Forget The Shiny Objects And Focus On Fundamentals - AdExchanger

Loading summary

Transcript47 lines

[00:00]
A
Foreign.
[00:13]
B
Welcome to Ad Exchanger Talks, the podcast devoted to examining the issues and trends in advertising and marketing technology that matter most to you. Alison I'm Allison Schiff. You're listening to Ad Exchanger Talks, and my guest this week is Sharona Sankar King, the Chief data and Product Officer at Havas Media Network. We'll talk about why more data isn't always better, understanding causality using AI so marketers can move beyond basic marketing mix modeling, Havas's black book, not black box approach to strategy and planning, and why Sharona thinks everyone in an agency should be an AI developer now. But before we dive in, Programmatic IO New York is coming up on September 28th and 29th at the New York Marriott Marquee in Times Square. It's time to unleash growth across the open web. And to do that, we've got a seriously good agenda lined up for you with excellent speakers including Chris Kane and folks from Horizon Media, pmg. Hey dude, Dentsu X, Razorfish, Goodway Group, and more. Podcast listeners get 10% off with the promo code POD10. So what you waiting for? Get your ticket and see you there. Sharona, welcome to the podcast.
[01:35]
A
Thank you for having me.
[01:37]
B
So what is something about you that not a lot of other people already know and that I couldn't find out just by googling you or reading your LinkedIn profile?
[01:49]
A
Well, one of the things is that I am a second generation data scientist. My mother worked in the field for her entire career in data integration and engineering and my father also worked in the computer science field and has a PhD in the area. So I'm next gen, so, so a bunch of dummies. It's a really fun field and it's, you know, it keeps evolving. So I don't think there's ever going to be a time where it's not very engaging, especially as our new folks are coming about coming out of school, there's even more exciting things that they're going to be able to do. So my son's going into engineering as well, but a different type of engineering, robotics, so it's a little bit different.
[02:45]
B
Interesting. I'm equal measures excited for the future and also a little bit worried for I guess, people your son's age because AI, there's so much opportunity there, but there is so much change and jobs are changing and what people do during year one and year two is changing. How do people learn? Where do they get their experience? Their so many open questions and we're going to talk about a lot of AI Related stuff. But before we get into that, I wanted to do a little quick resume rewind because I like your trajectory. So now I know that data science can be genetic. And then you got a master's degree in statistics from Columbia, which kind of sets the stage for your career, because statistics is about translating wrong numbers into clear stories in a way that reveals like, a truth about our world and about our lives. And my brain doesn't work that way, but it is very cool that other people's brains do. You were a data scientist at Time Inc. You spent a lot of time at agencies with a focus on data and analytics. They're resistant at Experian as a group director of digital analytics. You were a managing partner of mec, head of Marketing Science at bbdo, six years in consulting at Bain as a partner and head of customer engagement and marketing, Chief Customer and Data Officer at Hart Hanks. And then in February, you joined Havas Media Network as Chief Data and Product officer. And I do promise I'll stop talking in a second and ask a question, but how would you say this overview
[04:22]
A
there don't know what to add.
[04:25]
B
Well, I do have a question that I think that I think you will be able to shed light on. But how does this variety of experience, like being a quant and a consultant, inform your perspective on AI? And how do you find yourself communicating about its possibilities and its limitations to clients who are finding themselves at this very interesting inflection point?
[04:49]
A
Yeah, I think, you know, the years I've spent in media and advertising, being a practitioner really grounds you in what actually drives outcomes. And you get the fundamentals of how all of this works to drive growth for clients. So that is foundational. But my years I spent at Bain and Company really gave me a broader perspective on just the challenges that organizations face overall. And what it really confirmed for me is that the fundamentals for success continue to separate leaders from the rest. Where we think about how we drive profitable growth in market share. And it really comes from a solid understanding. And I know this is going to sound really cliche, but a solid understanding of the customer and really being very well aligned in your data strategy and the culture of test and learn in the organization in order to. To really extract the value in order to outperform competitors. We say this over and over again in our industry, but it's so true. And it still really baffles me that there are so many organizations that are chasing the shiny things and sometimes skipping over some of the fundamental things that drive success.
[06:16]
B
The shiny things are so pretty.
[06:18]
A
They are. And they're fun, too. So that's. That's the other thing.
[06:22]
B
So if you really want to take good advantage, though, of the shiny things, you do have to have a workflow in place and build the foundations and then you can go and play. But you can't do that and run until you have, like, a road to run on. Right?
[06:38]
A
That is. That is so true. And I think, you know, there is work that happens in what, at Bain, we would call the backstage or the back office. And those are the operational things. Getting the culture right, getting senior leader alignment, making sure you have the right KPIs, that everybody's driving towards having a North Star ambition for everyone to lock arms and actually work and row in the same direction. Those things cannot be discounted. And they can be facilitated both by leadership, but also at every level. There are operational things you can put in place and how you reward people, how you incentivize them to work together and collaborate that are really important to driving success. It's not just the data, it's not just the AI. There are really cultural and procedural things that can be put in place that can drive success.
[07:40]
B
And Havas has this philosophy, or maybe you'd call it an approach that you guys refer to as black book, not black box, which is very clever wordplay. So kudos on that. But there's also some real substance behind it. So what is that? What is it like when. What does it mean in practice for.
[07:57]
A
Yeah, it means a couple of things. It means, one, we, when we built our converged API platform, which is our tool that is end to end workflow from intelligence to design, to activate and to measure, and then feeding back into intelligence again. That workflow that we built is really just the underpinning and operation system that is underpinning our actual process that we have our people process. So how we work as an organization together, how we collaborate together, so the operating system can only reflect and support a good operational culture and spirit of collaboration across people. And so when we say a black book and not a black box, it means transparency at every single stage. It means that we built the operational system system and all of the data and all of the tools to be not locked into some big acquisition that we make so that we are now tied to a certain data set for life. Right? We can pivot from one type of data set to another or one type of tool to another, customized to what our clients need and not built around what we want to sell or build around what we think clients will need. So it's very bottom up outcome driven system and at every stage you can see how it's working. So as we look into a particular tool, let's say our audience planner tools, you can see how AI is actually building that audience step by step. And it explains to you what data it's using, how it's using it and how it's feeding to the final recommendation on a either Persona or profile or target audience or geographic insight. So that is what we mean by it being a black book. And then also what this allows us to do because we built it based off of client need. We created the architecture, the technology architecture that it sits on to have integration points with our clients. First party data to integrate with our clients also have technology. We don't want to be duplicative. So we didn't build a one size fits all system. And that means that the platform is flexible and it allows us to do more things that that I think some of the other platforms that our competitors have may have trouble because they built it for a specific their workflow and how they operate, but not really around as much how their different clients work based on industry, vertical, based on nuances of their business.
[11:09]
B
That does feel like saying Axiom. Without saying Axiom. That's just me saying that it does feel like the Holcos are shelling out for a lot of stuff right now and Havas decided to build rather than buy. And that explanation you gave makes perfect sense to me for what you're trying to do. I mean, buying is about moving as fast as humanly possible or like I should say agentically possible because the world is changing and it really, you know, behooves everyone to keep up with it, otherwise, you know, they're just going to be left behind. Would have us like, would it? I know this is not necessarily something that you, you get to decide, but product people and data people, they get pulled into these kinds of decisions because they get asked what's possible, what's doable, Would it make sense to buy something at some point or is it really a build over buy culture?
[12:08]
A
It is a, I'll say it build and partner culture. So we negotiate partnership deals with, for data with our providers and with our partners. We have over 500 partner integrations within our platform. We're constantly managing them and looking for new partner deals because I think when we partner we get to the best of the ecosystem. The ecosystem is very frothy and there's a lot of good stuff out there and there's also a Lot of garbage out there. And so I think the, the point of view we have is where we can partner with really excellent capabilities. We will partner with them and we will integrate them and build new solutions on top of them. So we can get exclusive, for instance, exclusive data access or exclusive access to build specific types of tools that will help our clients. And that gives us a competitive advantage, but it also allows us to help those partners even elevate their business. And we're often helping them build new capabilities on their side as well. So it is very mutually beneficial. And I think it's an exciting way to work because we're always able to see something new that we didn't think about. Right. And so we do build things as well. We build a lot of proprietary tools as well. So build and partner really is where we are. Doesn't mean that we won't ever. If there's something that really fits within our stack. I don't think, you know, we would buy a big data company or anything like that, but there might be some very nuanced AI capabilities that we might integrate and bring into the fold. I wouldn't rule that out.
[13:55]
B
Well, if you do, please tell us. We would be very interested to write about that. I want to talk a little bit about transparency, which you, you touched on briefly. Do you guys see transparency as a competitive differentiator? I don't feel like it should be, but it kind of is. I mean, everyone talks about the need for transparency. I mean, frankly, people won't sh up about it. But we're also at this moment in time when so much of the AI landscape is moving in exactly the opposite direction. And also the supply chain is still very murky. People seem to talk out of both sides of their mouth. And they spend a lot of money on pretty opaque systems like Performance Max and Advantage plus. They almost can't not spend on those things. But it feels like, I don't know, trying to keep two things in the air at the same time, two balls spinning.
[14:41]
A
Well, listen, we're agents of our clients, right? And we are working on behalf of our clients. And if we aren't protecting transparency, privacy and compliance, then I don't think we're doing our jobs very well. I think that's where our focus is. It's going to continue to focus on that. And for me, I think that when I step into the room with a CMO or a cto, this is the conversation that clients are very, very interested in having to make sure that that is maintained as we leverage AI or Agentic systems that could be doing autonomous things that could get in the way of transparency. So yeah, it's something we keep an eye on and we are committed to continuing to audit ourselves in order to
[15:30]
B
