AI Agents Have Brains, But Where Are Their Wallets?

Amias Gerety is a partner at QED Investors, a fintech-focused VC firm with $3.8 Billion under management. He was previously the Assistant Secretary at the US Department of Treasury. And he graduated from Harvard.

Amias's favorite book: The Origin of Species (Author: Charles Darwin)

(00:01) Introduction and Setting the Stage
(00:29) The Status Quo: AI Agents and Online Purchases
(02:20) The Current Payment Infrastructure Explained
(04:42) Automation Today: Auto-Complete and Recurring Payments
(06:09) Why AI Agents Need Dedicated Financial Infrastructure
(08:29) Historical Examples: Auctions, Bots, and Payment Systems
(09:32) Identifying and Verifying AI Agents in Commerce
(11:59) Risks of Agentic Commerce: Lessons from Knight Capital
(14:47) Stripe’s Role in Agent Payments: SDKs and Gaps
(19:46) Opportunities for Startups in Agentic Frameworks
(22:49) Challenges: Disputes, Chargebacks, and Reversals
(26:32) Regulation and Governance in Agentic Payments
(28:24) Building Merchant Trust in AI Transactions
(30:11) Needs of AI Developers: Reliability in Transactions
(33:04) The Future of Agentic Commerce and Microtransactions
(35:42) Building Consumer Trust in Agentic Systems
(38:45) Broader AI Applications in Finance
(40:13) Rapid Fire

--------
Where to find Amias Gerety:

LinkedIn: https://www.linkedin.com/in/amias-gerety/

--------
Where to find Prateek Joshi:

Newsletter: https://prateekjoshi.substack.com 
Website: https://prateekj.com 
LinkedIn: https://www.linkedin.com/in/prateek-joshi-91047b19 
Twitter: https://twitter.com/prateekvjoshi 

Transcript

Prateek Joshi (00:01.376)
Amias, thank you so much for joining me today.

Amias Gerety (00:04.354)
Thanks for having me. This is fun.

Prateek Joshi (00:06.668)
Yeah, this is a topic we have written about, we've discussed offline. So I'm really excited to dive deeper. So let's paint a picture of status quo for listeners who may not have read the piece. So what happens today when an AI agent tries to make a purchase online? Can you just walk through the current infra that is available?

Amias Gerety (00:29.506)
Yeah, so the first thing I think we have to note is that I don't think this is happening today. Right? So there's, there's a lot of discussion and I think a lot of excitement. think both of you, you and I are excited about the possibility of AI agents making purchases online. Right? When, when people talk about the idea of, Hey, I feel like kayak has their AI working against me. You know, can I have my AI work against them and be just at the perfect time? Right? We all feel that.

Prateek Joshi (00:34.946)
Right, right, right.

Amias Gerety (00:58.71)
But I think the truth is that today AI agents cannot make purchases or to the extent they are, it's only in very, very narrow examples. maybe the first thing I would say though is when we have AI agents making purchases today, probably what we're talking about is an organized fraud. We're probably talking about organized crime. We're talking about bots.

doing account takeover and trying to cram credentials into e-commerce. That's probably what we're talking about. And that, think, is exactly where we start here, Pratik, which is that the modalities of an AI agent acting on my behalf today look like an organized criminal gang acting with my stolen credit card information. So that's where we have to start.

Prateek Joshi (01:48.163)
Haha.

Prateek Joshi (01:51.692)
That is great. Right. That's one hell of a way to start. That's amazing. Okay. So maybe we can take a step back and just talk about, given a human, when I go to a website, I like something, I want to put in my card, I want to order it. Can you talk to, in plain English, like what happened? How does a transaction happen today? And then we can dive into the need for, why do agents need their own separate infrastructure?

Amias Gerety (02:20.288)
Yeah, exactly. So this is what we call merchant acquiring, right? Which is kind of a funny name for me putting my credit card information on the internet. but this is Stripe, this is PayPal, this is Braintree, this is Adyen. These are these, really big companies who go to merchants and say, Hey, let us create a gateway to safely take your customer's financial information, send it to the banks.

