No One is Talking about the Downsides of Agentic AI

Now what's happening with agentic AI is people are adopting it way faster than they were adopting internet.

AI agents are some of your best intent-to-purchase consumers you could ever get because they're ready to buy.

You'll get more transactions. You might think you hit the jackpot, and then all of a sudden those agents again are handling the chatbot conversations, conversations with your agents, uh, in chat sessions to get those refunds.

What's up, everybody? Welcome to another episode of Fraud Boxer. I have a big treat for you guys today. We're going back into the Agentek world. And we're going to be talking about some real stuff and what fraud and Agentic and how those things work together. But I got a heavy hitter. You know, I promised 2026 to be the year of heavy hitters. So I got Simon Marchand here today. How are you doing?

I'm great, Jordan. Thanks for having me.

Yeah. Sorry to put you on the spot right when you're taking a sip of your espresso there, you know, but, uh, it happens. So tell me, um, you and I have known each other for a little while now. Um, I think we actually started doing some calls in the COVID times. When we were all trying to figure that out, if I'm not mistaken. I think it's possible that we shared the stage at some point. Yeah, yeah, yeah. And, um, you have always been pretty good at what you do, so I'm really happy to have you on here. So why don't you just give, um, the audience, just in case someone doesn't know you, which they probably all do, a little bit about your history, where you've been, how you got here. You and I have been in the industry about the same amount of time, so, um, you and I are what we call OGs in this space.

So yeah, yeah, yeah, I've been there. Like, you can see some of the gray that Yeah, 16 years I've spent in that field is bringing me. But yeah, I started so 16 years ago and I started as a practitioner. So started with a small night shift running, you know, in a bank, a small team of 2 guys processing chargebacks. So very much transactional front focused. And then little by little, I moved from there to building a team, growing the team, starting to work on systems and implementing rules. And from transactional, we went into identity theft and eventually mortgage fraud and all the type of fraud you can imagine you have in a bank. So it kind of broadened my view on fraud and eventually I kind of reached what I felt was the limit. So I moved in telcos for Bell Canada, which, you know, for your American viewers would be like an AT&T. So when they're managing corporate security with a team of analysts that was looking at still some identity theft because it's still the biggest problem that telcos have, you know, when you open those accounts under a stolen identity, try to get the phones for free and then resell them. But then we had a lot of very specific fraud types, right? Account takeovers for SIM swaps, PBX hacking, SMS pumping, uh, dealer fraud. So, you know, dealers just colluding with fraudsters to facilitate—

All the mall kiosk folk.

Yeah, kiosk folks, you know, the ones that you take for a week, then they disappear and all of a sudden, you know, fraud's just exploding. So I did that for the first half of my career. Eventually I moved on the other side, uh, with a company called Nuance. So I was working on voice biometrics technology, which is used by most banks in the US now and in Canada even. And then I started evangelizing a lot more. So talking to people about what they can do with that technology, why they should use it, things like that. Did that for a couple years. Then I moved to a company called GeoComply, working on geolocation. And now I work with Datadog, working on agentic AI, agent trust, and all the kind of fraud that can leverage various types of agents and bots and automation online. So yeah, I, I played both sides of the, of the fence and I'm very happy that I got to see so many fraud types and so many technologies.

So, you know, Datadog's been a friend of the show for a couple of years now and has had been on the show a number of times. Very big fan of the product that they do. Been to the office in New York and SoHo, had a lovely lunch with them. So very big fan of all that they do. And one of the things that I do like to say about them is, They're usually forward thinking on the technology. So I had Benjamin on a long time ago, um, to talk about bots, what was happening in the bot world. That was even before his book came out. Uh, got a lot of really interesting information about bots, but over the last year, bots have evolved, haven't they? So now we have this new thing, which DataDome came on and talked about at the very beginning called agentic commerce, agentic AI, uh, which is kind of a form of a bot. But, you know, we have some episodes that have talked about it, but how it comes in and how it looks on websites looks like bot traffic to, like, a person that's not super familiar with it. So am I right or am I wrong in there on that one? Like, how do you— how could you frame it?

You're right in parts, right? There, in some cases, your agent will look like a bot. But in some cases, it will do everything it can to appear to be human. Because it wants to avoid some of the protections you have in place. So you know, even legitimate ones, even ones that, you know, a real person is using for legitimate purposes might try to avoid the defenses in place by acting human or appearing human. And that's because they wanna make sure that they can get access to all the data that they need to on your website. And yeah, it really evolved, right? Like 10 years ago, you, you had bots already. It was not new. What really changed in the past, let's say 3 years, is the ease with which you could use those. So yeah, it started with LLMs, the first versions of ChatGPT. Obviously, you could leverage those to ask them how to build your own bots, how to build your own malware, how to run your own fraud schemes. So that started being a concern. But now that we're seeing agents, we're seeing, you know, those very small autonomous components that can work together, orchestrate each other, that are purpose-driven, goal-driven, and that can self-improve in real time. So they can see and understand that one attempt is not successful. They can find another way to do it. They can decide to go through another path. And that's really what changes things. Uh, it's not as dumb as the original bots that were just kind of repeating an action. And as you had to go in, yeah, you had to go in, you had to tweak it, you had to make sure it worked again. And then when a patch, a security patch would come out, you had to figure it out. You had to adjust your bot. So it would take several weeks, you know, sometimes. But now that you have agents, uh, the problem is the agent will see that you're releasing a new security feature. It will adapt in hours. So it changes the way you fight.