ensure transparency and just sticking with Google's performance max Meta's Advantage plus for a moment. I mean they're the two most prominent AI powered ad products out there. A lot of platforms have been releasing their own maxes and pluses. It's like the naming conventions are so funny to me because they all sound like streaming services or healthcare providers, but they're unapologetically opaque. Like the promise is that you put money in and you get results out and how is more or less of a mystery and you've got to be okay with that because hey, we're performing, why question it? Because performance, when a client comes to you and says, yeah, like those black box systems are performing well for me, like what do you, what do you say and do you recommend those platforms? Like, how does that play in?
[16:21]
A
I think that we still have to go back to fundamentals. Right. We have a brand to ensure, to protect from a long standing equity standpoint and then we also have short term performance and gains that we have to manage. I think there's a balance and I think that anytime that we're working across all of these different platforms and tools, we want to always make sure that we are balancing the two and making sure that, you know, listen, a percentage of the mix can be, you know, running for performance, but then you have to think about it as a mix and understand that there are other goals that you have to ensure as across the full consumer journey. And so that is where I believe the, the science and the, a lot of the critical thinking has to go into it before you just set it and forget it.
[17:18]
B
So I guess it's like art and data science.
[17:20]
A
It's always been that way.
[17:23]
B
It hasn't changed and really it shouldn't. Although the pendulum really swings, right? Like brands get so focused on performance that they forget that there's a top to the funnel and then have to rejigger their investment. It seems to happen more than periodically.
[17:42]
A
We used to talk about being careful not to optimize yourself into a corner. Right. I can get a hundred percent sales off of one person if that, that one person's always buying from me. Right, Right. But what about that person that's, that's out there that could, I could acquire into my brand that doesn't look as, you know, immediate when it comes to getting a return? So I think we have to think about it in a mix. We have to think about it in not just the terms of performance, but in share gain, in brand protection and brand equity. And as long as we keep those in mind, I think we will always be driving performance for today and tomorrow.
[18:30]
B
So one more quick one. Well, maybe it's not that quick actually before we take a break, but there's a principle I know that Havas leans into hard, which is that more data isn't always better. So you shouldn't really be a, shouldn't be a data hoarder. You should focus on collecting and using fit for purpose data, which is about knowing what you need and just being more purposeful. You do hear more and more support for that notion, but in a way it's almost like heretical in an industry that really fetishizes scale and volume and cheap reach. And it has since the beginning. So make the case for me, why is fit for purpose data the right frame? And why is it really important for marketers to stop treating data like they're collecting nuts for winter or something? I love that.
[19:24]
A
Listen, as a data scientist, one of our very first things we always look at is garbage in, garbage out. And so there's a lot of data that can be useful for nothing. And there's some data that is extraordinarily useful in finding those treasures of gold that are actually going to drive our clients business in a more efficient and effective way. That is where the science comes in is really being able to find what's actually going to drive outcomes. And so there's data for data sake is just not a wise way of going about an efficient search for value. Right. And so we use AI to help us mine the data and find where those nuggets of gold are. And that's what, you know, predictive and causal AI actually does. It allows us to find what are those nuggets of truth and information that allows us to drive, drive understand causality between a data point and an outcome. And so what we've found with outcomes is that they're non linear, they don't behave always in the ways that we expect. And that's why machine learning and AI, hand in hand and statistics working together are actually how we get to the right answer.
[20:58]
B
Well, AI and measurement are two of our listeners and readers favorite topics and that's what we'll spend the whole second half talking about. So stick with us. All right, we're back and I want to spend a whole bunch of time now nerding out on measurement which is a topic where there's a lot of confusion, there's a lot of noise, big promises, snake oil over the last. There is. And if you could see my inbox, I mean, the pitches I get from companies, I'm like, if you could really do that, I would have heard of you every day. Exactly. But over the last bunch of years, media mix modeling has had this massive resurgence, like a post cookie renaissance. But you guys have been talking about something that goes a step further. And you talked about it too before the break, just causal AI. So you explained what it is. But I want to go a bit deeper. Like talk to me about how it's different from MMM and how it goes a bit further than predictive AI, because I feel like it does. And what makes it interesting to you there.
[22:23]
A
There are a couple things and you know, I'd like, I like to start. The way we like to start to talk about this with clients is thinking about the maturity curve of measurement. Right. Your. Your basics are being able to report as holistically on the data that you have and the outcomes that you can track in a very direct way. Right. And that those are your dashboards, your reports. Then you have basic mmm, which is trying to do a step better than that. We used to have digital attribution, which tried to count all the beans. But we have too fragmented of a ecosystem these days to be able to track cookies everywhere. Right. Because of walled gardens and other things. So market mix modeling got a resurgence. Some people say media mix modeling, some people say marketing mix modeling. It's all really the same way of building the models. These are econometrics models that are looking at things over time and trying to tease out the causal effects. So a basic market mix model is allowing you to look at what's happening as far as a channel mix. So across the mix, what's driving sales. Or you can build a model that is looking at how you drive brand equity and how that drives sales over the longer term. But where it falls short is where you have channels that are not a significant amount of the volume. So for instance, TV is a large volume, but you might have some social media activations that may be a much smaller volume in comparison to tv, which is a mass channel. And so a lot of those effects get washed out. So what the solve for that was is something called a nested mmm. And nested mmms are trying to look at all multiple different types of outcomes. So maybe sales or sorry, brand on sales, how a brand may impact sales at different time horizons. And you might want to look at channels that often get lost in the mix by creating these weights that you can nest into the, the overall model. So it essentially is multiple models stick stitched together with statistics to allow you to talk about different outcomes and different time horizons with one model. Well, with the nested models, there's actually a shortcoming. And that shortcoming is that because it's a statistically based model, it actually has the problem of multicollinearity. And what that means is that there are a lot of variables that are correlated or they're too correlated to tease out which variable actually is driving the causal outcome. And so that means that the people who are building these models have to throw out a ton of data that might actually be explaining what's going on. And so in order to make the models understandable. Right. And so what large causal models are, it's a new entire paradigm shift, is to leverage AI to solve this problem. And it's similar to a LLM in the fact that an LLM will predict from one piece of text we call a token to the next piece of text, to the next piece of text in a probabilistic way in order to form a sentence, a line of code, a paragraph, a book. Right. And what a large causal model is similar, but it's looking at one event to the next event in a probabilistic chain. We used to use something called Markov chains to solve this, but what this actually does is it looks at those chain of events in a multi layered fashion. So now you can look at multiple chains of events all at once across different layers. And when I say layers, I mean at the portfolio level. So if you have multiple types of brands that have halo effects on each other, and then the next level down might be the channel level of brand versus performance media. And then the next level down may be specific channels like TV and radio versus different digital activations. And then the next level down from there might be creative and content and call to action and so on and so forth. And so with a large causal model, actually what the AI is doing is because it's building these probabilities from one causal event to the next, you can actually cut the chain at any place that you want and that becomes your outcome. Right. So if you wanted to understand how TV activation is actually impacting your digital, right, you can look at that in a time horizon and the digital could be an outcome which is a halo effect of TV on some digital channel. You can look at how brand is affecting sales at different time Horizons one month out, two month out, three years out, five years out. Right. But because it's multi layer, it also allows you to understand the, the relationships in a multi layered way as well. So when you optimize and say, hey, listen, I am in a year where I have a competitor entering into the market and I need to reduce my budget, what is by X amount of dollars, right, or X percent because of some kind of economic pressure that I have on my business. But I still need to drive XYZ outcome or something, number of sales, tell me what I need to do and what, what this allows you to do is build those scenarios and all of those layers from the portfolio level all the way down to call to action can be optimized with one model instead of multiple. Having to build multiple models to answer that question and then stitch them together. So that is what large causal models are allowing us to do. And I would say don't get them confused with just causal inference or causal AI inference, AI, which there are several companies that have that those are more built for, you know, test and learn, uplift modeling. And that's quite different, which is just, you know, what we think of test and learn, but automated.
[29:21]
B
So when you explain this to clients, are they nodding and do they look confused at any certain point? It's very convincing. And I'm like, yeah, of course, why wouldn't you do that? It doesn't sound necessarily easy. And I wonder how much involvement brands need to have in the process if they're set up for a culture of experimentation. So how do you make what you just talked about reality for brands?
[29:50]
A
Well, first of all, it already is reality and it's driving results for quite a few brands that have adopted it. Some of the biggest logos out there. So it's not, this is a tried and true now at this point type of methodology, there are believers and there are, I would say nine out of the 10 clients that I talk to are very bullish on this because they are hitting up against a wall when it comes to MMMs. They've squeezed all the juice they can out of nested models and they're now reaching the limitations of what MMMs can do. And so for those types of clients that see where the limitations are and are feeling the pain of that, they are ready to go to the next stage in the maturity curve, which is a large causal model. And I would say that maybe, goodness, I think there's only been one client I've ever talked to who was hesitant. And the reason there was hesitation is because there's so much noise around causal AI and frankly a lot of the causal AI players in the space are not that good. And so there's a little bit of muddying up when it comes to the signal from the noise on what companies can actually do this well and which ones can't.
[31:17]
B
Oh no, fair enough. But if measurement is an area where there's a lot of hype and a lot of noise right now, I don't think it holds a candle to how much hype and noise there is around agentic AI. And there's also confusion and conflation of terms which really doesn't help the difference between automation and agentic AI and fully autonomous AI, for example, because I've heard people use the term agentic very liberally, I've done it myself to describe basically anything that AI does which isn't right and is definitely not nuanced enough. So let's define a few terms. So automation versus agentic AI versus, I don't know, autonomous AI or whatever the term is that you hear come up most often. And what are some of the practical implications of those differences for marketers? Like why is it really important for them to know the difference?