And then the banks will send you money. And the reason this is so important is because that that word safely is doing a lot of work in the sentence. Right? So we all know about Visa and MasterCard. Visa and MasterCard hold pretty high standards and they have set those standards over many years for what it takes to process payments on the internet. So there's safety guardrails about, well, how are you storing the information? How are you taking the information? How are you?

doing two factor authentication in certain instances. How are you controlling for fraud? Who has liability for fraud? This is a whole ecosystem. Obviously I'm a FinTech investor, so this is what we do all the time, but this is a whole ecosystem of quite big, quite sophisticated companies who are allowing regular old humans to type in their information on the internet and make a payment. Now,

The place this gets interesting, where we start to poke in the direction of AI agents, is already we've got auto-complete. If you think about in your browser, we're an investors in a company called Kudos. Kudos. One of the things that they do is they tell you which is the right card to use. And then they auto-complete that card when you're doing commerce. Stripe has an embedded tool called Link, where if you've made purchases with Stripe in the past, you can check out with Link.

And that is an auto-complete. Apple will do auto-complete. Chrome can do auto-complete. There's lots of auto-complete. So that's getting close. The second thing that we see is recurring payments. So recurring payments are what happens when Netflix, that's a recurring payment for us. Netflix will have an automated engine in the back, go talk to the banks, and

Amias Gerety (04:42.336)
charge your credit card. So instead of you putting your credit card into Netflix, Netflix is going back out and making a draw and that's an automated payment that looks like a bot. that's sort of what we start to see that we're poking in this direction of there are places where automation works, but most of the time, and I think this is the central point, most of the time, remember that question that always comes if you're checking out on an unfamiliar website. Are you a robot?

And the one thing we know about AI agents is they cannot say no truthfully.

Prateek Joshi (05:14.133)
Right.

Prateek Joshi (05:21.998)
That's funny. okay, so Netflix and with the bank, clearly some automation is happening, meaning clearly it's not a human going and requesting a bank, hey, can I please charge this amount to this card? So that automation is there. So they know how to handle requests from, let's call them bots for now, right? Now, what is the big deal if

I employ a board of my own to buy something on Amazon. I guess, can you maybe talk about where does the, parts of the financial infrastructure start to break and why do these agents that I employ as an average consumer, why does this agent need to be separate, like needs its own own infra?

Amias Gerety (06:09.538)
Yeah, exactly. Why is it separate? And I use recurring payments as a good way to start here. And because what happens with recurring payments is that they're actually separately tagged in the payment system. And so when you authorize a recurring payment, there is a special communication between the bank and the merchant that has authorization to recurring payment. And that has its own rules associated with it. So we see

that even in today's system, where there is some version of automation, there's also some version of specialized data at the very least, and specialized infrastructure, specialized rules, specialized permissions. So that's, think, where we start to see that if we're going to have AI agents act, they're going to need their own tooling. The second thing I'd point out, and again, just to use, this is probably less common today, but if you think about where eBay started.

where PayPal started, remember? PayPal and eBay start as part of the same company. Now think about how eBay works. And for those listeners who are maybe too young to remember, there used to be a thing where most of eBay was a true auction. Now most of eBay is buy it now. But in a true auction, some people started writing code. They started writing sniping algorithms.

that they would watch the price, they would watch the auction, and then they would bid at the one cent higher than the previous high price at the last possible second. that's sort of an example where we've had on the internet lots of experiments around sniping algorithms. That was a bot trying to make a bid on your behalf. Now the payment still came separately. And this is, again, why I think there's some indication that you might need your own infrastructure.

eBay controlled its payment infrastructure. It was one of the first companies on the internet, and that's why it birthed PayPal, to control its own payments infrastructure. And it was because eBay wanted to enable more advanced types of payments, like auctions, and then ultimately that led to the sniping algorithm. So these are the kinds of things that I think point in the direction critique of the idea that an AI agent can't simply walk up to an e-commerce site

Amias Gerety (08:29.14)
say, maybe slowly click and pretend to be a human because there's a lot of systems that have been grown up to prevent bots. And then similarly, where we have new types of behavior, we've tended to have new types of data and new types of rules and permissions. And those have often created new companies like PayPal and eBay example, like Stripe in the rise of internet e-commerce.

Prateek Joshi (08:57.466)
During our discussion offline, you made a very good point about how the entire current infrastructure is set up such that humans, good, robots, bad. Like that the whole thing is just centered on this thing. And now we are so deep that we don't know how to separate the good robots from the bad robots from the neutral robot. So how do we establish or how do we verify?

identity of a given AI agent if we were to implement a system where agents should be able to make basic simple purchases.