One of the things, like, I just want to level set here for everybody that's listening is when we talk about AI agents in this particular capacity, we're talking about consumer, like, AI agents, ones that are going to perform a task for a consumer, not like a chatbot on a website or some sort of backend thing that you might be using as an enterprise company that, like, having an agent do something for you as a work task. We're talking about a consumer. Asking an agent to go complete a task for them on their behalf, right? That's what we're talking about here. Exactly. Yeah.

When it started, you know, large language models and that, you know, first iterations of what we'd call AI, like in the modern years, uh, it, it was just an interaction where they would gather information, distill it down for you, and then reuse it. So that would be a chatbot. But slowly but surely, we saw those agents where very legitimately they can be used by consumers to, for example, crawl the web and find information on different products. If we want to build a plan for a trip, it can, you know, go book a hotel, go look at attractions, like touristy things you can do. It can book flights. I love that part.

That's my favorite thing about them, by the way.

And it's amazing the time it can save you. And it's great that you can give it a task, go do something else, come back, and then, you know, you have something very complex that was executed on your behalf. The downside of it is if an agent can access all of that information, whether it is to gather or make a purchase or a reservation on your behalf, a Froster can do exactly the same thing and they can do it very rapidly and they can do it at industrial scales, which, you know, was never a problem before. But yeah, like from a legitimate use, use case standpoint, like it, it, it's a great tool we have in our lives today and we don't even realize sometimes that those, those bots are working on our behalf. You know, when you, when they go crawl the web because you said find me the next phone that I wanna buy, there's something happening in the backend to give you those results, but that creates a threat for the different merchants and different websites that are being visited.

Yeah. And I think, you know, that is, is the excitement around these things, cuz I think that they're cool. The, the actual use case, I mean, you guys have some numbers here on the adoption of it right now. But the, the, the excitement about them is it's a cool concept. And then how many people are actually putting it to use is a different set of numbers. I will say, I think just like I've said it a couple of times, just like with Apple Pay, it's after you make your first one or you use it for the first time, then it's a lot easier to use it. I think that the, like, for me, I know personally, I usually, like, I just did a trip to Hawaii and I used Gemini to plan my whole trip. But I did stop short of having it book everything for me. I did still do that myself. Now, now I probably would have felt more comfortable allowing it to make some of those reservations for me, just knowing what I know now and, and seeing how accurate it was. Because for me, you know, there was, there was the, the trust, and we'll get on to the, the, the merchant trust, like from, from like my side of what an AI agent's doing. But from a consumer side of the trust, it was I was still a little leery about it messing up, but when I went through its entire list and booked everything personally myself, and it was all right and everything worked out perfect. Now the next time I go and do that, I would have a lot more trust in it doing and being correct about that, which once I do it the first time, then it's like, all right, go find me a new camera that can do greater than 29 minutes in 4K and export through USB-C to 4K. And then I'll just, and then I'll go pour a cup of coffee while it goes, finds it, buys it, and sends it to my house, you know? And that's like, it's like I had to dip my toe in before I went to test the water before I jumped all the way into the pool. But now I'd be more willing to jump into the pool. And, and do you have, can you read some of the numbers off that you guys might have on, on like consumer interest and possibly even adoption if you have it?

Just jumping on what you said, right? A few months ago, you weren't comfortable. Now you're comfortable. And I think that's really what is transformative from a lot of businesses' perspective. You know, when, when internet. Arrived, it took a lot longer for the general public to adopt it and start using it to transact and for it to become a new revenue stream. And the ones that picked up on the potential it had, the businesses that set up, you know, online stores and managed to accept payments securely very early on, they, they, it, it was the move of the century. Now what's happening with agentic AI is people are adopting it way faster than they were adopting internet.

'Cause people don't wanna miss out.

They're purchasing.

Like, yeah.

We're seeing, just to give you an example, right? We're seeing, um, just in 2025, you know, take the first half of the year, give or take, the bot traffic that we measured moved from, you know, 2 and something percent of the overall bot traffic being agents and quadrupled in 6 to 8 months to like 11%. So it's very, very rapidly shifting.

All bots.

Of all of us now, 11% are agents. So there, there are those advanced tools that are browsing on our behalf and making transactions on our behalf. So it shifts very, very rapidly. You know, Forrester talks about, uh, 33, 35% of consumers in the US saying that yes, they are willing to move to agents. They're willing to use those tools. Uh, you know, those are numbers we could never have thought of 2 years ago. So obviously businesses are putting a lot of pressure internally to, you know, adapt to this new revenue stream and this new growth opportunity, because if you're not ready to receive those agents, your income's going away, your growth opportunity is going away, but eventually your established audience will go away because your established audience will start using agents. And if you're not prepared, those agents will not go to your website. So you have no choice but to really be sure that you're ready and set to receive them because otherwise you'll just disappear.