[32:11]
A
Yes, there is a clear delineation between the three. And so automation is a rules based execution that is something that you can do with machine learning, you can do with other techniques, you can leverage AI to do it. But it's really where you're saying here, you know, you're finding where there are areas where you can build efficiencies and then you're creating rules and executing those rules in order to get the efficiency. Agentic AI is a little bit different because it is in some cases able to do chain of thought reasoning and it allows, it allows you to go out to other applications and perform actions. And so like the way a human would reason, okay, I need to go to this app and pull this piece of data, then I need to go to this other app and implement this and then actually make the action happen. And an agent workflow can actually do that. And what you want to have is not just one agent, you have a team of agents working together. So some of the agents are working as governance agents which are monitoring the outcomes or the outputs of the individual agents that are doing the steps to the action. And then you have some agents that are actually opening the doors to the APIs and checking the data and making sure that garbage is not getting poured into the process. So there's quite a few ways when you think about agentic AI but the number one characteristic is that it can perform actions based on off of, you know, reasoning and chain of thought and step by step working together with other agents. When you think about fully autonomous systems. Well, fully autonomous systems are a group of agents that work together without human oversight at all. And only 7% or so, based on a Bain study of corporations have any kind of systems that are running fully autonomously with agents. So we are far from being able to really realize the, the end to end autonomous agent world because there are a lot of unintended consequences of some of the agents actions and they still do need human oversight and governance because they're probabilistic and not deterministic. Right. A rule space is very deterministic. We can go and we can query the rules, we can look at them, we can make sure that it's being executed in the way that you intended. Whereas some of these autonomous systems as you can imagine, can be very complex and black box. And so it's very hard to tease out why an agent made a specific, did a specific action or used a specific piece of data. And so I think the, the often what we're, we're trying to do in the successful systems are you use agents to monitor the agents, but even those agents make mistakes, right? So that's where I think we're still got quite a ways to go to be fully autonomous, especially in this industry where we have a lot of risk tolerance to manage.
[36:03]
B
Where would you say most brands and agencies too are sitting in the AI adoption curve or on the spectrum or however you want to think about it, Just really honestly, because there is a gap, I think, between what I see in press releases and what people tell me and what people say on stage at conferences and what's actually happening.
[36:25]
A
I think on the client side there are a lot of promising point solutions within organizations, right? So some folks have done a really great job of building efficiency tools to help with productivity, to help with some of the tasks that are in their workflow. Where I don't see a lot of progress is the ability to take an end to end workflow across multiple teams and to break that down and actually start generating new business outcomes and growth from it in a consistent and reliable way. So it doesn't mean that it doesn't happen and there aren't companies that aren't able to unlock that. But I feel that, you know, when I, when I look across the full spectrum of companies out there and even, you know, there's Bain studies on this, that, that, that you know, you can look up, but they're, they're, they're also saying the same thing. And really the, the real reason why a lot of companies struggle to get the end to end is alignment at the top when it comes to corporate alignment with the leadership and getting the culture and the people on board to really unlock the full potential of what AI can do. And then also it's education, right? So there's quite a bit of training and understanding of being able to utilize and build enterprise grade systems and platforms that can go end to end. And then lastly, I think there's a willingness to test things and be okay with things not being perfect the first two or three times you do it and working your way to the right answer. So I think you have to have all of those in place and you have to also have a willingness to do the hard work on your data because AI consumes data and the quality of the data will give you the quality of your output. And that has, hasn't changed. I mean, quality of data has always been the number one driver for organizations for goodness, decades. Right. And, but now even more so because AI, that's the food that AI eats, right? So if we don't have good data, then AI doesn't have the fuel or the food it needs to do the jobs that it, it, it's being asked to do.
[39:07]
B
Guess we don't want it to eat out of, out of a garbage can.
[39:10]
A
Exactly. It's got to eat good nutritious food.
[39:14]
B
Gotta look at the food pyramid. What, what does an agency that's genuinely ready for agentic AI look like on the inside? Like, what are the markers? What work have they, or like have you people using AI across the org in a way that's actually helpful, that saves them time because I know it's a thing at Havas too to really democratize the use of AI. So it's not just engineering and data folks, but everybody, so it makes sense for their job.
[39:40]
A
Well, at havas, our philosophy is that everyone in our organization is an AI developer, right? And so what the wonderful thing about AI is in, in the generative AI era is that it's made itself accessible to people through natural language and through tools that are very intuitive. That said, you have to learn how to be able to prompt well and you have to understand what each component can actually do and what's possible. And so we do have a training called Prof. AI that everyone in our organization is required. Even new hires, when they come in, they take the training. We have an extensive HAVAS University of even more training for power users. And we have a leap program that gives each and every employee a certain amount of money to put towards courses as well. To take it even further, it's not a nice to have. It is a requirement for every single employee, including our admins and folks in jobs that you wouldn't think that they would need it, but everyone is required.
[40:55]
B
And what sort of guardrails need to be in place before. I don't know, you'd let an autonomous agent make a call that could impact a client's business? I know we're not really there yet with autonomous, like fully autonomous AI. It's just some small, less than 10% of. Of brands that are doing it.
[41:13]
A
But we, we have agentic layers. Um, so I don't want to go into the autonomous world because we're, we're really cautious about what we build in an autonomous fashion. Not that it doesn't exist, but we test. Test our way into it. Right. But we do have an agentic layer on our tools and we have a partnership with Accio that we work on building multiple layers of agentic architecture across our platform in order to take a lot of the manual tasks out of the things that we do. We've been able to reduce some tasks that take people three, four weeks. In fact, I had one planner tell me that, you know, something we built saved over a month of time for this individual and how it would normally happen. So we're taking a lot of the manual heavy lifting. And now what that's allowing our planners and our measurement and analytics folks or to really spend more time on the critical thinking around the insight. And I'll also say that spark of creativity that comes back when you're not spending all day in drudgery. Right. And so I think we have happier employees. Folks are way more excited about their work because they're able to just unlock value in ways that they couldn't before. And it's unlocking an amount of ingenuity and creativity that frankly, it's why I love my job when I see how excited folks are to work with all of these different tools.
[42:54]
B
Penultimate question. I can't believe we're basically out of time. And I promise this question is not coming from a place of sour grapes because I had to undergo a lot of drudgery and grunt work and heavy lifting. But isn't there some value in the drudgery? Right. Like you kind of have to do that so that you know how to prompt a system and you know what a really good Output looks like it's hard to have, you know, really creative thoughts and be able to think laterally about insights if you haven't done the grunt work, or at least there's an argument to be made about that.
[43:31]
A
Yeah, I agree. It's like, you know, math majors, right? Unless you're willing to actually do the math and derive formulas, right, you're not going to be a very good mathematician. Right? So here's what we do. We have Havas University that takes you through the full workflow from intelligence all the way out to activation. We still train folks in every step of the way. But what I think this is allowing folks to do because we built it as a black book and not a black box, because it's an open system and you can see how every decision is made. It's allowing folks to actually be more critical because now you have a better understanding of every single component that is driving a potential decision. Just because an AI serves something up to you doesn't mean it's the right thing to do. So our teams still have to be analysts, they still have to be scientists, they still have to be critical thinkers, they still have to understand the levers that drive success. They still have to understand the client's business, they still have to understand all of these different things. So when we take out the drudgery, what we mean is, listen, when back when I started in this business, we had, what is it, Lotus Notes before we had Excel spreadsheets, right? And a lot of people said, oh, well, people aren't going to know how to do things anymore if we have spreadsheets, right? Because the spreadsheet automatically populates the formula for you and all kinds of things. And no, it just unlock potential for folks to build the next level thinking, right? And so it just took out some of the work to get to the next level thinking. And I think that as long as we keep the. The black book and not black box, and we don't make it a completely closed system where humans can't interact with AI, we are learning from each other. The human is learning from the AI and the AI is learning from the human. And I think it's a symbiotic relationship now that we are going to continue to have with AI and that's how we're going to learn, and that's how we're going to actually build those skills. And so I'm actually very excited and hopeful for this next generation because they're going to get to learn things that we didn't get to learn. Because we were so inundated with the drudgery. Just like Lotus Notes versus Excel spreadsheets. I hope I don't get in trouble with. I'm sure it's okay saying that.
[46:20]
B
Well, last question. We were talking a little bit about different varieties of snake oil. What is the most just like annoying salesy thing that you hear AI startups say? Because I'm sure you see a lot of decks and a lot of pitches and I know we're early. I'll give you a second to think. Mine is that everyone says they're the first and therefore the best. Our associate editor, Joanna Gerber, she covers AI for us. She wrote this very spicy. Wasn't really that spicy, actually. It was just true op ed about. It was earlier this year about how these companies just need to stop claiming that they're the first. Like it's really okay to just have a product and explain what it is and not like brag about being the first. Like, how many first of its kind GEO platforms can there really be? Like, it can't be all of you. But yeah, what do you wish you wouldn't hear?
[47:17]
A
Again, I agree with that one. The first and the best. I think everybody claims that. But then I think that, you know, one of the things that really bothers me is when a company says, well, that they can just completely replace people without oversight. That is a red flag to me because I don't believe that we're quite there yet and everything that we build today still needs oversight. Even the biggest players in the world, Anthropic and others, they would tell you that humans need to be intimately involved in what AI is doing and to have as much ethics built into it as well. Because we know that human data that is generated out there is biased data. And we want to make sure that we're also not just getting the right answer, but we're getting the right and ethical answer so that there's fairness also built in.
[48:30]
B
So AI startups do not pitch Sharona on how you're cutting humans out of the loop because she will not like it.
[48:36]
A
No.

