[00:00:00] Anthony: So now you can ask the agent, Hey, what email will resonate with Phil and make him convert? It can go send those a hundred thousand emails to the simulated version of Phil and then come back with, here's the one that we think will perform the best because of all this, causality data, this agent context graph that we've built before. [00:00:17] you have to be snapshotting this over time. Where was Phil? What was the world doing? What, what did we try to send him? Did it work? [00:00:24] Did it do this thing? the earlier you start, the better, to have this foundation to actually do things better than let's just shotgun a bunch of emails at him over the next few weeks and burn our frequency caps and get him really angry. [00:00:36] And then, but maybe, oh, he engaged with that fifth one, but he engaged with an unsubscribe. Like that's not, that's not what we want to happen here. Right. [00:00:43] ​ [00:01:10] In this episode --- [00:01:10] Phil: What's up folks? Today we have the pleasure of sitting down with Anthony Rodeo, chief Data Strategy Officer at Growth Lube ~before Growth Lube, Anthony Head.~ Before Growth Loop, Anthony led all US marketing at AB InBev for brands like Budweiser, bud Light, ~Michelob Ultra, Stella Archis,~ and several others. [00:01:24] He's also a Harvard computer scientist. In a episode we cover how most marketing systems don't learn because they lack feedback loops and. We cover agent context graphs for drift detection ~and marketing systems.~ We'll also chat about the evolution of retail media networks and how they redefine targeting with governed data. [00:01:41] And Anthony also shares how agent to agent commerce operates inside our marketing funnels. All that, and a bunch more stuff after a quick word from two of awesome partners. [00:01:50] ​ [00:03:46] Phil: Anthony, thank you so much for your time, sir. Excited to chat. [00:03:50] Anthony: Uh, it's an honor to be here, Phil, uh, especially while the human perspective is still valued. You know, you with the trends we've been seeing, you might have to rebrand this to agents of MarTech soon.[00:04:00] [00:04:00] Phil: I hope not. Definitely hope not. [00:04:02] Anthony: Yeah. [00:04:03] Darrell: Uh, yeah. Good to have you on Anthony. Um, [00:04:05] 1. Journeying From Robotics to Modern Marketing Systems --- [00:04:17] Darrell: so you've been at Growth Loop for over six years and looks like you started at a customer facing role as Senior Solutions architect, and eventually you grew into chief customer Officer leading the entire function and for the last. Uh, but for the fast ~five, uh,~ three years, you've moved into a chief data strategy officer role where you get to work on, go to market with big cloud data providers and play at the leading frontier of AI product capabilities. [00:04:31] So why don't you tell us a little bit about your personal journey. Um, it sounds like you've loved solving data problems so much that you know for customers that this is something that you really decided to make a career out of. Um, tell us about us please. [00:04:44] Anthony: Yeah, it's, it's pretty funny when you think about, um, my background. I got my degree in computer science and I was, uh, you know, working with, uh, robots in college. I, uh, remember writing my first reinforcement learning agent and like my eyes lighting up. I'm like, man, this is like. Feels like it could be how the [00:05:00] human brain works or something like it. [00:05:02] And when you have this passion for technology and creating things and, uh, computer science, you know, the obvious next choice is to join a beer company like I did. Right? Um, so I didn't know what I wanted to build with all these cool tools that I had just learned. My dad was a carpenter. I looked at it as new tools in the toolbox. [00:05:19] Um, but I knew that I wanted to get, you know, to, to go ahead and like, learn about business and how I could apply and, and make value from this stuff. So I, I was thankful for the opportunities I had at AB and Bev. I was there for about six or seven years, but I was itching to get back to. Technology. And that's when I reconnected with David Josten, one of the co-founders at Growth Loop, who I had known from college and he was working on marketing technology. [00:05:44] And, um, for me, I'd kind of written like beer company in one corner and like AI and robotics in another corner on, on some axes, and said like, I wanna make sure the next steps is in that direction. So I, I said, let me come, uh, sweep floor for you guys, do whatever, whatever I can do at this, at this [00:06:00] eight, eight person company, you know, as we have a less than 1-year-old daughter at home. [00:06:04] Uh, a big jump, right? Um, but you know, they had some amazing customers. Google is an early customer, indeed is an early customer, and I saw what they were building with this technology that could plug into the data. And, um, I thought about the marketing I had done at AB InBev as a marketing director and this opportunity to actually bring the scientific method to marketing, right? [00:06:26] Uh, get more of the data involved in the audience creation, the journeys measurement. And drive optimization, uh, with some speed. The early, the early, uh, hints of what we call now the compound marketing engine, that was really exciting to me. And I thought, you know, if we have this foundation here, uh, eventually AI is going to play a role in this and I can help shape what that role looks like as we, as we go and build this stuff together. [00:06:52] So that's kind of how, um, how I got to growth loop Darryl. And then I think once I was there. You know, I, I got, I had the opportunity to work with you, uh, [00:07:00] when you were at Indeed. Right. Which was amazing. And, um, when, when GPT-3 came out, that's when it finally started to click that some of these jobs to be done that we were helping marketers with, for building audiences, for example, um, if you looked at what GPT-3 was really good at, in the early days, it was like translation and summarization. [00:07:22] And we looked at audience building and said, this is kind of a translation problem. Um, it's a business concept of something like, I want to target customers in the tri-state area at risk of churn and translating that into this SQL query on top of the data to actually surface those customers. Right. And, and our customers were like throwing Jira tickets back and forth. [00:07:42] And this process would take weeks or months, or take a really long time to get these campaigns launched. And we had already kind of sped that up by plugging directly into the data and letting people. Build these, these artifacts with a ui. And it started to work when we started doing initial, uh, experiments in this area in late 2020, [00:08:00] early 2021 to actually just ask the AI to kind of build these, these audiences for us. [00:08:05] And, um, there were a couple of conversations that I had that gave me a lot of, um, confidence that we were headed in the right direction here. I'll tell one story 'cause I think it's pretty funny. Um, a, a mutual friend had texted me and said. Uh, do you know Sam? And I said, Sam, who? And she said, Sam Altman. [00:08:26] And this was in, this was in summer of 2022. I said, no, of course, I don't know Sam Altman. You know, like I have a big open a I fan, and like I'm following what they're doing. So she said, he is gonna call you tonight. And like, he gave me a call and we started talking about a bunch of different things. I'd had the side project going on where. [00:08:45] I had, uh, you know, what, what they now call chain of thought before it was called chain of thought for AI that would participate in chat via Slack and you could talk to it. It had some long-term memory and he said, oh