So that's where I, I've done a couple of presentations this year on stage about, um, AI agents and a couple of different opinions that I've seen from different websites on, like, the original and, and initial opinion from, like, DevOps teams and any sort of bot security teams was block all the traffic because it's a bot. And I was screaming from the stage, don't do that, because if I, as a consumer, go out and ask for something and your website has like, let's, like, let's say I'm selling this light. Okay. Over here. My website has that light. Another website has that light. Me as a consumer is looking to buy that light. I, this website's blocking all of it or my website's blocking all bot traffic, but this website isn't in that list of results that comes back from the initial LLM. It's only going to be that other website. It's not going to be mine. I might have a better price. I might have a better product, but I'm not even going to be listed. So at the end of the day, like, bot, not all, you can't block all bot traffic. You need to be listed. Like you want to be listed. And I will say, and I think we had it on here too. Um, I don't want to jump ahead too much, but for anybody listening and we'll talk about it again, but these consumers that are using these agents to make a purchase that says, go buy me that light, they're buying. Right now. They're not window shopping. They're buying right now. So AI agents are some of your best intent to purchase consumers you could ever get because they're ready to buy. Now we'll get into the trust and the risky side of that. But I think that that's important for people to understand is like these agents, they're here to stay. It's growing and it, it, it's good traffic as long as you can identify it. Now that's where you come in to tell me about it.

Well, I'll tell you even like one last thing on that, like people need to understand that the same way that we saw social media compete to capture our attention and keep us in the platforms, the ChatGPTs and Geminis and Claude of this world, they want to keep us in their product. They want to keep us in because the longer we're there, the more chances they have, they can monetize information we're giving, train their model, improve. So they're working on making the shopping experience better. Directly in the LLM, directly in the ChatGPTs, Geminis, and the others. So if, as you know, retailers, we don't also recognize it's not just about the agents, it's about the people behind the agents that want to keep our customers there and not browsing on our website. Like, this is where it gets very, very dangerous from a business point of view if we don't understand what's really happening.

It's almost like they're doing the shopping cart optimization for you at the end of the day, like the actual agent, like like the ChatGPTs, the Geminis, and all those. I'll even say, like me, like, you know, I'm, I'm a big, obviously I grew up in the tech world. I Google everything, but lately Google hasn't been giving me the results that I want well enough. I've been dropping it in Gemini. Like I have, I've been trying to give away this giant plant that I have in my backyard by my barbecue. Huge. Takes up too much space. So I put it on the local Facebook group here. There's our social media callback and said free if anybody wants it. Now people ask me, what is it? I literally went out, took a photo of it in Gemini and said, what is this plant? And is it toxic to dogs? And it instantly told me exactly what it was. Yes, it's toxic to dogs. Thankfully my dog didn't eat it. And I was just like, I couldn't have just dropped that into Google. It wouldn't have worked as well. You know, Gemini told me in like a second. So then later that day, I found these little berries in the backyard. Same day. I did the same process, told me the same thing about the berries. And now I got to go tell the neighbor he's got to cut his berry plant back because it's going to make my dog sick. So it's just like, it's crazy how you're right. Like it used to just be, I needed it to do, to listen to my podcast and give me a summary of it. Now it's like, it's every day in my life. It's crept into every day of my life. They got me. They got me.

It got you. And now, you know, it's on your pocket. It's in your phone. It's on your computer. And, you know, it's even in the personal assistants that we have home, right? Like everybody was about Google Home and how you could talk to it and do stuff with it and interact with it. Truth is the promises of those initial assistants never really materialized, right? Nobody really started making purchases with Google Home or Alexa.

No, we just sell our blocks.

Now Google Home is embedding Gemini directly. They changed it.

I know, I tried to do it. I tried to do it with this speaker right here.

Yeah, it would work eventually. It will get— because the conversation will be much, much more natural. So it won't even be about, you know, being in front of a computer. We will eventually get there where we can have a conversation with an assistant that will also make transactions on our behalf, reservations, uh, purchases, or even just look for information and then send it somewhere else so we can read the summary later. So, so yeah, there's, it's everywhere. It's unavoidable, I think, to, to a certain extent.

Mm-hmm.

And, and we need to prepare for it.

So let's talk about preparing for it because we talked about all this awesomeness, all this goodness, all this potential that these agents have. But as a merchant in the fraud space, I have to go back to Well, how are they gonna screw me? So tell me, tell me how they're gonna screw me essentially at the end of the day. Like, how, how is my fraud day not gonna get easier because of AI? How's it actually gonna get worse?

Oh, it's gonna get way worse, way worse. And I think the, the first and main reason for it is the speed at which you can get attacked now. So if you have a team of 20 people reviewing orders, reviewing account creation, let's say you're in the financial sector or your neobanks or, you know, some services like that where you have some KYC process., and then your account gets open, and then from there you can get access to other services. You probably have a team of 10, 15, 20 people today. You have some legacy rules that are looking at very specific elements. Maybe you have some basic scoring or some AI-powered scoring that raises some of those cases, but you can still manage to do manual review of all of this because people are manually creating those accounts. Now, when agents start creating those accounts, it's not one every 5 minutes. It's one every second, and it's huge volume. So it's not just the volume of fraudulent attempts, it's also the impossibility to work through those cases that will eventually let some go through, uh, because of course you're not gonna put everything on hold until it's being reviewed by a human. Uh, you're gonna have to let it through eventually if the business doesn't want to collapse. And I think that's the main threat. Those agents today can orchestrate each other. You can have, you know, one agent that's gonna be talking to different agents that can look for information, uh, distill information from a list of stolen identities, go gather missing information on social media, uh, generate a fake ID, a deepfaked ID, then try to get through your process. If it gets through, then another agent might manipulate that account either to make a purchase, wire money somewhere,, do some smurfing, you know, if they're into money laundering at all. So now we're seeing the ability to drive those activities at scale. It's not giving you any kind of new fraud threats from a pure, like, what's the scheme perspective? It's the same fraud.