Amias Gerety (09:32.576)
Yeah, exactly. I think the truth is that neither you or I know for sure. We haven't yet started the company. We haven't yet built all out all the permissions, talked to all the various parties. But I think we have some ideas. And one of the ideas that I would start with is that you might need a whitelist. You might need to say, these agents, you might need a partnership between the merchant acquirer and the agentic.

company or the agentic framework company that people is allowing you. And they would say, OK, these are the bots that have given us permission. We can validate that there are identities, humans behind them. We have some frameworks. And then when you go out to the merchant, you're saying, hey, merchant, this is a bot. It has been sort of digitally signed.

And I think that would start to replicate the kind of trust that you see, you know, think about when Apple went into payment, they did a lot of partnership work to get banks and merchants comfortable with Apple pay. So that's the type of work that I think we would need to see, in order to do this. Now I have a separate idea. And if, by the way, if this is a good idea, come find me founders, it's a bad idea. Come explain why not. But you know how, the internet today in the rise of AI.

There's kind of a robots.txt file. It tells you whether you can scrape the permissions, that sort of thing. I think you could have a robots.merchantacquiring. You could have a line on a website that says, hey, if you are an agent, if you are a robot, don't try to brute force my current checkout. Instead, come over to this other special checkout.

that is specifically designed to test for good robot, bad robot. And I think if you think about the way that fraud systems are working today, they're spending a lot of effort to keep bad robots out, to keep organized crime syndicates who are doing a thousand cards in an hour. They're spending a lot of effort to keep those out. And so I think they might want to create a special tunnel, a special gateway for good robots to come in and indicate that somehow on the

Amias Gerety (11:56.598)
you know, somewhere in the HTML.

Prateek Joshi (11:59.998)
And in this case, want to quickly, a quick sidebar on if this goes a teeny bit wrong, well, actually it went very wrong in the case of Knight Capital. And you give a great example. So maybe for listeners who may not know that, you just kind of explain what happened there in plain English and what like just kind of highlight the risks associated with this?

Amias Gerety (12:20.086)
Yeah. So I think ultimately, so let's talk about why you would care, right? So the one thing we know about computers is they can do things very, very fast. And if you've done any programming at all, you know that computers can run into errors and they can run into loops, right? So, you know, in the age of AI, I've started programming.

I did a very little bit a long time ago, but now I can just build whatever I want. We'll come back to that. like, you know, think about the way programming works. You're always doing a try, right? Now try means that's telling the robot, telling the computer, hey, do this a bunch of times. If it doesn't work, keep trying, right? Or you can do a wait, right? And then you can put in gaps. But if you don't put those frameworks, you might just say, hey, keep doing this until it works.

So the reason that that matters is if you get simultaneously into a loop and maybe you've got a false error, like a false negative, you might make a thousand purchases in a second. And whose fault is that, right? Is that the person who programmed the agent? Is that the person whose card is it? Is it the bank that accepted the payment? Is it the merchant? Where does that go wrong? That's a new way to think about it.

People might say, that's not going to happen. But I use this example in the essay that we wrote together from Knight Capital. Knight Capital was a high frequency trading firm. And they lost billions of dollars in 45 minutes. And they had a glitch in their system. It had to do with the open. their computers went on a tear.

just bought everything in sight. And it bankrupted, you know, one of the largest funds on Wall Street, the largest market maker, one of the largest market makers, and it did it in 45 minutes. And that's a crazy fact. So if you think about you or me, Pratik, we're kind of relying on our bank, our credit card processor, the networks.

Amias Gerety (14:31.736)
You know, consumers generally live in a world where we assume that people are going to protect us from our own mistakes. And that means that as we move into an agentic world, we're going to have to find ways to not make, you know, buy a thousand pairs of rain boots in 22 minutes.

Prateek Joshi (14:47.458)
Right. Right. Right. Yeah, I need a warehouse to store all of that that happens. coming to, okay, so one of the good things about all this chaos and confusion when these paradigms shift is it opens up some opportunities for startups, which is great. Now, I want to start with the big gorilla in the room, Stripe. Clearly, it's in their wheelhouse.

We all, I some people disagree, but I think there's a general agreement that the agents are coming. Like good or bad, you'd like it or not, but we were in the future where a lot of like little verticalized agents would be doing a bunch of stuff on the internet. And we just need to know how to handle that. So let's start with Stripe's recent announcement, right? They released their SDK to enable agents to make payments. The timing of our essay is almost freshened looking back. So I'll take a little, we can take a little bit of credit for that, but so talk about.

What does Stripes SDK do just in simple plain English? And also what are the gaps? Like what didn't they address in this release?