yeah, we're, you know, we're working on this chat, this chat thing too. It [00:09:00] would, you know, obviously later would, would become chat GPT. [00:09:03] But I, through building that relationship, I ended up showing him what we had built with this natural language audience building, and he said, this is super cool. He got excited about it and then a few months later. Uh, Google invited us out to, to Sunnyvale and I sat down with some leadership there. We had showed them that, and then in 2023 at Google Cloud next we stood on stage with them and launched what they called degenerative marketing category, which was, you know, that, that to me was, okay, now, you know, go back to my two axes of [00:09:30] beer [00:09:30] Phil: Mm-hmm. [00:09:30] Anthony: to ai. [00:09:31] Now, like I feel like we have, uh, really good partner feedback and our customers are starting to see the value in. AI can help speed this thing up even more and maybe even more than an efficiency play, eventually drive top line compounding outcomes too. And that, that to me is when we kind of leaned all in. [00:09:49] And that's when my role became this, you know, made up title Chief Data Strategy Officer that focuses on a bunch of different things to try to make that reality, uh, valuable to our [00:10:00] customers and, and the market. So, long-winded answer, but it, it kind of was a long-winded experience over. You know, seven years at a BI and then, and then, uh, about six years so far here, here at Growth Loop. [00:10:11] Phil: No big deal. A little, uh, phone call conversation with, uh, Sam Altman in 2022. Probably like, yeah, yeah, this guy's a big deal. And you flash forward just a couple years, like it feels like, you know, two decades later that we've been with ai, but it's only been like [00:10:26] less than a half a decade that we've gotten the chance to play with GPD three and, and have that like chat ui. The speed of the, the, the pace of this industry has been super crazy and I love how you, like, you started that growth loop in a customer role, like you said, like, I'll, I'll do anything. Like, I just want to join the company. Um, but like, it's such a cool entry into the company because you had what, like probably hundreds of conversations with customers and I'm curious to ask you like. [00:10:53] Over the course of all of those conversations with customers, like over the past six years, I'm sure you've heard a lot of [00:11:00] folks complain about MarTech, complain about their tech, complain about silos of data and whatever. [00:11:05] 2. Most Marketing Systems Don’t Learn Because They Lack Feedback Loops --- [00:11:05] Phil: But you've actually said that the core failure in modern marketing is this idea of feedback, and so teams like are generating activity after they launch a campaign. [00:11:14] Data comes in, reports are populated, but very little actually. Compounds from that. There's no feedback loop to better inform the next thing that we do. The next campaign, let alone like shared broadly within the company. I still worked with plenty of companies that like, we're still in campaign mode, so to speak, like sales are low. [00:11:31] Let's launch a new sales campaign. Let's pull up a report to see how we did optimize it a little bit. Next month. Rinse and repeat, right? Like why have marketing systems historically? Optimized for execution instead of learning. And how, how do you think, like we get out of this, like why has this been a thing [00:11:49] Anthony: Yeah, I think, um, there's a lot to unpack in there. There are, so one is, um, there's a lot of short term pressures, right? Especially from these large enterprise [00:12:00] organizations that we work with, where, you know, they're reporting results quarterly. The average, uh, CMO tenure. I know I used to, I used to say this like when I was in the CU chief customer officer role, it was, you know, it was like 18 months then I think it's reduced since then. [00:12:13] Right. There's, there's pressure to perform, right? And, um, this idea of this long term, when you think about compound interest in, in finance, right? And like compound returns, like the initial phases of that are look kind of linear when you zoom out, right? And it's like. Folks are looking for those big bumps now, uh, which might not end up driving things that compound in the long term. [00:12:36] So one is just. Structure and the nature of these companies needing to show results quickly and publicly. That's one, um, one pressure. I think the other one is the, the technology and the challenges of the, the nature of the setups of a lot of these marketing technology stacks. Right? And there's a ton of silos. [00:12:57] That's the key, the key thing that [00:13:00] we, um. I think saw in the early days, right, is you have the marketing teams here, you have the data teams sitting over here, and if you do have your data in, in good shape where you know a lot about your customers, the transactions that you've had with them, the engagements, are they opening your messaging? [00:13:18] Are they interacting with your ads? Like even if you have that foundation and you've broken down the silos of like all these point systems that manage the communications, even if you have that. You had folks throwing Jira tickets over the wall, right? It's like, Hey, can you give me this, uh, audience that looks like folks who have done X, Y, and Z that's aligned to this campaign brief? [00:13:38] And then it would take, you know. A few weeks to like, for someone to interpret that, write some sql, throw it back over the wall. Oh, the audience isn't big enough, it's not enough revenue touched for to drive my short term results that I need to drive, throw it back over the wall. Right. In this iteration. [00:13:53] So like when you think about compounding like tight feedback loops on short time cycles with good [00:14:00] data, like are needed to actually drive that, right? So like step one is just. How much can you close that feedback loop, um, between the teams and the, you know, the messages that are going out to the marketer, the actual responses that they have to the interventions that you're putting out there and, um, learning and implementing the learning, right? [00:14:18] What do we double down on? What do we divest in? And that, um, in the MarTech world has, has historically been a really long cycle. And it's not like, um, the technology that was there wasn't good, right? When you look at the Adobes and the sales forces of the world, like they were built to solve a problem in a different time, right? [00:14:38] Like there wasn't this, um, this data cloud in BigQuery or in Snowflake or you know, in Databricks where you're like building up this source of truth for your customer data. So like Salesforce had to be that for a lot of companies and Adobe started to aggregate that for a lot of companies, right? Now when, you know, when we go into companies and they're kind of [00:15:00] like, Hey, here's the giant customer data repository that I've built in this amazing source of truth that my data team has invested in. [00:15:08] And we're gonna put it through a little straw, some of some of it through a little straw, and then have to map it to these fields in this other system before we can even do anything with it. How can you even hope to have a tight feedback loop and experimentation and compounding? So that was, that was the initial problem. [00:15:25] I think we, we set out to try to help with, uh, there Phil. [00:15:32] Darrell: Totally. Totally. Yeah. I, uh, I, I, I like your answer. I agree with it. I think that for me, a lot of times what I see too is, you know, companies have this competing incentives problem, like you said, short term. There's also. You know, it wants to cut costs and simplify systems and marketing and