It's just scale on them. It's scale and speed.

It's 1,000 or 10,000 worse and 10,000 times worse. Like, can you really grow your team by 10,000 people? No, of course not.

Definitely not. Uh, and that's why like everybody's going to say, probably say this, we'll just use AI to fight it back. But here's the thing is, you know, typically our, we all do, all of us use AI in some way to, to fight fraud on our site, but we rely on data in order to fight that fraud. And an agent, like to even to your previous, like life working in geolocation, like that's all stripped out. Of these, of these accounts, of these transactions. Like, we don't have that data anymore. So great. I could, let's say I had 10 people doing 100 accounts a day and maybe they could do 10 more, you know, but if the, if the quality of the data that they have to do their review is, is less than, how are they even going to be able to make a decision? Err on the side of caution and decline it when it's an agent's a good customer. Like there's, there's a lot more going into like how our days were and how they need to be. Going forward in the, the new, for this type of transaction. Does that make sense? I don't know.

Yeah. And I think you, you really can't work the cases the same way. Uh, I, I, I think the first key element is how do I make sure I can stop it as, as early as possible in the journey. Mm-hmm. So on the edge when that agent comes in.

Love that.

So, you know, the, and we see a lot of protocols being put in place, you know, Visa is putting out, you know, Trusted Agent Protocol and there's a lot of different ones like that.

That's right.

That want to certify that that the agent coming in is an authorized agent and was authorized properly by the person behind it. The problem with that particular protocol is it certifies the agent coming in as a trusted agent, but it doesn't certify it wasn't hacked. It's not certifying that, that agent will not try to do something malicious. It just makes like an early determination that it received proper permission to go on your website and make a purchase, for example.

So that's like, that is a thing that I, yeah, sorry. That's the thing I've talked about is like, like I, I love the idea of these protocols, but you know, one protocol to rule them all is going to, what's going to have to end up being at the end of the day, because everybody's got their own idea of what it needs to be. But, you know, it's come up on another, another one of these before is when was the last time ChatGPT asked you to log back in? So if someone in my house, or let's say I left my computer at the library and someone goes into my ChatGPT and orders, you know, a bunch of stuff, yeah, it's gonna be signed still as a trusted agent, even though it's not me making that purchase. And that opens up that whole new vector. Like, friendly fraud, as we already know, is running away with itself right now. This makes it even easier to claim that at the end of the day.

Way, way easier. And I think relying on only on those trusted agent protocols is like relying on checking your identity on a credit file. Yeah, of course you will have a match, but if it's a fraudster, it's still not a person you want to let through. So, so that's where, you know, at Datadome we talk about intent. Once your agent's coming in, it might be a certified agent, might be an agent you think you can trust, right?

It's—

could be a ChatGPT, but can I trust what that particular agent is doing in that particular session? Is there something nefarious in what they're trying to attempt? And that comes down to analyzing not just what the agent declares, but also how it's actually behaving throughout the journey and eventually trying to pick up on, is there a new signal? Is there something or a pattern that it's giving me that indicates that, yes, it might be, you know, coming from Jordan's ChatGPT session, it might have been done with credentials being provided accordingly, but there's something going on that tells me that the transaction at the end will be a fraudulent transaction. And that's the new layer you need to consider. And an anti-fraud tool, you know, toolbox, you know, there's no silver bullet. There's not the one single technology you have to have in place. But I think now that we have those agents being a real concrete problem, it's just a new layer you need to add. And you need to stop them very early. Like I said, you need to stop them before they click create account. Because if they do click that, it's getting in your other pipeline and you don't want that.