Amias Gerety (15:52.364)
Yeah. So first of all, think huge credit to Stripe because although we like our essay did at least come out before they released their SDK, the fact that they released it, you know, that, know, in a couple of weeks later means they were definitely working on it before we wrote our essay. So, huge, huge credit to them and their team for, for, for getting this out there. I do think given the excitement, it's, you know, just a demonstration of even a very large company now continuing to move quickly.

Prateek Joshi (16:06.356)
Right, right.

Amias Gerety (16:19.832)
I know they're partnered with OpenAI, they've partnered with others. So that's pretty exciting for them. But if you go into the details, there are two really important notes about what they've done. The first, which is very clever, is they use a virtual card framework. So a virtual card is a non-physical card, and you can create a virtual card

You know, there's good infrastructure for this where you create a valid card. It's a derivative of an existing card account. And then you can put special permissions on that card. And what it appears based on reading their blog post is that they're using that virtual card framework, probably for one use. And that's actually extremely clever because it's the, it's the simplest way to prevent the example that we just.

Walked through of you a thousand purchases in a minute Because they're gonna create a single use virtual card In the SDK and then that SDK can and then that virtual card can be used in a purchase. That's that's very clever The the second thing though I will say is if you read in the blog post They they still say please run this in a test environment

So it's clear that they, like us, are not quite sure what the parameters of failure could be. so I think huge credit to them for getting this out there. But also, I think if you read the blog post, you can tell that even they haven't quite figured this out. And I think that continues to be the opportunity

The last point that I think is interesting about Stripe is that they did this on the issuing side, meaning on the side that sends the money, not on the acquiring side. And that's also very interesting. And it points to this possibility of an opportunity of a different type of merchant acquiring. Now, Stripe has so many merchants, right? They're already so dominant on merchant acquiring.

Amias Gerety (18:39.0)
that it does seem like they potentially would be better placed than anybody to make this work. But if you read the release that they've done, it's clear that there's still a lot to work out.

Prateek Joshi (18:51.566)
I think that that's great. I think the observations are already astute here. And the one good thing is Stripe just validated that this is a problem worth at least talking about, at least thinking about, because clearly they see the wave of agents coming and they don't want to be caught flat footed. So they did something and obviously they have left some gaps. Now Stripe in the next release can very well cover those gaps or it could be an opportunity for startups.

If you were to identify reasonable gaps that Stripe can do it, but so it's a very big company. It's a giant company and they know by design, they won't be able to just do every little thing. what role do you see for startups here? If they want to build, founders are thinking about this. They want to build something useful. What are the areas to attack, if you will?

Amias Gerety (19:46.552)
Yeah. So I think the, there's probably two big ones that I would think about. And I do think that Stripes for a does change the landscape significantly. So, so let me talk about two that I would see given this. The first is there are a lot of companies experimenting with agentic frameworks.

It is natural for people who are experimenting with the Gentic frameworks to worry about error control, containment, alignment, hallucination. But these are machine learning engineers. These are AI engineers. The chances that they also deeply understand the payment system are relatively low. So I do think that

Just as Stripe will be trying to work with some of these companies, I do think there's an opportunity for payments experts to really go in there and build deeply into the agentic frameworks and the companies that are building agentic frameworks in order to connect that. that last mile, that connection point, is going to be, I think, rich and varied. And I think there's going to be a lot of opportunity for great product and great product insight there.

The second thing that I would say is, know, Stripe certainly controls a reasonable number of the merchants on the internet, but there are hundreds and hundreds of other payment getaways. There are hundreds and hundreds of other checkout companies. And I think many of those companies have clearly not made announcements about facilitating agenda commerce. And so I think that's the other thing that I would do is I would move in both of those directions and say,

you know, to anyone who considers themselves a competitor of Stripe. How can I help you enter this agentic world with a software layer, but really building on top of maybe those merchant relationships that already exist. So if I were a founder today in this space, I'd probably look in those two directions because I do think this is a splashy announcement. It validates the problem. It shows that you've got to move quickly and aggressively.

Amias Gerety (21:55.884)
But it also, because it validates the problem, it will create a competitive response. And I think that there's room for a very set, a smart set of team or teams to go into on both sides and work around those last mile problems and work parallel to the announcement that Stripe has made.

Prateek Joshi (22:17.334)
Right, that's wonderful. And whenever there is, in commerce, there's money involved, which means people expect, all parties expect strong governance because it's money. can't just be casual. So I have two questions in this section. One, for AI agents, what new challenges do these agents introduce with respect to disputes and chargebacks and reversals? So maybe talk about what's the status quo today and does that...

Do we need a new framework for agents?