sales want to move [00:16:00] faster and, and activate more touch points. [00:16:03] So that, that competing problem, I think makes it really challenging. The, the competing incentives make it really challenging. And then [00:16:10] 3. The Martech Engineering Talent Gap --- [00:16:10] Darrell: honestly, I think there's like a talent problem too. And I think, I think a lot of the. You know, technical talent gets, that gets thrown on these problems. Maybe just aren't really interested in, you know, what works for MarTech today or like how single source of truth works. [00:16:29] You know, I always have this, this theory, you know, and Phil, you can cut this if you want, but I always have this theory that the MarTech teams get like the B talent [00:16:41] Phil: Hmm. [00:16:42] Darrell: B and C talent. [00:16:43] Phil: Of engineering, you [00:16:44] mean? [00:16:45] Darrell: Of engineering and of data? [00:16:47] Phil: Because no engineer [00:16:48] wakes up and wants to work on marketing. They want to build product. [00:16:51] Darrell: there are a few but, but there're few and far between and most, I think data, people want like, I don't know, analyze [00:17:00] customer trends or something and inform, like building the product and so do engineers. And so I feel like you get this group of people that you know, honestly probably never worked on a MarTech project before. [00:17:11] You know, and that's why, and that's why it was so great to work with Anthony and his his team as well, because they'd seen this stuff before and they knew what they were talking about. So I think that that that also causes a lot of the issues [00:17:24] that you can't really get to where you want to go. [00:17:25] Anthony: I think that's spot on Darryl. And I think, uh, you know, without revealing specifics and things like that, but 'cause it was, uh, it happened when I worked with you and it's happened with many other customers, is like, these teams are at odds, right? And like you have, um, you have the, you know, the tech team that maybe wants to build a bunch of things. [00:17:42] You have the marketing team that maybe want, sees a solution that makes a lot of sense for 'em and just wants to go and move fast. Um, you have the MarTech team that's kind of caught in the middle. Um, and when. Uh, what I, what we found and we found this, uh, with you at Indeed as well, is like when you actually get those folks in the room, you [00:18:00] know, you just talked about the marketing team wants to move fast, the tech team wants to efficiency. [00:18:04] If you go to any tech team, like no one, none of them are gonna tell you they wanna move slow. And if you go to, and if you go to the marketing teams, none of them are gonna, are gonna tell you that bottom line and efficiency isn't important for the business. It's just you have to un, you have to try to unlock that and get folks in the same room and tell the data teams, Hey, all this work that you've been doing, we're gonna make your work a revenue center for the company, not a cost center anymore. [00:18:26] Like, we're gonna be able to align revenue and engagements and all these things that are important for the business. Back to all this work that you've been doing on your. Amazing identity resolution algorithm and model that you've built internally or, um, the data ingestion efficiencies that you've built for like your pipes that suck all the data in, right? [00:18:44] Like the machine learning models to, to tell you propensity for a customer to take X, Y, and Z. We wanna get those off the shelf and like empower your marketing teams to use this data directly, right? Like it's a, there's a huge people element, uh, in this and. [00:19:00] That's the change management piece where you talked about like sometimes it feels like the market, the MarTech teams get the be the B team folks from the engineering side. [00:19:08] What we've had to look for is like who is the maverick or the change agent within these companies and like who wants to change that paradigm? And you kind of go latch onto them and you find, like we talk about an ICP, which persona is the right one? Like, and we have definitions of those. But then every, every now and then you'll get in the room with someone and it'll click and they'll be like, oh, Darrell wants, Darrell wants to change things here, and we think we can go help him do that. [00:19:30] And that's the, that's the biggest piece is like, how do you break that momentum or, or start that momentum in a new direction. Uh, and technology helps, but it's a big people exercise too, as you say. Darryl. [00:19:41] Darrell: Yeah, totally. [00:19:43] Phil: it's what we are called the humans of MarTech. At the end of the day, it's not [00:19:46] all [00:19:47] Anthony: I love it. [00:19:47] Phil: The humans obviously play a role in, in all of this. Uh, [00:19:51] 4. AI Will Amplify Whoever Has the Cleanest Causal Feedback Loop --- [00:19:51] Phil: Anthony, can, can you maybe like, unpack that, the feedback loop thing a little [00:19:55] bit more? Like, um, the, the one thing that I was like reading when you were sharing that [00:20:00] like. [00:20:01] The cleanest loop is, is something that you wrote, right? Like you, you argued that like AI won't automate marketing, but it's going to amplify whoever has the cleanest feedback loop. And I'm curious what you mean by that exactly. Like is it hypothesis driven, controlled experiments? What are all like the required pieces of the system for it to learn properly? [00:20:20] Like are you talking about besian inference, like proxy metrics validated over time? Like walk us through that. [00:20:26] Anthony: I I'm amazing. You're going very deep right now. So, um, let me, let me walk through this and pause anytime with questions if we need to, uh, spell something out or if I'm not being clear. 'cause I'll, I'll try to use some analogies that may or may not work. So we'll test some things. I'll, I'll start by cheating and I'll use someone else's analogy. [00:20:42] So, um, I heard Mark Andreesen on Lenny's podcast. Uh, talking about alchemy of all things, right? And like when you go back to Isaac Newton, uh, you know, brilliant scientist like his, did amazing things to change the world. He was also obsessed with the concept of like turning lead [00:21:00] into gold and alchemy, right through the philosopher stone. [00:21:03] If you read about him, you read any of his biographies, you read some of the Neil Stevenson books that deal with fiction of his time. Like it's in there as a common theme. Um. Um, the analogy that Mark Andreessen used is that, you know, we kind of have alchemy now with ai, right? Where we're turning this abundant, uh, substance sand into thoughts, um, with ai, right? [00:21:26] And if you take it in the limit, like I'll say, okay, we'll go beyond sand and we'll call it energy and like, what's the biggest source of energy that we have? This giant fusion reactor in the sky, we call the sun. So we're turning sunlight into thoughts in action is like where this is headed. Right. And, um, when you think about it as turning sunlight into thoughts and action, right? [00:21:48] Like this whole paradigm of what it means to actually like, do work changes. And if we, if we t take it back from the limit to now in 2026, and we think about what AI is doing [00:22:00] well, unlike long running tasks, we think a lot about coding. We think about math in some senses, right? Science, it's starting to do science a little bit. [00:22:09] I still hear engineers when they look at what, like Claude code or, or, um, cursor or even, uh, like molt now do, and they like look at the, the thought process in between those. Oh, this was so inefficient. It burned 40,000 tokens going all over the place before it got here. And you know. Do we care as much