So that's exactly, I mean, so I'm doing that right now at my own company is we're coming up with a strategy on, you know, agent-created accounts essentially at the end of the day, because, you know, it can be at scale and we, we're in a position where we like to talk about our user numbers and, you know, if anybody watched the last season of Succession, you know, where they talked about the user numbers on that account, half of them are bullshit. You know, you don't want your site to be in a position like that where you have to declare like, yeah, we have 800 million users a month, you know, but it's really 23 guys in a bot, you know, in a ChatGPT terminal at the end of the day. One of the things that, you know, I, I, I've talked about it a little bit, um, before, but I just was on a bunch of sessions with the, with like Visa, MasterCard, Stripe did one, uh, too, that was talking about their strategy and what they're doing for merchants. In the AI agent space, you know, and Stripe's obviously was talking about ACP, their protocol that they have and how they were trying to, like, they were open sourcing it so that like add-ins and like other and like checkouts and all those people could plug into it and, and use their sort of thing. Obviously they're saying it's free now. We all know, you know, it's always, it's always free for now. Look at Apple just changed yesterday on my Final Cut Pro subscription. Now I gotta pay for that every month, even though I thought I paid for it once. But the, I think that like, there's, there's, it's weird to see everybody's approach to it and how, like, if people are going the protocol side, some people are offering to send the prompts through for any chargeback mitigation, which I think, first of all, like, is crazy because I don't want my prompt being sent. Uh, and I think anybody that like finds out their prompt is being sent is gonna like that the LLM provider is gonna have a bad day. Um, but also like, You know, this is a funny thing is they always say, like, when I worked at Ticketmaster, the question was for fraudsters, like, when we got a fraud chargeback was, well, they used the tickets. Well, of course they were gonna use the tickets. Like, that was always the intent was to buy the tickets, either resell them or use them themselves and claim fraud. So, like, the use of the tickets doesn't indicate fraud or not. It, it's just, it's a non-factor on that. And like, the same thing is going to be like, it doesn't matter what the prompt is because you don't know who put the prompt in. There's not a photo of the person putting in it. So it doesn't prove that it wasn't fraud at the end of the day. So I think that like the companies that are doing that sort of thing need to rethink that particular strategy because I do think that like until we have the liability shift through these sorts of transactions, we're not going to hit like real nirvana. But like, once we do, and I think we'll get there, I do think we'll get there, is when we will start to really see it explode, especially from the merchant standpoint where we'll allow things. So go ahead.

Yeah, you're spot on, right? I think we have lots of conversations on, okay, how do we accept those agent transactions? When you talk to the big brands and you tell them, okay, what, like, where does the liability lie if you're, for example, testing that this transaction coming in is done by a trusted agent, you know, quote unquote, do you take the loss if it's then declared fraud? Is there a new chargeback code that they can use if, you know, on the issuer side they're receiving a claim and then it's tied to an agent transaction? And what you realize is that those big brands, they don't have it quite figured out. They want to be first on certifying an agent and facilitating those transactions. And you have, you know, Stripe making their, like you mentioned, their own protocol, but like, who takes the loss? Ultimately, everyone wants to build a highway and they want to say it's my highway and they want to benefit from it. Then you want to monetize it. Like you said, it's free today. It's never going to be free all the time.

Yeah, that's right.

We haven't figured out the structure of how the consumer ultimately that would suffer the consequences of a fraudulent attack. Like, how do we help those people? How do we make sure that the claim is processed correctly? How do we differentiate from abuse and from fraud? You know, if it's going through an agent, it might get much easier for people abusing to claim, you know, a refund or return or, you know, merchandise not as described. We already see it, right? You already see people using, you know, Gemini to edit a photo to claim a DoorDash credit because they're saying like my steak wasn't cooked properly, but it's really a deepfaked version of the original photo. So we know people will use it. For their own benefits, they're going to have tons of first-party fraud, opportunistic fraudsters, people looking at a TikTok or video on Instagram, following some new trend, thinking it's a life hack, quote unquote, when it actually is fraud, right? And it's another issue, right? Non-agents facilitate quite a bit these types of attempts because they give you the impression that you're not so connected to the merchant. You know, you're adding kind of an abstraction layer in the relationship. It makes people comfortable.

That is such a good thing is like, that's such a good call out because I've said for years that you wouldn't go in and take a shirt off the rack at your local, yeah, with JCPenney, bad example, and walk out the door with it. But like, people have no hesitation to issue a chargeback on a shirt that they bought online from the same retailer. But to your point, to do these things at scale, and that's what they're doing to brute force a refund or brute force some sort of like, um, rewards redemption. To your point, to have that terminal and that prompt in front of them, between them, they feel like they're not doing anything wrong at the end of the day. They feel like they're exploiting because that's, that's, that is the TikTok word, right? Is it's a hack. It's an exploit. You know, it's a glitch when it's not. It's Crime. It's actually crime.

Yeah. It's crime. We, we had it with the Chase money glitch. Was it like 2 years ago when people started depositing, you know, empty envelopes to withdraw money because the Chase money glitch allowed you to deposit and then withdraw committing felonies without even realizing it, thinking it was a hack. Um, so, so we know that people are, the problem is here, lots of people are getting access to technology that's increasingly more powerful. You know, just not just our phones, but now agents that can do stuff for us without having the proper literacy of the consequences of some of the things they might do. So yeah, it's going to trigger increased volumes of opportunistic fraud, which we're also not accustomed to. We're not necessarily good at detecting fraudulent intent from a real person. So, so we'll need to start looking at, okay, well, how is the agent behaving for maybe the real person, maybe for legitimate type of transaction and signing authorized agent. So maybe there's something there that indicates that it's a first-party fraud attempt. So yeah, it's, it's all about the, ultimately all about the intent and the way it behaves itself as it goes through your website or your organization.

I could see like, yeah, because basically we're going to have to now biometrically like monitor an agent at the end of the day, how we used to do people. Where are they clicking? Are they going to the account details page first? Like, what are they doing? But I could see, you know, like to, to go back to my thing and now this has just popped into my head right now. Is like on the refund, like a refund might be more, might be more helpful to see a prompt because it could say like refund this guitar that I just bought yesterday, but make me keep it and don't take no for an answer. Now, if I saw that come through on a prompt, you know, then I might know. And I think that might be a little more useful on that.