Summary7 min read

AdExchanger Talks Summary

Episode: Forget The Shiny Objects And Focus On Fundamentals
Date: July 21, 2026
Host: Allison Schiff
Guest: Sharona Sankar-King, Chief Data and Product Officer at Havas Media Network

Episode Overview

This episode centers on steering the advertising and marketing technology industry away from “shiny objects”—the surface-level trends and unvetted tech solutions—instead urging a renewed focus on foundational practices, transparent strategies, and a thoughtful approach to AI and data in driving real client outcomes. Sharona Sankar-King shares Havas Media Network’s approach to data, measurement, and AI, expounding on how to combine time-tested fundamentals with emerging technologies.

Key Discussion Points & Insights

Sharona’s Background & Perspective

Second Generation Data Scientist (01:49):
Sharona explains that data science runs in her family, highlighting her mother’s and father’s careers in the field, and her son’s current pursuit of robotics engineering.
Straddling Quant & Consulting (04:49):
Sharona notes that her broad career—from agency practitioner to partner at Bain & Company—helped her see “the fundamentals for success continue to separate leaders from the rest” and that many organizations still “chase the shiny things” rather than focus on what drives business.

"It really confirmed for me is that the fundamentals for success continue to separate leaders from the rest... and it still really baffles me that there are so many organizations that are chasing the shiny things and sometimes skipping over some of the fundamental things that drive success." – Sharona (05:07)

On Foundations Before Tech

Building Foundations (06:22–07:39):
Success in leveraging tech begins with strong operations: senior leadership alignment, a test-and-learn culture, North Star goals, and effective KPIs.

“The operating system can only reflect and support a good operational culture and spirit of collaboration across people.” – Sharona (08:00)

Black Book Not Black Box: Havas’s Approach

Transparent Technology & Flexible Platforms (07:56–11:08):
Havas has built its Converged API Platform for workflow transparency—from data intelligence to activation, all visible and modular, customized for each client’s needs.

“So as we look into a particular tool, let’s say our audience planner tools, you can see how AI is actually building that audience step by step. And it explains to you what data it’s using, how it’s using it and how it’s feeding to the final recommendation…” – Sharona (09:37)
Partnership over Acquisition (12:07–13:55):
Havas prefers to “build and partner,” integrating the “best of the ecosystem” through 500+ data and tech partners, rather than acquiring and locking in data or tool silos.

“We negotiate partnership deals…We have over 500 partner integrations within our platform...We do build things as well. We build a lot of proprietary tools as well. So build and partner really is where we are.” – Sharona (12:16)

The Real Value of Transparency

Transparency as a Differentiator (13:55–15:29):
Sharona emphasizes that true transparency is not just a buzzword but a client expectation, especially as more AI-powered “black box” ad products (Performance Max, Advantage+) proliferate.

“If we aren’t protecting transparency, privacy and compliance, then I don’t think we’re doing our jobs very well... clients are very, very interested in having to make sure that that is maintained as we leverage AI..." – Sharona (14:48)
Balancing Performance & Brand (16:20–18:30):
Marketers should beware of optimizing into a corner and remember the mix: combine short-term performance with long-term brand health.

“We used to talk about being careful not to optimize yourself into a corner.” – Sharona (17:42)

Data: Fit for Purpose, Not for Hoarding

More Data Isn’t Always Better (19:24):
Advocating against mindlessly collecting data, Sharona stresses purposeful, “fit for purpose” data selection and using AI to identify valuable data that actually drives client outcomes.

“There’s a lot of data that can be useful for nothing. And there’s some data that is extraordinarily useful in finding those treasures of gold that are actually going to drive our clients' business...” – Sharona (19:33)

Measurement: Beyond MMM to Causal AI

Limits of MMM, Advantages of Causal AI (22:23–29:49):
Sharona reviews the “maturity curve of measurement:”
- Dashboards/Reporting → Media Mix Modeling (MMM) → Nested MMM
- Nested MMMs solve some issues but still suffer from multicollinearity, often discarding relevant data.
- Large Causal Models (Causal AI): Move beyond by modeling multi-layer chains of events (portfolio, brand, channel, creative) using probabilistic, AI-driven chains, allowing much deeper, precise, and actionable scenario planning.
  