anymore? [00:22:28] Right? Like if I go to sleep at night and an optimist robot has my dishes done in five minutes, or five hours, if I wake up and they're done, like what does it matter? Right? Like you can, you could spin up, you could spin up 20 agents working on 10 tasks when you go to sleep, and as long as they have tight definitions of done specific tests, they have to pass like. [00:22:46] If they wander and meander all night long, as long as that work's done in the morning, like that's incredible. And this idea of like having slack like in the work process of these agents, I think is underappreciated right now. Um, but it does require [00:23:00] these really good definitions of done and specs. So 2026 is gonna be the year of, um. [00:23:06] Huge, huge progress and many tasks falling within what we call verifiable domains. Is this something you've heard of before? Verifiable domains. So this is what we're talking about with, with coding, with games, with math, you can test the outcome, right? Like. You can go ahead and have the thing write code and then see if that code runs and passes the tests that were there with games, right? [00:23:27] Like go and alpha go that Google solved years ago. It's like the games have rules that are strict and it's bounded and you can go and see if the rule is valid and like play the game out and simulation and see if it gets to an end state where someone won, right? It's a very specific rules of what, what the definition of done is and, and winning or or achieving the outcome you have. [00:23:47] Now, the problem with marketing. Hey, go write an email that will increase the response and, and like conversion rate of this customer. It is not a verifiable task. Um, [00:24:00] now there's two things you can do to try to make it. One, one is you can let it actually send the emails and go ahead and like blast fill with a million different things. [00:24:10] And then, oh, but we had one that got, no, this is a very, this is a very bad idea, right? Like, uh. And there are actually some like AI decisioning, um, modalities that are starting to do this. They have these long explore phases and yeah, you can bound it with the options that you give it, but you're still kind of like just firing off a bunch of random stuff and learning online. [00:24:30] So that's not necessarily the best approach for folks who care about their brand, their relationship with their customers, things like that. So that the way we are trying to make this a verifiable domain is through what we call. Causality data or our agent context graph. And what this is, is think about the state that you're in right now and everything a company might know about you. [00:24:54] Um, so you start snapshotting, Hey, this is everything we know about Phil right now. [00:25:00] Here's the intervention or message that we tried with him today. And here's the outcome. Here's the actual uplift in whatever KPI that we care about or set of KPIs that we care about. And you do that constantly with every experiment that you run. [00:25:14] And you build this context graph over time. That's not just a, a snapshot of the data we have today, but every single, at every single point in time. Where was Phil? What was the world doing? What, what did we try to send him? Did it work? Did it drive this purchase in this, you know, this category that we care about? [00:25:30] Did it do this thing? And when you have that over time. You can answer what are called counterfactual questions. You can answer the what if questions. What if we send Phil an email that looks like this, his state right now, like, and given the the types of content he's received before, we can say with confidence what will happen in the future. [00:25:49] And this is better than just doing this on observational data with machine learning models and classifiers. This is truly based on this causality data. So now you can ask the agent, [00:26:00] Hey, what email will resonate with Phil and make him convert? It can go send those a hundred thousand emails to the simulated version of Phil overnight and then come back with, here's the one that we think will perform the best because of all this, uh, causality data, this agent context graph that we've built before. [00:26:18] So that's the. When I say the cleanest feedback loop, it's like you have to be snapshotting this over time. And like the earlier you start, the better, um, to have this foundation to actually do things better than let's just shotgun a bunch of emails at him over the next few weeks and burn our frequency caps and get him really angry. [00:26:36] And then, but maybe, oh, he engaged with that fifth one, but he engaged with an unsubscribe. Like that's not, that's not what we want to happen here. Right. [00:26:44] ​ [00:28:47] Phil: that's the problem I've had with propensity models for email specifically, like to stick to your example is that even though something did, uh, like an uplift for a certain cord of people in the [00:29:00] past. Does not mean that it's still going to happen in the future. [00:29:04] Those are different users. We're sending this at a different time. There's different stuff going on in the world. Like we're missing that context that you're, you're saying it is part, is part of this like contextual, what'd you call it? Like, uh, the context. [00:29:16] Anthony: The context graph. Yeah, the, [00:29:17] 4.2 Agent Context Graphs for Drift Detection in Marketing Systems --- [00:29:17] Anthony: the agent context graph. And I think, uh, you know, you'll, I think you're gonna have, uh, my colleague Toby on at some point, so I won't spoil too much of probably what he'll talk about. But the, the other thing this allows you to do is kind of monitor for drift, right? Like, if you have an experiment, you have a treatment group in a control group, you wanna scale to a hundred percent, but you still wanna measure, you can measure using this simulated. [00:29:39] Set of customers from the causality data, but the, but you're constantly grounding it in the real transaction data. 'cause again, you're sitting on top of the data warehouse where all your transactions are, [00:29:48] Phil: Mm-hmm. [00:29:48] Anthony: so you can see if there's drift. Like you don't have to assume that your assumptions back then were real, were, were still true. [00:29:54] Right? The world changes, the market changes, the response to your products change. All, all these things [00:30:00] change. So you get this kind of, uh, built-in early warning system for our, my are my assumptions, um, still strong. So that's, you know, it's some of the things we try to do to box in this very unbounded problem of. [00:30:12] Marketing, which reduces to human communication, right? It's like, can we, can we make it such that, uh, you know, Chris sell, one of our co-founders used to say marketing should eventually feel like, you know, grabbing a coffee with an old friend. Like, and that's how the brand should be talking to you, right? [00:30:28] It's like you have all the background about that person. You don't, you don't need to start from scratch and like, you're not always trying to sell them something, right? Like maybe you're just saying thank you and, uh, giving them an offer. Like it, you know, it's not necessarily, um. I think it's gonna be very different, right? [00:30:43] When we have intelligences that are increasing exponentially that you can deploy to go talk to Phil and just be like, Hey, like if Phil's not interested today, maybe we back off today. And then when you go back to him tomorrow, right? It's like it can become much more human. Where, um, you know, the company could [00:31:00] focus much more on providing the products and services, building the products and services that will be helpful to, to fill to Darrell. [00:31:07] Um, rather