Amias Gerety (22:49.376)
Yeah. So first of all, I think these are exactly the right questions. So charge backs, reversals, by the way, there's a lot of first party fraud happening even apart from this. you know, one of the big issues in payments today is like TikTok videos that say, Hey, by the way, do you know that if you can go into a department store, buy a lot of goods and then call your bank and tell them that it wasn't you, if you've never done it before, they'll probably just refund you and you can keep the goods.

So this kind of fraud modality is happening on the internet already. And it's a real big problem in payments. So now let's go to the agent. I did this in agent experiment, right? So there's both the truthful and the untruthful version of this, where different banks will have to design their own risk parameters and their own customer service parameters on like, you know, do I think Pratik is lying about his agent having made a mistake?

and he's just trying to get free money or do I believe Pratik because he's been a good customer and I want to be part of the future. I don't want people to switch credit cards because you know, their agent made a mistake. So that dividing line is one really important piece of governance that I don't think we've figured out yet. And I think the banks are afraid. I think the banks are not sure how to handle it because of these, know, do I want to be part of the future? Do I want to disappoint my customers? And do I want to just be a mark?

for fraudsters, whether first party or third party. So that's probably the most important thing. The second thing that is happening though, and it's the holiday season, right? People are buying gifts for each other. And I think the other really important governance question is, when was the last time you were super excited, like you got the perfect gift? And that's not even your money.

You know, like, was the sweater, sure, it's nice. Like your partner got you a sweater, that's fine. But like, did it really fit perfectly? Was it exactly the right color? And now, and that's like, that's like your partner, your friend spending their own money. Now imagine it's your money, and it's your AI agent. And you get that same pit in your stomach of like, that's not really what I wanted. You know, right? So I think

Prateek Joshi (24:53.39)
you

Amias Gerety (25:09.43)
I think that is actually a really underappreciated challenge in these agentic frameworks and the place where it will come out, I know we're laughing about it, it will come out in charge backs.

Prateek Joshi (25:20.824)
Right. That is hilarious. I need to make sure my wife doesn't listen to this episode because that is a very real thing. that's why my friends, I kind of in my, circle, I famously like don't you're forgiven. Nobody please give me anything because not one of you, my friends have given me a thing that either fits well or I like the color. So like, just like, thank you, but no thanks. That's, that's a hilarious thing. Right.

Amias Gerety (25:46.08)
Yeah, now that we're all grown-ups, it's like, look, if I want it, I'll buy it, you know?

Prateek Joshi (25:50.446)
That is exactly right.

Amias Gerety (25:52.8)
And now think about it's your AI agent. Are you really sure your AI agent is gonna get everything right?

Prateek Joshi (25:58.446)
boy, that's hilarious. And the second part of the question is a little more about regulation. And in another many famous Jamie Dimon interviews, he just said something to the effect of, when it's money, the government needs to know who has it, where it's going, how much of that is. So they have to know it. So when it comes to AI agents handling money, are there any...

regulatory concerns or changes like can we just build and shape or are there new rules to be aware of?

Amias Gerety (26:32.92)
So I think there probably will be new rules. And again, we talked earlier about recurring payments, how there's special data tags, special rules inside Visa and MasterCard for how to handle recurring payments. And I would expect something similar develops over time. But I think this is also part of why special, I think why this is an opportunity for startups. Because I think getting agentic frameworks correct is going to be so hard, they are not going to

also want to take on the regulatory burden of becoming a payment processor. And so I do think, at first, there's a whole host of rules around what people call PCI compliance, how you store the credit card data. There's KYC. There's fraud rules that Visa or MasterCard impose on you. So there's disputes and chargebacks and how you handle those things. Those are

regulatory. So I think there's a whole thicket of, you know, this is a regulated industry and to do it right, you've from the data to the way the money moves to the way you make sure it's not moving for criminals, there's a host of laws and regulations and rules. And so I think you're going to that's partly why you're going to have specialist firms that sit in between the developers building agentic frameworks and the payment system.

Prateek Joshi (27:55.566)
Another point that you highlighted that is very interesting is the potential for network effects here. So I want to start with just understanding the needs of the merchants and on the other side, the needs of an AI developer platform that is building these many agents for customers. So from the merchant perspective, what do merchants need to feel confident in accepting these AI-initiated transactions?