  “With a large causal model...the AI is...building these probabilities from one causal event to the next...you can actually cut the chain at any place...and that becomes your outcome.” – Sharona (26:45)
  
  "Nine out of the 10 clients that I talk to are very bullish on this, because they are hitting up against a wall when it comes to MMMs." – Sharona (29:54)

Demystifying Automation, Agentic, and Autonomous AI

Definitions & Practical Implications (32:11–36:02):
- Automation: Rule-based, efficiency-focused, often deterministic.
- Agentic AI: Able to perform actions using “chain of thought reasoning,” sometimes as a team of agents (with governance and QA agents monitoring others).
- Autonomous AI: Agent teams working end-to-end with no human oversight. Still rare (<10% of companies) due to complexity and risks.
- Full autonomy is not yet viable for most; human oversight and agentic guardrails remain critical.
"We are far from being able to really realize the end to end autonomous agent world...they still do need human oversight and governance because they’re probabilistic and not deterministic.” – Sharona (34:48)
Where the Industry Really Is (36:02–39:07):
Most brands have pockets of AI productivity but lack end-to-end workflow transformation. Why? Leadership alignment, education gaps, data quality hurdles, and limited appetite for experimentation.

“There are a lot of promising point solutions within organizations…Where I don’t see a lot of progress is the ability to take an end-to-end workflow across multiple teams and to break that down and actually start generating new business outcomes...” – Sharona (36:29)

Democratizing AI Across the Agency

Everyone as an AI Developer (39:14–40:55):
Havas requires all employees—admins to analysts—to be AI literate. Continuous learning and hands-on upskilling are the norm.

“At Havas, our philosophy is that everyone in our organization is an AI developer...” – Sharona (39:39)

Benefits of Agentic AI—But With Guardrails

Agentic Success Stories (41:13–42:54):
Through partnerships and layered agentic systems, Havas reduces weeks of manual work to days or hours, freeing planners to focus on insight and creativity.

"We’ve been able to reduce some tasks that take people three, four weeks....I had one planner tell me that something we built saved over a month of time..." – Sharona (41:37)

Does Drudgery Still Matter?

Retaining Analytical Rigor Without the Tedium (42:54–46:20):
Even as AI automates routine work, Havas ensures employees understand the full workflow—comparing the evolution to the move from Lotus Notes to Excel—emphasizing human/AI symbiosis.

“Just because an AI serves something up to you doesn’t mean it’s the right thing to do...as long as we keep the black book and not black box...we are learning from each other. The human is learning from the AI and the AI is learning from the human.” – Sharona (44:23)

Industry Snake Oil and Closing Thoughts

Annoying AI Startup Promises (47:16–48:29):
Sharona flags the "we can replace humans entirely" sales pitch as a red flag, stating full oversight and ethics remain essential.

“...when a company says, well, that they can just completely replace people without oversight. That is a red flag to me because I don’t believe that we’re quite there yet and everything that we build today still needs oversight.” – Sharona (47:34)

Notable Quotes & Moments (with Timestamps)

"Chasing the shiny things and sometimes skipping over... the fundamentals that drive success." – Sharona, (05:07)
"The operating system can only reflect and support a good operational culture..." – Sharona, (08:00)
"We say a black book and not a black box, it means transparency at every single stage." – Sharona, (08:33)
"There’s a lot of data that can be useful for nothing. And there’s some data that is extraordinarily useful." – Sharona, (19:33)
"With a large causal model...you can actually cut the chain at any place that you want and that becomes your outcome." – Sharona, (26:45)
"At Havas, our philosophy is that everyone in our organization is an AI developer." – Sharona, (39:39)
"Just because an AI serves something up to you doesn’t mean it’s the right thing to do." – Sharona, (44:23)
"When a company says...they can just completely replace people without oversight. That is a red flag to me..." – Sharona, (47:34)

Timestamps for Important Segments

Sharona’s Personal and Professional Background: 01:49–04:24
Foundations vs. Shiny Objects: 04:49–07:39
Black Book, Not Black Box: 07:56–11:08
Build vs. Buy & Transparency: 12:07–15:29
Performance Platforms & Media Mix: 15:29–18:30
Fit for Purpose Data: 19:24–20:57
Measurement & Causal AI Deep-Dive: 22:23–29:49
Automation, Agentic & Autonomous AI Explained: 32:11–36:02
Democratizing AI at Havas: 39:14–40:55
AI & Drudgery/Analytical Rigor: 42:54–46:20
AI Startup Red Flags: 47:16–48:29

Conclusion

Sharona Sankar-King and Allison Schiff deliver a highly insightful discussion, urging the industry to resist distraction by surface-level innovations (“shiny objects”) and instead double down on cultural alignment, transparent tech, and purposeful, fit-for-purpose data practices. Listeners are left with a technically rich yet pragmatic blueprint for harnessing AI and measurement in a way that elevates both client outcomes and agency talent. Guardrails, transparency, and continuous learning emerge as the cornerstones for agencies wanting to thrive amid waves of technological change.