than like, what is the next best email for me to send to a. 10, 10,000 people to optimize for the aggregate outcome of the group. Right. It, it, it should become much more personal if we're successful and inboxes should become much better. And inboxes of course will eventually be filtered by agents on your end too. [00:31:26] And the, the concept of the inbox, how long will that be around? There's, there's all kinds of fun places where this goes, but the foundation of it is if you, if you've been snapshotting this data, if you have this agent, agent to context graph like. You could be much more personal with your, your prospects and your customers. [00:31:44] Darrell: Absolutely. Absolutely. So Anthony, let's switch gears a little bit and talk about, um. You know, you've been, [00:31:51] 5. Humans Will Set Hypotheses, AI Will Accelerates Iteration --- [00:31:51] Darrell: you've been said to say humans will put forward the hypotheses and AI will lead the iteration acceleration. Um, but there's always been times where, you know, we've been through scenarios where companies have changed their goals abruptly, mid-cycle, you know, a new C-level executive steps in and overrides everything that. [00:32:13] We've been hap happening any, any learnings that we've had. When do you think that AI can also help stabilize decision making versus just assisting? [00:32:23] Anthony: Uh, this is a good one. This is a really good one, Darrel. So, um. Whimsical decisions are much more prevalent when results are bad. Yeah. Or can we agree? Can we agree on that one? So, um, when things are performing really well. [00:32:39] You'll have a lot less of a tendency from C-suite, from leadership of the company say, let's rip everything out and try something new or completely change direction, right? So one is like if you have a system that's driving results, um, that makes like whimsical big changes in direction [00:33:00] much harder. [00:33:00] Unless you see external threat of disruption from somewhere and you need to go on the go on the offensive, right? Um, so. I think the closest analogy I have here, and I'm sure I'm not the, I know I'm not the first to ever make this is like programmatic media, right? It's like, yeah, you probably had like a long time ago people saying um, Hey, this placement on this website is super important and like getting into the nitty gritty details of, Hey, we want to specifically define where our ads are showing across these different publishers and content and things like that. [00:33:33] Now with a lot of like. Programmatic results. Driving up into the right is, that's a piece that's a lot less touched, right. And you just kind of make sure it has the data it needs, the conversion data, all the right things to like make good decisions, and you kind of let it go. You'll see that type of, uh, adopt, uh, that type of response in different areas of marketing as the automation actually proves results now until it does like, yeah, like trust me, I'm the first one [00:34:00] to have seen firsthand that. [00:34:02] Our initial version of natural language audience building that I talked about [00:34:05] Phil: Mm-hmm. [00:34:06] Anthony: like looking back, like I, I said it, I tried to say it then, and I don't think it resonated with people. I said, this is the dumbest this will ever be. And it truly was like probably 50% of the time it would generate things that were not so useful. [00:34:18] Right. And now as we have, oh, this, like a supervisor swarm model of these long running agents and AI studio that do all these amazing things, like it's gotten, uh, it's gotten pretty remarkable for, um. For some of the, uh, some of the customers that are, that are driving some pretty big outcomes with it. But, um, these things kind of like if they don't perform, like everyone is going to have this, uh, this hangover effect. [00:34:45] We saw it last year a lot in late 24, early 25. It's really hard to go from pilot and prototype to production, right? It's really easy to make this pilot and prototype work well. A lot of people were sold technology that then [00:35:00] didn't actually work and practice didn't scale. So there was this kind of AI hangover. [00:35:04] Um, but you're seeing in these verifiable domains now, um, things driving a lot more value. Um. And a lot more tangible value for the enterprise. And that's, that's a place where I think results Darryl really make, uh, the conversation shift from what do we need to do differently to how do we, uh, continue down this path with more robustness, with more data, with more information? [00:35:31] What are the new hypotheses we should test? Right? What are the new things we should try? Within the system, rather than saying, maybe the system doesn't work, let's go shift to something else. Let's go rip out this technology to plug in a new one. Um, so I think that's, I think that's an important part of it. [00:35:46] Obviously results, you know, results don't lie. [00:35:49] Phil: Anthony, [00:35:50] 6. The Evolution of Retail Media Networks --- [00:35:50] Phil: do you think retail is one of those [00:35:52] verifiable domains or is increasingly becoming more likely to be verifiable? Like we, we [00:36:00] talked about, uh, recently, like marketing as a whole is like a bit tricky to put in that box, but do you think retail specifically like media networks is a bit more in that box? [00:36:09] Anthony: Yeah. So it's, uh, when I think about retail and how like verifiable, like predicting a purchase for example, is, um, I've seen this long term trend, like when I was at AB InBev. Um. I was a marketer, marketing director, but then part of my time as a marketing director was in experiential marketing. And then I went and ran [00:36:29] Phil: and stuff [00:36:30] like [00:36:30] that? [00:36:30] Anthony: yeah, festivals, like all the chef stuff that we did for all the major brands that you've heard of, right? [00:36:36] For Budweiser, Stella and Michelob Ultra, bud Light, all those things. It was a crazy 18 months of my life. I was appreciative of the opportunity. I never, never want to go, never want to go through five days a week, you know, working hard at the desk and then two days of, you know, schmoozing on the weekend at the festival. [00:36:53] I was very lucky to go to a bunch of these things, but, you know, there was a time when I just never wanted to see a, a festival or a [00:37:00] concert again. Um, but it was, it was, it was great to, uh, to be a part of that. But, you know, then I went and ran our retail business and retail, um, retail margins have been under. [00:37:12] Pressure for a long time, especially brick and primarily brick and mortar businesses. When I inherited that business, it was, you know. Minus eight or 9% EBITDA per year was losing money. Um, it was, uh, pretty flat in terms of growth. The quality scores were really high, so we, you know, we li loved that. We said, okay, people are coming in trying our craft beers and then go buying them, going and buying them somewhere else. [00:37:39] But my task was very simple, is make it grow, make it profitable, and don't screw up the quality scores. Right. To get to, you know, uh, you know, high single digit positive, uh, margins before I gave it back was like a, it was a, a pretty big deal at the time for the company in a, in an industry where, you know, retail margins were 2, 3, 4, [00:38:00] 5%. [00:38:01] So that's, so that's where I think, um, like retail pressure has been something that's important and close to me for at least a decade. Um, but with the experiential marketing. Side of things. I've seen this shift for a long time where it's kind of, you have this continuum of like, commodity products to like high experience products, high end products. [00:38:22] It's, it seemed like the middle's kind of been