So, and here's the thing, right? As, so we think of people using it to try and break, you know, through our refund policies, for example, but you know, even one step before that, people will say, find me a guitar and make sure to pick it from the store that has the easiest return policy that I can then exploit. So your organization, like, these agents will look at your policies, what's being said about you on different forums, eventually identify you as an easy target, even for opportunistic fraudsters. You'll get more transactions, you might think you hit the jackpot, and then all of a sudden those agents again are handling the chatbot conversations, conversations with your agents, uh, in chat sessions to get those refunds and make sure they have the right script to really hit, uh, all the checks in your procedure.

And then they just blast it. Then they do it again and again and again. I had a, um, a client a couple of years ago that had their average sales on their website was like $2,000. And then one month they had $130,000. And the guy calls me, he's like, look, isn't it great? And I was like, let me see that traffic right now. I'm like, it's not great. I'm like, I'm like, did you ship any of this stuff right now? He's like, yeah, yeah, yeah. It was out the door. I'm like, I was like, oh, I was like, you just lost everything, like everything on that. Like, and it was to the point is like they found one transaction that worked and then they tried it two more times and then they just ripped it. They just like, it was like a full-blown hockey stick on that. And it was absolutely crazy to see that. And then. You know, to smaller brands, to non— like all of us enterprise merchants, if we see something really slide out of band, like we're like, ah, something's wrong here. But like to like to mom and pop stores, they just think like they're, they just, they went viral on TikTok, hit the jackpot when in reality they're about to lose their whole business at the end of the day.

Yeah, yeah, exactly. And you know, you're talking about that. It just pops it into my head. Like, right now we could probably have rules in place for, you know, if we have 3 accounts shipping something to the same house. So yes, in that case, you know, if, if Frosters are shipping, they're hitting a mom-and-pop shop, they're all trying to ship in the same place, someone will pick up. That doesn't make sense that, you know, 20 new buyers are shipping in the same place. But those agents, what they also allow is ship it to the billing address and then my agents will change the destination call UPS in a process that's not monitored either with the merchant directly or with the, the UPS or FedEx that's handling the delivery. So, so, you know, it, it also means that processes we have today, which rules that are well established and might work, well, we have to think about what would happen if at scale someone was trying to get to an agent and have them do a manual operation behind the well-structured systems that are looking at, let's say, shipping address, billing address, things like that. So, so yeah, it, it's really opening a hell of a lot of new threats that we have to keep an eye on.

So how can you help me not have to deal with any of these threats at Datadog?

Yeah.

Well, you have to deal with them.

I, I, I think it's unavoidable. I think, like I said earlier, that's a good point. Accepting agent traffic is like accepting transaction on the internet.

Uh, I think, yeah. And I think we need to call that out to everybody is You need to accept these transactions. You need to, like, you shouldn't think this isn't bad. These are your highest intent buyers. It's good traffic. It's just making sure you aren't that guy that gets ripped by the hockey stick level of, of increase overnight. Like it's going to be a gradual increase. I mean, it's going to be like a normal gradual increase that we would see, but these are transactions that you need to be taking. You just need to put a strategy in place to take them.

So, and I think that strategy can be broken down in like 3 main pillars, right? There, there's a technology layer.

You sound like Alexander Hall with your 3 pillars.

Oh, there you go. I was just talking to him yesterday.

Yeah. He always has 3 pillars.

He always has 3 pillars. Um, but the first one is technological. You need to have that layer at the edge. So the agent's coming in, you need to figure out what the intent of that agent is. So not just trust the signature that you're receiving, not just trust the the protocol in place to know if it's a trusted agent, but really track what that agent is doing all the way through the journey. So I think that's one key. It's unavoidable the same way that, you know, we have some basic infrastructure to prevent DDoS attacks and we have basic, you know, transactional fraud tools in place and we have identity verification. You now need to have agent intent verification and monitoring and making sure you don't have automated text Now that you have that though, that will technically most of the time reside outside with the fraud team. So there's an organizational transformation that needs to be operated, which I think should be led by the fraud teams, and it will be breaking down the silos. And we always say break down the silos, but between cybersecurity and fraud. And that means making sure that you understand which platforms are used for cybersecurity, including you know, eventually an agent intent detection platform to make sure that the fraud team has access to those new tools and they understand it and they have access to tools that are not meant just for cybersecurity teams, but are designed with fraud users in mind with the right forensics tools, investigation tools, case management, making sure that it can actually make sense of what's being presented. I think the third pillar is operational in the team itself. We have a need to educate our existing members, grow them, make sure that they step up their knowledge on that particular type of threat. You know, you know how it is, right? We probably have lots of teams that are relying on people that have 6, 7, 8, 15 years experience in the business. They don't get to know new stuff and learn about new fraud attacks and new technologies so much. So there's a need operationally to start looking at, okay, what's the skill set I have? Do I have people that can make sense of this? Do I have people that are more proficient with AI threats that can understand what bots are doing, what agents are doing, how it works. So you either grow your existing team and develop their skills, or you start bringing people with very, very specific skill sets to enrich the fraud team. Because I think the last thing you want to see is being completely blind, you know, going dark on that new threat because you don't have the right understanding, you don't have the connection with the cybersecurity team, and then they start doing what they think is best. But really with no context on fraud. They just do it from a pure cybersecurity perspective. So, you know, just to recap my 3 pillars, technology layer, organizational transformation to break down the silos and start working together, sharing information, even in some cases become one team, cybersecurity and fraud. Yeah. Fuse it together. And then operationally, making sure you have the right talent and the right skills, whether you develop the people you have or you bring new people in.