Amias Gerety (28:24.566)
Yeah. So the first thing they need is they need trust that these transactions will perform like other transactions. Right. So if I start accepting, let's imagine not a world where I get confused and I accidentally accept agentic transaction, but I go out and I say, Hey, look, your agent can purchase things from my store. Now, if my charge backs, if my reversals.

on those things, if my margin, right? Maybe those agents are super good at finding out when I have a sale and then I just get a flood of purchases when I have a sale and I never get purchases otherwise, right? So from a merchant, you need to know that these transactions, you're not going to lose money on them. Nobody wants to do a transaction when they lose money. So that's the first thing. And then the second thing is...

You know what a merchant wants from a merchant acquirer from a partner like Stripe. They want it to be easy to implement. They want it to be fast to get their money. They want their, they want their merchant acquirer to control fraud on their behalf and to have an open dialogue with them about fraud. And so these are the types of things, both in terms of what the merchant wants. They want to make sure that the agent is actually doing the purchase, right? No one wants to ship a t-shirt and then have the

T-shirt come right back because UPS is not going to refund you on shipping both ways, right? So so that's one piece and then the second is they're gonna want that merchant acquire to be sophisticated about fraud and about the separation between good bots and bad bots

Prateek Joshi (30:11.042)
And on the other side, the people who are building these AI agents, AI developers, what do they need from the merchants, from the infrastructure to just say, hey, I have no malicious intent. I just want to have an AI agent make normal, regular purchases. So what about this side? What do they need?

Amias Gerety (30:32.576)
Yeah. So they need to know that the payments will be accepted. And I think the mirror of these two needs is why there's some potential for network effects. And let me, let me give an example of critique from, you know, just being a regular person in a country that's not the United States. So in Brazil, for example, credit card acceptance rates. like a regular person making a regular purchase with a credit card.

The acceptance rates, the chances that goes through are like 60, 70, maybe 80%. In the U S that would be 97, 98, 99%. Right? So this is actually a real problem in the world. There are many countries, there are many people who find, have the experience of I tried to make a purchase and it didn't work. Now, imagine you're an agentic company. Your whole brand promises the agent is doing this for you.

Well, also, you know there's going to be a lot of steps. That's the whole challenge of agentic frameworks. There's going to be a lot of steps between typing the prompt and getting the result. And so if those steps can fail in seven different ways, and maybe at the end, you have a card acceptance rate of 60%, 70%, that's going to destroy the brand value of the people who are trying to build these agentic frameworks.

Prateek Joshi (31:57.07)
All right. And the potential, and that's a great point. Actually, many times you just think at least like, you just swipe the card, it works every single time. So you just don't, you you don't stop to think about cases where, my God, that this is a, it's not like a guaranteed thing. It's a probabilistic thing, like 60, 70%. That's not, so 30 % of the time you have to figure out how to pay. The card doesn't work. It's a, that's crazy. And no, going forward,

Amias Gerety (32:23.917)
That's right.

Prateek Joshi (32:27.339)
agent commerce, like just in general, let's say agents start working, people figure out how to accept payments. So how do you envision agents shaping commerce? And one of the couple examples that you mentioned that was interesting was like humans, buy, you know, maybe one, two, three items a day, a week, very, very low frequency.

agents can make a giant number of microtransactions across a bunch of websites and to them it's completely normal, right? The speed doesn't, they're not bound by the human limitation. So how do you envision agents shaping commerce?

Amias Gerety (33:04.354)
So I think the first thing is that I think this initially is going to look more like search than we think. This is even kind of a meta search modality before it becomes a fully automated modality. I mean, think about interacting with a travel agent. Hey, I want to fly from DC to Chicago on this day. I want to get the earliest flight that's still direct. The travel agent.

always shows you the flight before they purchase. Okay, I think this is the flight that you want. And there's a lot of nuance to that, right? So I think the first thing I would observe is that this is going to look more like search than we think, or than we hope, I should say. And the second thing is, and I think your point about volume is fascinating, right? Something we didn't cover in the essay, but imagine

an agentic framework that is saying, actually I'm acting on behalf of a thousand people. And they're all making the purchase this morning. Maybe I, as the AI agent could negotiate a discount with this retailer. I mean, that would be a fascinating thing if imagine an AI agent aggregates demand and then goes to Walmart and target.

Prateek Joshi (34:19.703)
Right.

Amias Gerety (34:30.954)
interacts with their AI agent and says, hey Walmart, I've got a thousand orders right now. I can either send them over to target or I can bring them to you. Can you do something for me? I mean, that would be a fascinating pro consumer. would change the competitive landscape in a lot of different ways. It's still downstream from search, but if you think about the types of interactions that are possible, you know, who knows? That would be a fascinating outcome.

and speak to the possibility of network effects if you could really aggregate demand across these agents.