Loading summary

Transcript47 lines

[00:00]
A
Foreign.
[00:13]
B
Welcome to Ad Exchanger Talks, the podcast devoted to examining the issues and trends in advertising and marketing technology that matter most to you. Alison I'm Allison Schiff. You're listening to Ad Exchanger Talks, and my guest this week is Sharona Sankar King, the Chief data and Product Officer at Havas Media Network. We'll talk about why more data isn't always better, understanding causality using AI so marketers can move beyond basic marketing mix modeling, Havas's black book, not black box approach to strategy and planning, and why Sharona thinks everyone in an agency should be an AI developer now. But before we dive in, Programmatic IO New York is coming up on September 28th and 29th at the New York Marriott Marquee in Times Square. It's time to unleash growth across the open web. And to do that, we've got a seriously good agenda lined up for you with excellent speakers including Chris Kane and folks from Horizon Media, pmg. Hey dude, Dentsu X, Razorfish, Goodway Group, and more. Podcast listeners get 10% off with the promo code POD10. So what you waiting for? Get your ticket and see you there. Sharona, welcome to the podcast.
[01:35]
A
Thank you for having me.
[01:37]
B
So what is something about you that not a lot of other people already know and that I couldn't find out just by googling you or reading your LinkedIn profile?
[01:49]
A
Well, one of the things is that I am a second generation data scientist. My mother worked in the field for her entire career in data integration and engineering and my father also worked in the computer science field and has a PhD in the area. So I'm next gen, so, so a bunch of dummies. It's a really fun field and it's, you know, it keeps evolving. So I don't think there's ever going to be a time where it's not very engaging, especially as our new folks are coming about coming out of school, there's even more exciting things that they're going to be able to do. So my son's going into engineering as well, but a different type of engineering, robotics, so it's a little bit different.
[02:45]
B
Interesting. I'm equal measures excited for the future and also a little bit worried for I guess, people your son's age because AI, there's so much opportunity there, but there is so much change and jobs are changing and what people do during year one and year two is changing. How do people learn? Where do they get their experience? Their so many open questions and we're going to talk about a lot of AI Related stuff. But before we get into that, I wanted to do a little quick resume rewind because I like your trajectory. So now I know that data science can be genetic. And then you got a master's degree in statistics from Columbia, which kind of sets the stage for your career, because statistics is about translating wrong numbers into clear stories in a way that reveals like, a truth about our world and about our lives. And my brain doesn't work that way, but it is very cool that other people's brains do. You were a data scientist at Time Inc. You spent a lot of time at agencies with a focus on data and analytics. They're resistant at Experian as a group director of digital analytics. You were a managing partner of mec, head of Marketing Science at bbdo, six years in consulting at Bain as a partner and head of customer engagement and marketing, Chief Customer and Data Officer at Hart Hanks. And then in February, you joined Havas Media Network as Chief Data and Product officer. And I do promise I'll stop talking in a second and ask a question, but how would you say this overview
[04:22]
A
there don't know what to add.
[04:25]
B
Well, I do have a question that I think that I think you will be able to shed light on. But how does this variety of experience, like being a quant and a consultant, inform your perspective on AI? And how do you find yourself communicating about its possibilities and its limitations to clients who are finding themselves at this very interesting inflection point?
[04:49]
A
Yeah, I think, you know, the years I've spent in media and advertising, being a practitioner really grounds you in what actually drives outcomes. And you get the fundamentals of how all of this works to drive growth for clients. So that is foundational. But my years I spent at Bain and Company really gave me a broader perspective on just the challenges that organizations face overall. And what it really confirmed for me is that the fundamentals for success continue to separate leaders from the rest. Where we think about how we drive profitable growth in market share. And it really comes from a solid understanding. And I know this is going to sound really cliche, but a solid understanding of the customer and really being very well aligned in your data strategy and the culture of test and learn in the organization in order to. To really extract the value in order to outperform competitors. We say this over and over again in our industry, but it's so true. And it still really baffles me that there are so many organizations that are chasing the shiny things and sometimes skipping over some of the fundamental things that drive success.
[06:16]
B
The shiny things are so pretty.
[06:18]
A
They are. And they're fun, too. So that's. That's the other thing.
[06:22]
B
So if you really want to take good advantage, though, of the shiny things, you do have to have a workflow in place and build the foundations and then you can go and play. But you can't do that and run until you have, like, a road to run on. Right?
[06:38]
A
That is. That is so true. And I think, you know, there is work that happens in what, at Bain, we would call the backstage or the back office. And those are the operational things. Getting the culture right, getting senior leader alignment, making sure you have the right KPIs, that everybody's driving towards having a North Star ambition for everyone to lock arms and actually work and row in the same direction. Those things cannot be discounted. And they can be facilitated both by leadership, but also at every level. There are operational things you can put in place and how you reward people, how you incentivize them to work together and collaborate that are really important to driving success. It's not just the data, it's not just the AI. There are really cultural and procedural things that can be put in place that can drive success.
[07:40]
B
And Havas has this philosophy, or maybe you'd call it an approach that you guys refer to as black book, not black box, which is very clever wordplay. So kudos on that. But there's also some real substance behind it. So what is that? What is it like when. What does it mean in practice for.
[07:57]
A
Yeah, it means a couple of things. It means, one, we, when we built our converged API platform, which is our tool that is end to end workflow from intelligence to design, to activate and to measure, and then feeding back into intelligence again. That workflow that we built is really just the underpinning and operation system that is underpinning our actual process that we have our people process. So how we work as an organization together, how we collaborate together, so the operating system can only reflect and support a good operational culture and spirit of collaboration across people. And so when we say a black book and not a black box, it means transparency at every single stage. It means that we built the operational system system and all of the data and all of the tools to be not locked into some big acquisition that we make so that we are now tied to a certain data set for life. Right? We can pivot from one type of data set to another or one type of tool to another, customized to what our clients need and not built around what we want to sell or build around what we think clients will need. So it's very bottom up outcome driven system and at every stage you can see how it's working. So as we look into a particular tool, let's say our audience planner tools, you can see how AI is actually building that audience step by step. And it explains to you what data it's using, how it's using it and how it's feeding to the final recommendation on a either Persona or profile or target audience or geographic insight. So that is what we mean by it being a black book. And then also what this allows us to do because we built it based off of client need. We created the architecture, the technology architecture that it sits on to have integration points with our clients. First party data to integrate with our clients also have technology. We don't want to be duplicative. So we didn't build a one size fits all system. And that means that the platform is flexible and it allows us to do more things that that I think some of the other platforms that our competitors have may have trouble because they built it for a specific their workflow and how they operate, but not really around as much how their different clients work based on industry, vertical, based on nuances of their business.
[11:09]
B
That does feel like saying Axiom. Without saying Axiom. That's just me saying that it does feel like the Holcos are shelling out for a lot of stuff right now and Havas decided to build rather than buy. And that explanation you gave makes perfect sense to me for what you're trying to do. I mean, buying is about moving as fast as humanly possible or like I should say agentically possible because the world is changing and it really, you know, behooves everyone to keep up with it, otherwise, you know, they're just going to be left behind. Would have us like, would it? I know this is not necessarily something that you, you get to decide, but product people and data people, they get pulled into these kinds of decisions because they get asked what's possible, what's doable, Would it make sense to buy something at some point or is it really a build over buy culture?
[12:08]
A
It is a, I'll say it build and partner culture. So we negotiate partnership deals with, for data with our providers and with our partners. We have over 500 partner integrations within our platform. We're constantly managing them and looking for new partner deals because I think when we partner we get to the best of the ecosystem. The ecosystem is very frothy and there's a lot of good stuff out there and there's also a Lot of garbage out there. And so I think the, the point of view we have is where we can partner with really excellent capabilities. We will partner with them and we will integrate them and build new solutions on top of them. So we can get exclusive, for instance, exclusive data access or exclusive access to build specific types of tools that will help our clients. And that gives us a competitive advantage, but it also allows us to help those partners even elevate their business. And we're often helping them build new capabilities on their side as well. So it is very mutually beneficial. And I think it's an exciting way to work because we're always able to see something new that we didn't think about. Right. And so we do build things as well. We build a lot of proprietary tools as well. So build and partner really is where we are. Doesn't mean that we won't ever. If there's something that really fits within our stack. I don't think, you know, we would buy a big data company or anything like that, but there might be some very nuanced AI capabilities that we might integrate and bring into the fold. I wouldn't rule that out.
[13:55]
B
Well, if you do, please tell us. We would be very interested to write about that. I want to talk a little bit about transparency, which you, you touched on briefly. Do you guys see transparency as a competitive differentiator? I don't feel like it should be, but it kind of is. I mean, everyone talks about the need for transparency. I mean, frankly, people won't sh up about it. But we're also at this moment in time when so much of the AI landscape is moving in exactly the opposite direction. And also the supply chain is still very murky. People seem to talk out of both sides of their mouth. And they spend a lot of money on pretty opaque systems like Performance Max and Advantage plus. They almost can't not spend on those things. But it feels like, I don't know, trying to keep two things in the air at the same time, two balls spinning.
[14:41]
A
Well, listen, we're agents of our clients, right? And we are working on behalf of our clients. And if we aren't protecting transparency, privacy and compliance, then I don't think we're doing our jobs very well. I think that's where our focus is. It's going to continue to focus on that. And for me, I think that when I step into the room with a CMO or a cto, this is the conversation that clients are very, very interested in having to make sure that that is maintained as we leverage AI or Agentic systems that could be doing autonomous things that could get in the way of transparency. So yeah, it's something we keep an eye on and we are committed to continuing to audit ourselves in order to
[15:30]
B