moving out in both directions. Um, like a lot of things that are purchased cyclically that you can just like get on Amazon or get on, you know, from an e-commerce website to like the things that you want to go touch and feel and shop for. And experience yourself, and it feels like as more things move into this kind of, hey. [00:38:47] If I just know that I need, uh, toilet paper and deodorant and like some of these products that like, I know what our t-shirt, we were talking about black t-shirts before we started recording here, right? It's like I buy these black t-shirts at like probably [00:39:00] 10 at a time and like, yeah. And in that, in that place, I think that is a very predictable, verifiable realm, right? [00:39:08] Where you can just say like. In the future, it'll be, Hey, AI agent, go get me the best deal on shirts that meet these specs. And I don't really, I don't really care as much to think about it at the other end. Uh, the value of, you know, human engagement, being there in person, like touching and feeling things like the value of those things will go up, um, over time. [00:39:27] And that I think is a little bit harder to predict. Um, it maybe it won't be forever, but, um, I think there is this bifurcation of kind of commodity and, and experiential, but. Uh, but in retail, you know, this, this whole thing, um, about the margin contraction has been a, a long running trend and it's hard to kind of, um, see a way that that margin for actual purchases doesn't keep trending in that direction as so much of the production and things get automated and like technology and things like that, is you're, you [00:40:00] removing. [00:40:01] You're removing opportunities for kind of the, the gap in the margin between the production and like the end state, right? And, and that's, that's a good thing for consumers. Like, don't get me wrong, it's a good thing for consumers, and that's a tough thing for retailers. So retailers have turned to this concept of media networks, um, as a, a great potential, uh, source of both growth and profitable growth. [00:40:25] When you think about this, uh, this concept of a retail network, what it means is essentially okay. Um, AB InBev, you know, I, I, you know, I'm selling beer. I, I don't sell that beer directly to you. I sell that to a distributor who sells it to a retailer, who sells it to a consumer. Sometimes there's even an extra tier in there. [00:40:42] Um, so it's very valuable for, um, for AB InBev to go ahead and say, Hey, Mr. Retailer, would it be possible for me to target the people who have bought Budweiser in your store over the last 12 months directly? And the retailers have [00:41:00] realized, oh, I can sell access to that audience and do it in a governed way that respects the consumer rights and everything they've opt into and the local laws and everything else, but they can make 70, 80% margins on that. [00:41:11] They, you know, now you're selling data access. And, um, that's something that, uh, if you look at the retail media, we call it retail media and retail, it's commerce media more. Uh, you know, when you abstract a level and talk about, you know, all the different providers that can do this, um. It's a high growth, high margin business, and it is the darling of, um, especially the retail industry right now because of, um. [00:41:37] The new growth and profitable growth that it's infused into, uh, into the market. And I rem I remember, you know, I, I used that AB InBev, uh, analogy. I should have used one that was real for me. I ran our, I ran our sponsorships with our sports, [00:41:51] Phil: Mm-hmm. [00:41:52] Anthony: our, our sports partners for a few years. Actually it was a few months, it was before we brought in the guy who was really good at this. [00:41:59] And then, uh, they [00:42:00] kicked me over to music and food. They're like, you go do this, this stuff that, that's over there. Um. But I was talking to nascar. I, you know, I went to Daytona 500. It was amazing. I was in the pits, like it was, it was a crazy experience. And we got back and it was like, okay, we're renegotiating the sponsorship. [00:42:16] And, uh, we were looking to reduce our amount that we were spending. But the, um, the thing I wanted most was to target NASCAR fans who had. Certain beer buying, uh, behaviors, and you couldn't, you couldn't do that a decade ago. They're like, well, we'd love to sell it to you. And like, but we can't. And like I was like, I'll pay you more for the sponsorship if we can do it. [00:42:38] And it just wasn't possible. So to see it now possible to be building technologies that help make this possible for our customers across industries is pretty, is is pretty remarkable to me to kind of take that full circle. It's the same, uh, you know, it, it's very similar challenges. It's, Hey, the data's over here. [00:42:57] We wanna make it available to our vendors. [00:43:00] Um, we want to be able to like, build audiences quickly and make them available to these folks who wanna buy the audiences and get access to them. And it's even, it's even more urgent than marketing in some cases because. If you don't make them, those audiences available quickly, like those vendors can go to other retailers or other folks and get access to similar audiences. [00:43:20] So one is like speed to value is incredibly important for like time to first dollar. That's one of the key metrics for them. The other one is, uh, you know, can you measure what happened? Like did folks actually change their purchase behavior in your store and provide that to these folks? 'cause that's incredibly valuable to them. [00:43:36] One of the. One of the largest regional retailers in the country. I was speaking to 'em a few days ago. They said they talked to their three top vendors and they said they won't buy an audience at all if they can't, uh, if they can't provide back measurement in terms of what's happening. So again, having this feedback loop, having this, uh, incrementality measurement, having this agent context graphic, all of these things are incredibly important to be able to say, [00:44:00] if you do this with this audience, you can expect these returns. [00:44:03] And that's. That's a huge edge we haven't seen a lot of folks talking about, um, in the market today is, you know, if you're a retailer and you're snapshotting this data over time now, or yesterday or last year, like some of our customers have been doing, you can do this now. Right? And like you can get ahead of it and you can, you can predict much with much. [00:44:22] Much better accuracy what might happen, which makes your audiences more valuable, it makes the communications that go out more valuable to the customer, right? It ends up being a win-win on both sides. And that's, you know, that's what we're all here to try to do, is to kind of like nudge these, these technologies in a trajectory that's a little bit more helpful for, uh, for humanity. [00:44:41] I, I think that would be a nice, nice outcome here. [00:44:43] Phil: Yeah. No, it's such a cool space in the industry. Like I, I was telling you before we press record that I knew you're like super passionate about this space and I spent a lot of time ahead of our interview just getting more familiar with retail media networks and commerce media networks. So for like the folks [00:45:00] that aren't like super into retail, like you've had the pleasure of being for, for most of your career. [00:45:04] I wanna make sure I have this [00:45:05] right. So, [00:45:06] um, [00:45:07] we're like. [00:45:07] 7. How Commerce Networks Redefine Targeting With Governed Data --- [00:45:07] Phil: Like retail networks and, and commerce networks aren't like selling this data to retailers, right? Like we're using anonymized first party data and we're selling advertising