Prateek Joshi (35:04.142)
Right, that's a fascinating outcome, as you said. And the thing is the construct of, hey, I have aggregated demand, now give me a discount. That is very old. People know that. Like if you find a way to aggregate a bunch of, yeah, of course I'll give you discount because you're giving business. So think that part, given that that is well known, I think agents are just replacing other nodes who are usually humans or companies to aggregate demand and demand discount. That's a fascinating,

It was very curious to see how that plays out. Okay.

Amias Gerety (35:35.83)
Yeah, we'll see, but these are the kinds of things that I think make this area so exciting.

Prateek Joshi (35:42.87)
Yeah. And we talked about merchants. We talked about AI developers. Now, consumers who are the ones who actually buy, they pay money. They're very, as consumers, I'm finicky. Like I have no patience. I don't want to wait and I want everything now. And I trust nobody. So what would it take to build trust with me, the consumer, and also does...

some other party in this value chain, do they need to underwrite so that consumer never, I don't, I never had to pay for any bad things that the agent does. Like by mistake, it purchases like five items. I needed only one and I don't want to be on the hook for the other four. So you, somebody else figured it out. So how does that play out?

Amias Gerety (36:28.978)
The honest truth is I don't know. mean, think about even something as sophisticated as searching on Amazon. Right? Sometimes you want to search only prime. Sometimes you only want to search one size. Sometimes actually like, know, medium or a large either would work for me here. Or, you know, I really want red. I want a red sneaker. I'm glad to have nine and a half instead of 10.

Right? Well, if it's Adidas, yeah, that's fine. If it's Nike, only 10 or 10 and a half. Right? There's so much nuance baked into our preferences and into our own brains that I do think it is going to be exactly as you describe a lot of work for consumer oriented agentic companies to build the trust that they actually can carry the full nuance of a consumer's preferences.

Prateek Joshi (36:54.167)
Right.

Amias Gerety (37:22.698)
around with them. And by the way, it's not, and this is what's so exciting about AI, it's not a data problem and it's no longer necessarily a reasoning problem. Now it's a communication problem. What are your true preferences? know, what are the nuances that go in your head? How do you trade off, you know, United, a United flight, maybe that's where you get more miles, but it might get you there.

an hour earlier than you really need to be. And so now you have to wake up an hour earlier. So which do you really prefer? Do you want Delta, but you have to wake up at 5 a.m. or United and you wake up at 7? I don't know. What's the risk of the meeting on the other side? So there's so much nuance in every purchase. And I think we don't yet know how to carry the full nuance of a consumer's preferences around the internet.

Prateek Joshi (38:17.934)
Amazing. I have one final question before we go to the rapid fire round. And this is more about the future. So going beyond agent to commerce, just overall, the AI is coming into finance and that's such a big trend. So what advancements are you most excited about? And also what opportunities or what corners of this big area excites you the most?

Amias Gerety (38:45.026)
So I think the most important thing is just productivity, right? Finance is downstream of the real economy and it's a very big chunk of the economy, but a lot of that chunk is made up of human beings and processes that could be more efficient. And so it would be a win for global commerce if financial services actually became a smaller chunk of the economy and still carried the same amount of money from point A to point B.

And that's why I think, you know, there's a lot of excitement, especially around compliance, around customer service, things where a computer, a sufficiently smart computer could be both faster and better than humans. And these are not like the most exciting jobs in the main, right? It's not actually that fun to work in a call center and have someone yell at you. You'd prefer

to only deal with the complex customer problem. So I think there's a lot of room for productivity gain in financial services. we still, think most financial services companies don't have enough confidence with the error rates, or they will only use it for what I would call fault tolerant systems, places where there's a lot of give in the joints. But I think, and this is true for the broader economy.

This is a potentially huge deal for productivity and productivity does increase human welfare. That's great news.

Prateek Joshi (40:13.432)
Right. Yeah. No, I love the way how you put it where you want this entire sector to pull the same amount of weight, but with this small chunk of the pie or like just spending less, meaning more efficient systems, more efficient people. that's wonderful. All right. With that, we're at the rapid fire round. I'll ask a series of questions and would love to hear your answers in 15 seconds or less. You ready? All right. Question number one. What's your favorite?

Amias Gerety (40:37.025)
Alright.

Amias Gerety (40:41.56)
My favorite book is The Origin of Species by Charles Darwin. It's an old book, it's so good, and he's actually a good writer, and just how much of biology is in that one book is incredible.