ensure transparency and just sticking with Google's performance max Meta's Advantage plus for a moment. I mean they're the two most prominent AI powered ad products out there. A lot of platforms have been releasing their own maxes and pluses. It's like the naming conventions are so funny to me because they all sound like streaming services or healthcare providers, but they're unapologetically opaque. Like the promise is that you put money in and you get results out and how is more or less of a mystery and you've got to be okay with that because hey, we're performing, why question it? Because performance, when a client comes to you and says, yeah, like those black box systems are performing well for me, like what do you, what do you say and do you recommend those platforms? Like, how does that play in?
[16:21]
A
I think that we still have to go back to fundamentals. Right. We have a brand to ensure, to protect from a long standing equity standpoint and then we also have short term performance and gains that we have to manage. I think there's a balance and I think that anytime that we're working across all of these different platforms and tools, we want to always make sure that we are balancing the two and making sure that, you know, listen, a percentage of the mix can be, you know, running for performance, but then you have to think about it as a mix and understand that there are other goals that you have to ensure as across the full consumer journey. And so that is where I believe the, the science and the, a lot of the critical thinking has to go into it before you just set it and forget it.
[17:18]
B
So I guess it's like art and data science.
[17:20]
A
It's always been that way.
[17:23]
B
It hasn't changed and really it shouldn't. Although the pendulum really swings, right? Like brands get so focused on performance that they forget that there's a top to the funnel and then have to rejigger their investment. It seems to happen more than periodically.
[17:42]
A
We used to talk about being careful not to optimize yourself into a corner. Right. I can get a hundred percent sales off of one person if that, that one person's always buying from me. Right, Right. But what about that person that's, that's out there that could, I could acquire into my brand that doesn't look as, you know, immediate when it comes to getting a return? So I think we have to think about it in a mix. We have to think about it in not just the terms of performance, but in share gain, in brand protection and brand equity. And as long as we keep those in mind, I think we will always be driving performance for today and tomorrow.
[18:30]
B
So one more quick one. Well, maybe it's not that quick actually before we take a break, but there's a principle I know that Havas leans into hard, which is that more data isn't always better. So you shouldn't really be a, shouldn't be a data hoarder. You should focus on collecting and using fit for purpose data, which is about knowing what you need and just being more purposeful. You do hear more and more support for that notion, but in a way it's almost like heretical in an industry that really fetishizes scale and volume and cheap reach. And it has since the beginning. So make the case for me, why is fit for purpose data the right frame? And why is it really important for marketers to stop treating data like they're collecting nuts for winter or something? I love that.
[19:24]
A
Listen, as a data scientist, one of our very first things we always look at is garbage in, garbage out. And so there's a lot of data that can be useful for nothing. And there's some data that is extraordinarily useful in finding those treasures of gold that are actually going to drive our clients business in a more efficient and effective way. That is where the science comes in is really being able to find what's actually going to drive outcomes. And so there's data for data sake is just not a wise way of going about an efficient search for value. Right. And so we use AI to help us mine the data and find where those nuggets of gold are. And that's what, you know, predictive and causal AI actually does. It allows us to find what are those nuggets of truth and information that allows us to drive, drive understand causality between a data point and an outcome. And so what we've found with outcomes is that they're non linear, they don't behave always in the ways that we expect. And that's why machine learning and AI, hand in hand and statistics working together are actually how we get to the right answer.
[20:58]
B
Well, AI and measurement are two of our listeners and readers favorite topics and that's what we'll spend the whole second half talking about. So stick with us. All right, we're back and I want to spend a whole bunch of time now nerding out on measurement which is a topic where there's a lot of confusion, there's a lot of noise, big promises, snake oil over the last. There is. And if you could see my inbox, I mean, the pitches I get from companies, I'm like, if you could really do that, I would have heard of you every day. Exactly. But over the last bunch of years, media mix modeling has had this massive resurgence, like a post cookie renaissance. But you guys have been talking about something that goes a step further. And you talked about it too before the break, just causal AI. So you explained what it is. But I want to go a bit deeper. Like talk to me about how it's different from MMM and how it goes a bit further than predictive AI, because I feel like it does. And what makes it interesting to you there.
[22:23]
A
There are a couple things and you know, I'd like, I like to start. The way we like to start to talk about this with clients is thinking about the maturity curve of measurement. Right. Your. Your basics are being able to report as holistically on the data that you have and the outcomes that you can track in a very direct way. Right. And that those are your dashboards, your reports. Then you have basic mmm, which is trying to do a step better than that. We used to have digital attribution, which tried to count all the beans. But we have too fragmented of a ecosystem these days to be able to track cookies everywhere. Right. Because of walled gardens and other things. So market mix modeling got a resurgence. Some people say media mix modeling, some people say marketing mix modeling. It's all really the same way of building the models. These are econometrics models that are looking at things over time and trying to tease out the causal effects. So a basic market mix model is allowing you to look at what's happening as far as a channel mix. So across the mix, what's driving sales. Or you can build a model that is looking at how you drive brand equity and how that drives sales over the longer term. But where it falls short is where you have channels that are not a significant amount of the volume. So for instance, TV is a large volume, but you might have some social media activations that may be a much smaller volume in comparison to tv, which is a mass channel. And so a lot of those effects get washed out. So what the solve for that was is something called a nested mmm. And nested mmms are trying to look at all multiple different types of outcomes. So maybe sales or sorry, brand on sales, how a brand may impact sales at different time horizons. And you might want to look at channels that often get lost in the mix by creating these weights that you can nest into the, the overall model. So it essentially is multiple models stick stitched together with statistics to allow you to talk about different outcomes and different time horizons with one model. Well, with the nested models, there's actually a shortcoming. And that shortcoming is that because it's a statistically based model, it actually has the problem of multicollinearity. And what that means is that there are a lot of variables that are correlated or they're too correlated to tease out which variable actually is driving the causal outcome. And so that means that the people who are building these models have to throw out a ton of data that might actually be explaining what's going on. And so in order to make the models understandable. Right. And so what large causal models are, it's a new entire paradigm shift, is to leverage AI to solve this problem. And it's similar to a LLM in the fact that an LLM will predict from one piece of text we call a token to the next piece of text, to the next piece of text in a probabilistic way in order to form a sentence, a line of code, a paragraph, a book. Right. And what a large causal model is similar, but it's looking at one event to the next event in a probabilistic chain. We used to use something called Markov chains to solve this, but what this actually does is it looks at those chain of events in a multi layered fashion. So now you can look at multiple chains of events all at once across different layers. And when I say layers, I mean at the portfolio level. So if you have multiple types of brands that have halo effects on each other, and then the next level down might be the channel level of brand versus performance media. And then the next level down may be specific channels like TV and radio versus different digital activations. And then the next level down from there might be creative and content and call to action and so on and so forth. And so with a large causal model, actually what the AI is doing is because it's building these probabilities from one causal event to the next, you can actually cut the chain at any place that you want and that becomes your outcome. Right. So if you wanted to understand how TV activation is actually impacting your digital, right, you can look at that in a time horizon and the digital could be an outcome which is a halo effect of TV on some digital channel. You can look at how brand is affecting sales at different time Horizons one month out, two month out, three years out, five years out. Right. But because it's multi layer, it also allows you to understand the, the relationships in a multi layered way as well. So when you optimize and say, hey, listen, I am in a year where I have a competitor entering into the market and I need to reduce my budget, what is by X amount of dollars, right, or X percent because of some kind of economic pressure that I have on my business. But I still need to drive XYZ outcome or something, number of sales, tell me what I need to do and what, what this allows you to do is build those scenarios and all of those layers from the portfolio level all the way down to call to action can be optimized with one model instead of multiple. Having to build multiple models to answer that question and then stitch them together. So that is what large causal models are allowing us to do. And I would say don't get them confused with just causal inference or causal AI inference, AI, which there are several companies that have that those are more built for, you know, test and learn, uplift modeling. And that's quite different, which is just, you know, what we think of test and learn, but automated.
[29:21]
B
So when you explain this to clients, are they nodding and do they look confused at any certain point? It's very convincing. And I'm like, yeah, of course, why wouldn't you do that? It doesn't sound necessarily easy. And I wonder how much involvement brands need to have in the process if they're set up for a culture of experimentation. So how do you make what you just talked about reality for brands?
[29:50]
A
Well, first of all, it already is reality and it's driving results for quite a few brands that have adopted it. Some of the biggest logos out there. So it's not, this is a tried and true now at this point type of methodology, there are believers and there are, I would say nine out of the 10 clients that I talk to are very bullish on this because they are hitting up against a wall when it comes to MMMs. They've squeezed all the juice they can out of nested models and they're now reaching the limitations of what MMMs can do. And so for those types of clients that see where the limitations are and are feeling the pain of that, they are ready to go to the next stage in the maturity curve, which is a large causal model. And I would say that maybe, goodness, I think there's only been one client I've ever talked to who was hesitant. And the reason there was hesitation is because there's so much noise around causal AI and frankly a lot of the causal AI players in the space are not that good. And so there's a little bit of muddying up when it comes to the signal from the noise on what companies can actually do this well and which ones can't.
[31:17]
B
Oh no, fair enough. But if measurement is an area where there's a lot of hype and a lot of noise right now, I don't think it holds a candle to how much hype and noise there is around agentic AI. And there's also confusion and conflation of terms which really doesn't help the difference between automation and agentic AI and fully autonomous AI, for example, because I've heard people use the term agentic very liberally, I've done it myself to describe basically anything that AI does which isn't right and is definitely not nuanced enough. So let's define a few terms. So automation versus agentic AI versus, I don't know, autonomous AI or whatever the term is that you hear come up most often. And what are some of the practical implications of those differences for marketers? Like why is it really important for them to know the difference?