You see, um, like Target combine their fraud and cyber teams into one team, uh, which they did a couple years ago, which I thought was a brilliant move. Me at my company, uh, you know, I just posted yesterday a job for an abuse role. Uh, so if it's open, everybody that's listening to this and you want to apply, you can go ahead and apply. But, um, it is to your exact point too on there is to— we need to break down the silos. And that's what we've kind of done on my team is like, I don't want to own the cybersecurity piece. Like that cybersecurity team that we have is They do great, but there's crossover. And every 2 weeks I meet with my cybersecurity team and we talk about what we're working on, what they're working on. They're working on different types of things that I need to work on. But this is one where there's a crossover because, you know, they do the bot prevention right at the edge, to your point, all day long, right on our Cloudflare. They're looking for the bots, but they've seen these agents and they don't have the capacity to manage agent behavior. They just don't. But it does land, that piece of agent behavior does land more in my, in my wheelhouse at the end of the day at my company, because if they're trying to, like, I want to own everything that's trying to transact and I need to staff a person now that to your point on your point 3, which has that expertise to know and learn and experience. Cause like, I don't, I don't, I, me as Jordan, I can't do that every day. I need to say, and I don't know, I don't pretend to be an expert in this. I know a lot about agents, but I don't pretend to be an expert. I'm going to hire somebody that's an expert to come in, work with my team and work with that team and continually, and this is the thing, is continually update policy as these agents and these threats evolve. And that's going to be their whole job is not working on my fraud tools. I even put on the posting the other day, not looking for a fraud person. I'm specifically looking for a person that knows. This type of edge traffic to be able to sort to, to start to understand what we need to do and, and break down the silos. So I'll finally have a person sitting in the room with the cybersecurity team, but works for me, you know, my own human agent, if you will.

But I think you're a few months ahead from, you know, a lot of people. And, and we see that you mentioned Target, which is like the biggest move where you bring everything together. We see a lot of people doing exactly what you mentioned. Bring someone and make them sit with the other team. So at least you have like that liaison that's perpetually connecting the dots and bringing back the information. I think that might be like the quick fix for a lot of organizations. Just do exactly that. Just find someone that has that proficiency but fraud knowledge and background and just sit them with the other team.

I love that. So with Datadome, you guys are already working on a lot of this agent behavioral fingerprinting. You guys are an edge-based company, correct? Yeah.

I mean, we were in place, uh, you know, originally for bot threats, right? So automation, DDoS attacks, uh, you know, of scraping of the websites, undesirable scraping, right? Not, not your, your, your Google crawler, but so it was put in place for that. So because of our privileged position in a lot of organizations' tech stack, when we started seeing those agents come in, we said, What do we know from our own internal expertise, from what we're gathering? How can we start trying to process these new types of traffic and help our existing customers, you know, be certain that that new revenue stream will not become just a loss factor? So that's exactly what we do, you know, prevent from the very typical threats that you'd have— automated attacks, DDoS, scraping. But then we have that layer of Fraud prevention, making sure that you're— there's no agents coming in with the intent of creating a fraudulent account, with the intent of taking over an account, with the intent of making a fraudulent purchase, you know, under a guest kind of path. So completely at the edge, the minute someone hits your website, tries to get in, we're already looking at them, we're already trying to determine if that person should be let through, challenged, or just blocked altogether.

And like my team, like let's say I had this new person and they were working inside your tool. It's a pretty straightforward tool, right? Like you guys are pretty, like doing all the heavy lifting for us so they can, they can set the type of traffic that they want to have or the type of agents they want to have. Is it, is it that granular?

It, it can be that granular where you can build your own specific rules for specific types of agents and specific, you know, provenance, you know, geographically speaking. But the models we built, and there's tons of different models in place, even out of the box, they're so effective, the false positive rate is so low that we can hard decline the vast, vast, vast majority of the nefarious traffic that you're going to receive. So really, you don't need to do so much in the solution. The minute you plug it in, what comes out of the box already works day one. And then you can go and monitor what's being done, what's happening. Is there a spike in traffic? Is there something new or new trend that I should have an eye on? And then if you need to build specific rules. You go in the interface, build your rule, very, very user-friendly, right? All, all fraud companies will tell you, uh, you need to develop something about as easy as can be because your typical fraud user might be using Excel and Outlook in their day-to-day. So you want to have low-code, no-code, as easy as possible.

That's right.

So, so that's what we focus on, right? Like, and understanding who the user is, is, is part of the success of having, you know, a good anti-fraud and cybersecurity platform. Because they're not the same users with the same skill sets and the same level of, of, of IQ.