spots. So like, it's not customer data per se, and they're selling the ability to show ads to specific audiences. [00:45:26] So like a company would say, tell us who you wanna reach. We'll show [00:45:30] your ads to those people inside our, [00:45:32] environment. so like. [00:45:34] in Uber for example, like, um, Uber knows where their users have traveled. They know like where you travel, how often, what food you order on Uber Eats. So like a Starbucks could. Work with Uber and say like, I wanna reach more commuters on a weekday, on mornings who ride between like six and nine at least three times a week with an ad. [00:45:55] Encouraging them to like put a coffee order as their driver is like coming to pick them up. Like [00:46:00] another example [00:46:00] that I had was like Walmart, like they know what you buy, how often you buy online or in store, like all this first party data that [00:46:08] a company like Tide would be able to say, I want to sell a new eco-friendly detergent. [00:46:13] And they can decide. People that, uh, they wanna see people that have like bought a detergent in the past, but it wasn't from Tide, who also buys skews of products that are tagged eco-friendly. Am I thinking that the right way? Like is that, do I have that [00:46:26] Anthony: You have very quickly become top 10% of those who understand commerce media in the, in the industry. Right now, Phil, that is spot on. And I do wanna, I do wanna double down on what you said at the beginning. Yes. These companies are not just saying, here's a bucket of my customer data and throw it over the fence to, to the, to the top, top bidder, right? [00:46:44] Like, no, no, no, no. It's all, this is a, Hey, I wanna, I wanna target folks who have these, these buying habits and things like that. Um, and then it's up to the retailer to decide whether or not they want to make that access available to go ahead and place, place those ads or make that audience [00:47:00] portable in an anonymized way to go target it offsite and things like that. [00:47:03] So yes, it's all, the governance is a, is a key piece of the puzzle. Um. All the logging and, and observability and things like that to make sure that, um, you're doing the right thing. So, but yeah, the use cases are Exactly, exactly. I love the, I love the eco-friendly one that you mentioned because it's like both category purchase history is like, have they bought this brand or this category before? [00:47:25] And kind of this behavioral. Do they buy other products that are related to eco-friendliness or sustainability? Like those are the exact, um, use cases that are kind of hard if you ask your internal team, like, go write some SQL to surface this. But it's, uh, it's actually pretty easy if you type that into like AI studio and have an agent go and like look up the data and like build that data. [00:47:47] Um, you know, 'cause things that have been tagged, AI or things that have been tagged, sustainability or eco-friendliness. You might not have a tag like that. But the world knowledge of the large language [00:48:00] models and the foundation models can go figure out which product skews and, you know, are related to those things. [00:48:05] So it's a, it's a fun use of ai. It's really helpful, um, to see these things kind of churn out. These really remarkable, well-reasoned, um, marketing artifacts from very little context, or in some cases, many, many briefs in different languages and things like that. We've seen some pretty, pretty remarkable things spit out at skill. [00:48:26] 8. How Agent to Agent Commerce Operates Inside Marketing Funnels --- [00:48:26] Darrell: something that's on everyone's mind right now is machine customers, and I'd be interested. you know, in this area, if I was in consumer and retail, you've said that marketing has evolved from one to many, to one to one, to zero to zero, which is agent to agent marketing plus commerce, where assistant agents become the new storefront and the new funnel. [00:48:47] Do you think zero to zero is more of a parallel mode funnel and not necessarily a successor? Like what, what do you think about that? [00:48:57] Anthony: So, um, if we go [00:49:00] back to the story I had mentioned where we went out to Sunnyvale and talked to some Google execs about this generative marketing stuff, back in the day before the meetings, I sat down with Yasmine and Ahmad there who. You know, she's, she's an incredible human being. She has an incredible career. [00:49:16] Like, you should go look her up if you don't know who she is. She's, uh, I'm very thankful for the time I got to spend with her. And, um, she said, you know, how are you guys thinking about one-to-one personalization? And I said, like, I'll happy to answer how we think about one-to-one personalization, but I think one-to-one will be very short-lived and will quickly leapfrog to zero to zero, right? [00:49:38] This concept of agent to agent and. The, analogy I made for her, I said, you know, your colleagues on the workspace team who manage like Gmail. Like you have a spam filter there, right? Like that's a very tangible thing that exists today. How long is it before? Instead of a spam filter, you just say, here's an agent that's going to be the first filter for you as an individual, as a consumer, right? [00:49:59] And [00:50:00] you give it parameters, you give it some context, you give it some system instructions, right? All the things that you were calling, calling these, these pieces of context that you'd, you'd give to an LLM back then and just say. You have to get through this if you want to get to me. Right? And at that point, you know, we were already doing some early versions of agentic marketing on the, on the, on the business side. [00:50:19] And it's like, okay, you have the agent talking to the agent, how much is going to get through? And, um, that's where I think you're starting to see some of that become reality now. Like Google just launched, um, Gemini Enterprise for customer experience, where they launched it with the Home Depot. They launched it with, uh. [00:50:37] A bunch of, a bunch of partners where essentially you can just transact directly through Gemini, right? Like you can just talk to Gemini and transact right there in the, in the platform. And when that starts to happen, you say, oh, uh, wait a minute, like. We've built up all this marketing infrastructure to, you know, manage the conversation with the customers and things. [00:50:59] But [00:51:00] now one provider of these AI agents to consumers is going to be the single point of interaction between my business and, uh, and the customer. That's a little bit scary for folks, right? And obviously it's, you know, it's minuscule right now, but I remember when I was at a BI, we were talking about. [00:51:18] Folks were talking about how small Amazon and e-commerce were and what percentage of the market it was, or like that happened quickly, and this will happen much faster, more likely is my personal, my personal opinion. Um, when this does happen, like those agents still need to know things about your products and services, right? [00:51:37] They still need to know things about. Your history with the customer. If you do have this causality data about their state and their purchase history, their interventions, you've tried, and you can supply that to these agents in a way that's useful. Like your products are probably gonna get surfaced more often and like the messaging you can provide around your products is gonna be more useful. [00:51:57] Right? So even in this world where it [00:52:00] does shift to more agent to agent and machine, machine shoppers, I think you called it Daryl, or machine buyers, it's like the, the companies who have. They're. Data in order and can make the right features available to these agents. Um, are, are one going to likely have more success in terms of the revenue they can drive through these channels, but two, also have much more control, right? [00:52:24] If you understand the elasticity of your customers. If you get the customer ID back from this agent and you can make an informed decision about, ah, we'll take a discount with this one. 