[32:11]
A
Yes, there is a clear delineation between the three. And so automation is a rules based execution that is something that you can do with machine learning, you can do with other techniques, you can leverage AI to do it. But it's really where you're saying here, you know, you're finding where there are areas where you can build efficiencies and then you're creating rules and executing those rules in order to get the efficiency. Agentic AI is a little bit different because it is in some cases able to do chain of thought reasoning and it allows, it allows you to go out to other applications and perform actions. And so like the way a human would reason, okay, I need to go to this app and pull this piece of data, then I need to go to this other app and implement this and then actually make the action happen. And an agent workflow can actually do that. And what you want to have is not just one agent, you have a team of agents working together. So some of the agents are working as governance agents which are monitoring the outcomes or the outputs of the individual agents that are doing the steps to the action. And then you have some agents that are actually opening the doors to the APIs and checking the data and making sure that garbage is not getting poured into the process. So there's quite a few ways when you think about agentic AI but the number one characteristic is that it can perform actions based on off of, you know, reasoning and chain of thought and step by step working together with other agents. When you think about fully autonomous systems. Well, fully autonomous systems are a group of agents that work together without human oversight at all. And only 7% or so, based on a Bain study of corporations have any kind of systems that are running fully autonomously with agents. So we are far from being able to really realize the, the end to end autonomous agent world because there are a lot of unintended consequences of some of the agents actions and they still do need human oversight and governance because they're probabilistic and not deterministic. Right. A rule space is very deterministic. We can go and we can query the rules, we can look at them, we can make sure that it's being executed in the way that you intended. Whereas some of these autonomous systems as you can imagine, can be very complex and black box. And so it's very hard to tease out why an agent made a specific, did a specific action or used a specific piece of data. And so I think the, the often what we're, we're trying to do in the successful systems are you use agents to monitor the agents, but even those agents make mistakes, right? So that's where I think we're still got quite a ways to go to be fully autonomous, especially in this industry where we have a lot of risk tolerance to manage.
[36:03]
B
Where would you say most brands and agencies too are sitting in the AI adoption curve or on the spectrum or however you want to think about it, Just really honestly, because there is a gap, I think, between what I see in press releases and what people tell me and what people say on stage at conferences and what's actually happening.
[36:25]
A
I think on the client side there are a lot of promising point solutions within organizations, right? So some folks have done a really great job of building efficiency tools to help with productivity, to help with some of the tasks that are in their workflow. Where I don't see a lot of progress is the ability to take an end to end workflow across multiple teams and to break that down and actually start generating new business outcomes and growth from it in a consistent and reliable way. So it doesn't mean that it doesn't happen and there aren't companies that aren't able to unlock that. But I feel that, you know, when I, when I look across the full spectrum of companies out there and even, you know, there's Bain studies on this, that, that, that you know, you can look up, but they're, they're, they're also saying the same thing. And really the, the real reason why a lot of companies struggle to get the end to end is alignment at the top when it comes to corporate alignment with the leadership and getting the culture and the people on board to really unlock the full potential of what AI can do. And then also it's education, right? So there's quite a bit of training and understanding of being able to utilize and build enterprise grade systems and platforms that can go end to end. And then lastly, I think there's a willingness to test things and be okay with things not being perfect the first two or three times you do it and working your way to the right answer. So I think you have to have all of those in place and you have to also have a willingness to do the hard work on your data because AI consumes data and the quality of the data will give you the quality of your output. And that has, hasn't changed. I mean, quality of data has always been the number one driver for organizations for goodness, decades. Right. And, but now even more so because AI, that's the food that AI eats, right? So if we don't have good data, then AI doesn't have the fuel or the food it needs to do the jobs that it, it, it's being asked to do.
[39:07]
B
Guess we don't want it to eat out of, out of a garbage can.
[39:10]
A
Exactly. It's got to eat good nutritious food.
[39:14]
B
Gotta look at the food pyramid. What, what does an agency that's genuinely ready for agentic AI look like on the inside? Like, what are the markers? What work have they, or like have you people using AI across the org in a way that's actually helpful, that saves them time because I know it's a thing at Havas too to really democratize the use of AI. So it's not just engineering and data folks, but everybody, so it makes sense for their job.
[39:40]
A
Well, at havas, our philosophy is that everyone in our organization is an AI developer, right? And so what the wonderful thing about AI is in, in the generative AI era is that it's made itself accessible to people through natural language and through tools that are very intuitive. That said, you have to learn how to be able to prompt well and you have to understand what each component can actually do and what's possible. And so we do have a training called Prof. AI that everyone in our organization is required. Even new hires, when they come in, they take the training. We have an extensive HAVAS University of even more training for power users. And we have a leap program that gives each and every employee a certain amount of money to put towards courses as well. To take it even further, it's not a nice to have. It is a requirement for every single employee, including our admins and folks in jobs that you wouldn't think that they would need it, but everyone is required.
[40:55]
B
And what sort of guardrails need to be in place before. I don't know, you'd let an autonomous agent make a call that could impact a client's business? I know we're not really there yet with autonomous, like fully autonomous AI. It's just some small, less than 10% of. Of brands that are doing it.
[41:13]
A
But we, we have agentic layers. Um, so I don't want to go into the autonomous world because we're, we're really cautious about what we build in an autonomous fashion. Not that it doesn't exist, but we test. Test our way into it. Right. But we do have an agentic layer on our tools and we have a partnership with Accio that we work on building multiple layers of agentic architecture across our platform in order to take a lot of the manual tasks out of the things that we do. We've been able to reduce some tasks that take people three, four weeks. In fact, I had one planner tell me that, you know, something we built saved over a month of time for this individual and how it would normally happen. So we're taking a lot of the manual heavy lifting. And now what that's allowing our planners and our measurement and analytics folks or to really spend more time on the critical thinking around the insight. And I'll also say that spark of creativity that comes back when you're not spending all day in drudgery. Right. And so I think we have happier employees. Folks are way more excited about their work because they're able to just unlock value in ways that they couldn't before. And it's unlocking an amount of ingenuity and creativity that frankly, it's why I love my job when I see how excited folks are to work with all of these different tools.
[42:54]
B
Penultimate question. I can't believe we're basically out of time. And I promise this question is not coming from a place of sour grapes because I had to undergo a lot of drudgery and grunt work and heavy lifting. But isn't there some value in the drudgery? Right. Like you kind of have to do that so that you know how to prompt a system and you know what a really good Output looks like it's hard to have, you know, really creative thoughts and be able to think laterally about insights if you haven't done the grunt work, or at least there's an argument to be made about that.
[43:31]
A
Yeah, I agree. It's like, you know, math majors, right? Unless you're willing to actually do the math and derive formulas, right, you're not going to be a very good mathematician. Right? So here's what we do. We have Havas University that takes you through the full workflow from intelligence all the way out to activation. We still train folks in every step of the way. But what I think this is allowing folks to do because we built it as a black book and not a black box, because it's an open system and you can see how every decision is made. It's allowing folks to actually be more critical because now you have a better understanding of every single component that is driving a potential decision. Just because an AI serves something up to you doesn't mean it's the right thing to do. So our teams still have to be analysts, they still have to be scientists, they still have to be critical thinkers, they still have to understand the levers that drive success. They still have to understand the client's business, they still have to understand all of these different things. So when we take out the drudgery, what we mean is, listen, when back when I started in this business, we had, what is it, Lotus Notes before we had Excel spreadsheets, right? And a lot of people said, oh, well, people aren't going to know how to do things anymore if we have spreadsheets, right? Because the spreadsheet automatically populates the formula for you and all kinds of things. And no, it just unlock potential for folks to build the next level thinking, right? And so it just took out some of the work to get to the next level thinking. And I think that as long as we keep the. The black book and not black box, and we don't make it a completely closed system where humans can't interact with AI, we are learning from each other. The human is learning from the AI and the AI is learning from the human. And I think it's a symbiotic relationship now that we are going to continue to have with AI and that's how we're going to learn, and that's how we're going to actually build those skills. And so I'm actually very excited and hopeful for this next generation because they're going to get to learn things that we didn't get to learn. Because we were so inundated with the drudgery. Just like Lotus Notes versus Excel spreadsheets. I hope I don't get in trouble with. I'm sure it's okay saying that.
[46:20]
B
Well, last question. We were talking a little bit about different varieties of snake oil. What is the most just like annoying salesy thing that you hear AI startups say? Because I'm sure you see a lot of decks and a lot of pitches and I know we're early. I'll give you a second to think. Mine is that everyone says they're the first and therefore the best. Our associate editor, Joanna Gerber, she covers AI for us. She wrote this very spicy. Wasn't really that spicy, actually. It was just true op ed about. It was earlier this year about how these companies just need to stop claiming that they're the first. Like it's really okay to just have a product and explain what it is and not like brag about being the first. Like, how many first of its kind GEO platforms can there really be? Like, it can't be all of you. But yeah, what do you wish you wouldn't hear?
[47:17]
A
Again, I agree with that one. The first and the best. I think everybody claims that. But then I think that, you know, one of the things that really bothers me is when a company says, well, that they can just completely replace people without oversight. That is a red flag to me because I don't believe that we're quite there yet and everything that we build today still needs oversight. Even the biggest players in the world, Anthropic and others, they would tell you that humans need to be intimately involved in what AI is doing and to have as much ethics built into it as well. Because we know that human data that is generated out there is biased data. And we want to make sure that we're also not just getting the right answer, but we're getting the right and ethical answer so that there's fairness also built in.
[48:30]
B
So AI startups do not pitch Sharona on how you're cutting humans out of the loop because she will not like it.
[48:36]
A
No.