I do love that. And one of the things that's like, I love that it's happening, but it's like making me a little nervous is like all of like the fraud tools have all put in some natural language prompt to get like reports or to take an action or what am I supposed to be looking at here? Which I think is cool. It's really awesome, but it's going to take so much of the like, like our old school knowledge that we had to have to find these things, it's gonna rip it right out, you know? But that is the future, you know? Like, we had to get rid of the horse and buggy at some point, right?

So it's the future, but here's the problem, right? Like, yes, you can make it easier to, to, to write the prompt to get to the end result, but the reality is the data structure behind it, especially in like very heavy corporations, you know, take a telco, take a bank, there's so many systems with so many structures and table structures and architecture that you need to have that deep knowledge of what's happening in the backend. Where does the data reside? How is it put? How is it structured? Because ultimately you want explainability, you know, you want to be able to explain why is this number, the number that you're giving me for my losses, where is it coming from? So I think if you just have a bunch of people that are really good prompt engineers, assuming that the prompt itself. Doesn't make mistakes, doesn't, you know, emit—

Hallucinate a number.

Table or system. You, you still want someone that can explain exactly what's in it. So I don't think we can avoid having deep understanding of the backend data.

So this is really interesting. Now you guys have actually put together for your silos and all that we just talked about, you guys have actually put together a guide, a resource that like people can go and check out on their own for free, right?

Yeah. Yeah. It's a, it's a guide. A guide. It's called the Guide to Readying Your Business for Generative AI. Very, very simple. Says everything. But it's basically a guide that will run you through the different steps that your organization should reach when it comes to maturity to handle agentic traffic. So you already have, you know, like a checklist of things you can look at. Where am I today? Where's my organization like? What should I aim to be at, you know, in 3 months, 6 months, 12 months from now? So, so that's one very good report that we produce that I would recommend, you know, everyone have a quick read to try to assess where their organization is at. Today for Agentic AI.

So, everybody, I will slide up QR codes for those that are watching, and then there will be a link to that in the description box below for people that are just listening. It's the Agentic Commerce Guide, everybody. Go check it out. I'll probably have slid a QR code up a couple of times on there, and then I will slide it up one more time at the end as we wrap this up. Which, speaking of wrapping this up, I ask you and everybody, like, their big predictions for what's happening in the future. Since we are at the beginning of 2026, Give me your top 3 trends that are happening that you think for 2026, especially even if it relates to this or even if it's outside this, but, you know, probably will relate to this.

3 trends. There's gonna be so much happening. I think the main one is going to be, you know, deepfake-powered attacks. So we're gonna see a lot more of deepfake IDs, you know, deepfake avatars. Used to try and open accounts in organizations that would otherwise be left out of, of those agentic attempts. So banks, telcos, people that require some regulatory compliance, you know, iGaming, for example. I think we're at a point where it's getting so easy and cheap to generate, um, those types of IDs and, and visuals and videos that we're gonna start seeing it being used at scale. So that's gonna be one thing. It's gonna be pure subscription fraud issue, but in areas that are not too much feeling the threat of agent AI. So it's going to be one. The second one I think is going to be in internal fraud. So there's not going to be a lot of people being threatened by it, but it's also going to be leveraging, you know, the Geminis of this world where you can now generate, you know, receipts, you can edit receipts, you can, you know, defraud your own organization by trying to max out your expense reports, go through the limits of the policy, generate new attendees where they aren't there. I think that's going to be maybe a big threat. And the third one I would say is probably going to be on abuse, abuse of return policy, abuse of refund policy, because a lot of the time we ask for photo proof, right? Show me that the item is damaged. And now it's so easy to generate those photos that I think we're going to see on the opportunistic side of things, a lot of attacks, and there's going to be a big spike there.

We're already seeing that even at my company. And like, we don't have like, you know, big ticket items. It's small ticket stuff. And we're already getting the AI like damaged boxes and damaged bottles and things like that, which is super interesting. So anything you want to say to wrap this up here about Datadome, Agentic AI, we're all going to die because all the fraud's going to happen to all of us all the time. Anything about any of that?

Yeah, I'll say what I say to everybody, right? The best way to prepare is to talk about it. And, you know, I love to get in touch with people, understand what they're seeing, and try to like brainstorm on what could be and what they might be missing and what I see, you know, with other people I talk to. So, you know, I leave people with this, dude, get in touch, more than happy to have a chat. I want to talk about what you're seeing. If you don't think you have a threat today, I want to understand why you don't think you have a threat. That's always interesting to me, but yeah, let's, let's talk about it. Let's make sure we don't keep the information to our own organizations. I don't think fraud prevention is a competitive advantage at all. We should all be sharing what we know. So let's talk about it. And for people that want to talk about it with me, just get in touch.

Yeah. And I will obviously link him in the description below alongside the Agentic Commerce guide as well. Simon, I know you're a busy man and I want to thank you so much for your time. I know it's late where you're at too. So thank you for your time. Thank you for coming on my show. Simon Marchand, everybody.

Thanks, Jordan. Always a pleasure.

Like I said, everybody, I'll link everything in the description and we'll see you all in a couple of weeks.