Prateek Joshi (40:54.784)
Amazing. What has been an important but overlooked AI trend in the last 12 months?

Amias Gerety (41:02.53)
think AI-oriented investors are overlooking the competitive pressures that will come from AI-driven companies. So a lot of investors are looking at older legacy industries saying, hey, we can automate. We can completely change the cost curve for this industry. Let's start a company to do that. But if because that cost change is coming from the power of AI, the competition associated with that cost change will be

could become very intense, very fast, and the overall margin of the industry might stay the same.

Prateek Joshi (41:36.812)
Yeah, yeah, that's a good one. All right, next one. What's the one thing about agentic commerce that most people don't get?

Amias Gerety (41:47.138)
I think we've already talked about this, it's that the nuances of consumer preference are so, they're so varied and they're so detailed.

Prateek Joshi (41:57.28)
Right, Yes. What separates a great AI product from a merely good one?

Amias Gerety (42:05.377)
Honestly, it is product, so to speak. I am amazed by how important the quality of the design choices and still how important the quality of the human to machine interaction is to making the magic happen even with these AI systems.

Prateek Joshi (42:23.648)
What have you changed your mind on recently?

Amias Gerety (42:27.586)
So I used to not read fiction and I've started reading fiction again. So, you know, I'm a very analytical fact oriented person and I used to think, you know, only products of the world, no products of the mind. And I've changed my mind on that.

Prateek Joshi (42:31.662)
Right.

Prateek Joshi (42:44.396)
Maybe a quick part B to the question, why did you change your mind?

Amias Gerety (42:49.25)
So I think part of it was just realizing that I'm very focused. I read a lot of old books. That's mainly what I read. And that the of the classics of human knowledge were just as often fiction as nonfiction, especially as you go back in time. Right. So whatever your great myths are, I feel like we need to read those myths, not just read about them. And so I started reading fiction again.

Prateek Joshi (43:18.702)
Amazing, love it. All next question. What's your wildest AI prediction for the next 12 months?

Amias Gerety (43:26.018)
So I think engineers are going to start losing their jobs faster than we think.

Prateek Joshi (43:34.818)
Interesting.

Amias Gerety (43:35.616)
I think that the, and I'm not sure that it will actually see a decrease in employment because I think productivity is going to go up, but I think it is already true that product and engineering have been significantly collapsed and we're already seeing, I started using cursor a month ago. I know this is a very hot product out there, but it is amazing. I am literally coding a web app. I'm not a technical person.

I have no engineer working with me and I'm using natural language prompts and a little bit of reasoning to generate like a pretty performant web app with no engineer. And I think that's, I think that's a little bit under the radar. How much of engineering is just like painstakingly typing out like the word, phrase div over and over again in, your react app, right. Or, or in your HTML.

Prateek Joshi (44:15.575)
right.

Prateek Joshi (44:32.066)
Bye bye.

Amias Gerety (44:33.721)
And we don't have to do that anymore. And I think that's going to change the employment levels for engineers fast. So that's my wildest prediction. I think we might be closer to the end of human engineering than we think.

Prateek Joshi (44:45.206)
Right, there's a funny meme that recently is like, somebody goes and asks like a very senior web dev, like, hey, do you know CSS? And the response is, can anyone truly know CSS? It's like, so yeah, so first of all, nobody wants that. And if that can be automated by cursor or like AI coding products, then yeah, good riddance because...

Amias Gerety (45:06.53)
I mean, I have hundreds of lines of CSS code, CSS, and I don't know the first thing about it, Pradeek.

Prateek Joshi (45:12.366)
Yeah, and that's not going to change even if do it for like 20 years because my friends tell me that they've been doing it for a long time and I don't have an idea. it's funny. All right, final question. What's your number on advice to founders who are starting out today?

Amias Gerety (45:29.88)
So the number one piece of advice is you need an insight that makes things that are hard for other people easy for you or your team. If you don't have that insight, competition is going to come get you really fast. So you have to look for problems where everyone else is doing it wrong or they're making it hard when it doesn't need to be hard. And until you have that insight, it's hard to build that

first just lightning rod of

Prateek Joshi (46:01.942)
one of those spreads where long shot like it has to be easy for you and it has to be difficult for others. So that Delta, the whitening Delta kind of gives you a real advantage. So it's very interesting. Amayas, this is a brilliant discussion. It's even better than, than, you know, a past discussion, the thing we wrote, it just came to life. So thank you so much for coming onto the show and, yeah, talking about

Amias Gerety (46:24.802)
Thank you.