust. Is the decisioning the one that it's getting the most of the revenue out of this new ecosystem? That question is still up in there to be solved within a couple more episodes or even waiting some time, because I'm not sure who's going to be winning right now.

Abril: There are a couple of different perspectives that just embedding the decisioning and the layer of trust within Stripe is going to be easy because they already have the payment flow or you are gonna build a new category on top of it.

Alex: Well, that's exactly how I want to finish this episode where we'll keep looking for answers. I'll keep watching what's happening and Abril thank you very much for coming over to the show and I enjoy the conversation a lot.

Abril: Thank you, I enjoyed it too, looking forward.

Alex: And for the listeners don't forget to hit the like button on YouTube, subscribe to the channel and leave a review at the platform you're listening this to and see you in the next episodes. Bye bye. Bye.