'cause we know their LTV is really high right now and it's important that we prevent churn with this customer. Things like that, like Right. [00:52:40] The use cases that we think about, um, are going to change a little bit, but the foundation of being able to. You know, hydrate, uh, these technologies with the right data that you have about your customers will be your edge versus your competitors in the market. I believe at least that's, that's what I believe about where this is going.[00:53:00] [00:53:00] Phil: Super fascinating space. Do you think, like what, like, I know that [00:53:04] 9. Google Universal Commerce Protocol Explained --- [00:53:04] Phil: you said like two of the catalysts in this space here are like Gemini Enterprise for customer experience, but you kind of teased out already like Google's universal commerce protocol. Um, do you think UCP and and Gemini are basically saying like. [00:53:18] We want to mediate, not just like finding products, but also completing and resolving like the whole commerce experience for folks. Like what are your thoughts there? [00:53:30] Anthony: Yeah, personal thoughts. Again, Google's in a unique position 'cause they already have Google shopping, right? So they have like the retail, like product library and things like that. So I think for them it's, uh, it's more about like connecting the dots, making it really good for the consumer, making it good for the retailer. [00:53:47] Um, so I, I think that they will. They will try to make the experience as smooth as possible for the consumer, right? Like when the consumer's trying to buy something, if they can [00:54:00] transact right in the system. Yeah. They have Google Wallet, they have Google Shopping, like all these pieces plug into it to make it good for the, the consumer to be able to go ahead and just transact there. [00:54:10] Now the, the where the companies I think need to be really focused is how do they make sure that they're. Products and services are the ones being being surfaced and, uh, making sure they get their fair shake within that system. And that's where having all this data supply, uh, is going to be really important. [00:54:26] Phil: Super cool. Anthony, I feel like we, uh, I'm looking at the time there and like we're, we're flying through this. This has been super fun. Got a couple last questions for you, but one that we like to end on, uh, with a bunch of folks on the show is like a happiness spin on this. Like, we talked a lot about tech here, but we teased out like the human component here. [00:54:43] 10. Personal Happiness System --- [00:54:43] Phil: But I wanted to give you a chance to, to, to chat about like your personal happiness system, like you're a chief data strategy officer at work. But at home, you're a girl dad of two, you're also an avid home chef. You got a ton of stuff going on. One question we ask everyone is how do you decide what deserves your energy at any given [00:55:00] moment and like, what's your personal system for staying aligned with what actually makes you happy? [00:55:05] Anthony: Oh man, what a great question. So, I don't know if you can see in frame here, but my two daughters, my two daughters are up there seven and four years old. Um, and you know, them and my wife, like, that's our, you know, that's our, that's our clan, that's our family here. So. The more I can do to support them. And especially, you know, in a world where it's, it's hard to see what's coming beyond, you know, I used to, used to be like 15 years, 10 years, five, I, you know, it's hard to see past that kind of, uh, that kind of one to two year horizon at this point. [00:55:34] It's just trying to do the right, trying to do the right things by them and set them up for, for success, I think is, is really important. And that, that gives me a lot of happiness. Um. And then I, I had mentioned earlier, you know, like trying to nudge, trying to nudge things in a direction that could be helpful for humanity. [00:55:49] It's a big, big, lofty goal, but, um, kind of being in the mix a little bit, trying to, trying to, trying to do what we can. Those are kind of the two, the two structural things for me is [00:56:00] try to leave humanity in a better place than I found it. And that, that drives little decisions too, right? It's like we just had a giant snowstorm. [00:56:05] If you're seeing somebody trying to dig their car out, you go over with your shovel and help 'em, right? And like, yeah. You talked about cooking. If I can, if we can invite. You know, another family over and we make meals, you know, make meals that we could share with them. Like that human element is so important. [00:56:19] So, um, you know, connection, strong human relationships and trying to leave, trying to leave things in a better place than, than I found is, um, is what really drives me. Phil, I. [00:56:29] Phil: Awesome. I love it. [00:56:30] 11. Favorite Books --- [00:56:30] Phil: You got a [00:56:30] bunch of books behind you there, Anthony, you got a favorite nonfiction or a favorite fiction book to recommend folks? [00:56:36] Anthony: Um, so I actually, these are back there. So I, I'd start with fiction. I think with so much uncertainty in terms of like human, human role if, you know, labor's disrupted and stuff like that. I love the e and m banks, uh, the culture series, it's, uh, it's, you know, it's pretty far out there like space travel and like in the future. [00:56:54] But there's also a plausible AI human, uh, element to it. Like, [00:57:00] yeah, AI could turn out this way and it could be positive, very, very like Star Trek of. And a, an abundant, uh, future, right? And an amazing abundant future that could be a place where humans still thrive and find meaning by, by doing, you know, different things that are, that are different than a nine to five job. [00:57:16] So I love, I love recommending that one to people who like sci-fi and can get into stories like that, but also have questions about, is there a plausible, plausible, optimistic outcome like in the, in the very, very, uh, far future. That, that one is like, at least the. An existence, proof of someone who's thought through something that, uh, that could be, could be positive. [00:57:37] Um, and then nonfiction. Um, I'll talk about two. So one extreme ownership. Like I love, I love that book. Um. You know, I think there's a lot of good lessons in there. I played sports, uh, growing up and, um, one of the things that sports taught me was kind of like accountability and like you watch game film after college football and like you have to stare it in the face and kind of talk about it. [00:57:58] And, [00:58:00] uh, being an owner and like really trying to find ways to. Change outcomes rather than come with excuses is, uh, is important. And then of course, I'll plug the one up top there. First party data activation. Uh, a co co co-author David Josten, one of our co-founders. So go check that out. If you wanna see kind of, uh, some strategies for getting your, your data foundation in order in, uh, a rapidly changing ai ai world that we live in, is, uh, if, if that foundation's no good, it's gonna be a lot, lot harder to apply these things with, uh, with value. [00:58:32] Phil: Love it. Anthony, thank you so much for your time today. This is a blast. Really appreciate it. [00:58:36] Anthony: Man, this was a lot of fun. Thank you both. I thank you both so much for having me on. It was, uh, it was a great time.